Nonlinear Blind Source Separation of Signals with Noise Using Neural Networks
- 格式:pdf
- 大小:197.87 KB
- 文档页数:4
独立分量分析及其应用研究梁端丹;韩政;郝家甲【摘要】独立分量分析是近年来兴起的一种高效的信号处理方法,主要解决的问题是从观测到的混合信号中分离或提取各个源信号.简要介绍了独立分量分析的模型、数学原理等基本问题,详细分析了解决独立分量分析问题的优化准则及对应的算法,最后介绍了独立分离分析的主要应用领域,并对独立分量分析问题的研究方向进行了展望.【期刊名称】《现代电子技术》【年(卷),期】2008(031)003【总页数】4页(P17-20)【关键词】盲源分离;独立分量分析;优化准则;高阶统计;信息论【作者】梁端丹;韩政;郝家甲【作者单位】西安通信学院,陕西,西安,710106;西安通信学院,陕西,西安,710106;西安通信学院,陕西,西安,710106【正文语种】中文【中图分类】TN9111 引言盲源分离(Blind Source Separation,BSS)是在源信号和传输通道参数未知的情况下,根据输入源信号的统计特性,仅由观测信号恢复出源信号的过程。
当源信号各个成分具有独立性时,此过程又称为独立分量分析(Inependent Component Analysis,ICA)。
所谓的“盲”,是指源信号的特性及传输通道的特性都是未知的。
盲信号分离的概念最早提出20世纪80年代,随后从90年代开始,盲信号分离技术广泛的应用于无线通信、雷达、声纳、图像、语音、医学等领域,迅速成为国内外信号处理研究的热点。
2 ICA的基本理论概述2.1 BSS与ICA的关系BSS是指仅从观测的混合信号中分离出各个原始信号,而ICA技术主要利用了源信号统计独立等容易满足的先验条件,为了解决BSS问题而发展起来的。
在源信号相互独立时,BSS和ICA具有相同的模型,但实际情况下,两者目标上稍有不同。
BSS的目标是分离出源信号,即使他们并不完全互相独立;而ICA的目标则是寻找某种变换,使输出的各信号之间尽可能的独立。
ICA是BSS的一种方法,但是解决BSS问题还有其他的理论方法,如非线性主分量分析、稀疏分量分析(SCA)等,不局限于ICA。
眼电伪迹自动识别与去除的新方法李明爱;郭硕达;田晓霞;杨金福;郝冬梅【摘要】为了改善脑电中的眼电伪迹过估计问题及环境干扰耦合引起的非线性混合对眼电去除效果的影响,提出一种基于快速核独立成分分析(Fast Kernel Independent Component Analysis,FastKICA)与离散小波变换(Discrete Wavelet Transform,DWT)的眼电自动去除方法,即(Fast Kernel Independent Wavelet Transform ,FKIWT)方法。
首先,利用 FastKICA 方法对脑电信号进行分离得到独立成分,并以相关系数为依据识别出眼电伪迹;进而,基于 DWT 对眼电伪迹进行多分辨率分析,将逼近分量置零,而细节分量保持不变,使得重构所得眼电伪迹成分保留更多有用脑电信号;最后,利用 FastKICA 逆变换重建眼电去除后的脑电信号。
实验结果表明:FKIWT 不仅有效改善了眼电过估计问题,增强了抗干扰能力和鲁棒性,而且在线性混合和非线性混合情况下,均得到较好的伪迹去除效果,特别是在非线性混合时优势更为明显,适合于实际在线应用。
%In order to improve the overestimation of ocular artifacts (OA)in electroencephalogram (EEG)and the OA removal effect of nonlinear mixture caused by environmental interference coupling,a novel automatic removal method is proposed based on fast kernel independent component analysis (FastKICA)and discrete wavelet transform (DWT),and it is denoted as FKIWT.The independent components are separated from the mixed EEG by using the FastKICA algorithm,and the correlation co-efficient is applied to identify OA component;Then,the Multiresolution analysis of OA is achieved with DWT,the approximation wavelet coefficients are set to zero and the detail wavelet coefficients are not changed.So more usefulEEG is remained in the re-constructed OA component;Furthermore,the clean EEG is restored with the inverse algorithm of FastKICA.The experimental re-sults show that FKIWT can effectively improve the overestimation of OA and has perfect anti-interference ability and robustness. Meanwhile,the better effects of OA elimination are also obtained on the condition that the linear or nonlinear mixed model is a-dopted,and the latter’s advantage is especially obvious.The FKIWT is suitable for on-line application.【期刊名称】《电子学报》【年(卷),期】2016(044)005【总页数】8页(P1032-1039)【关键词】非线性混合模型;快速核独立成分分析;离散小波变换;眼电过估计;鲁棒性【作者】李明爱;郭硕达;田晓霞;杨金福;郝冬梅【作者单位】北京工业大学电子信息与控制工程学院,北京 100124; 计算智能与智能系统北京市重点实验室,北京 100124;北京工业大学电子信息与控制工程学院,北京 100124;北京工业大学电子信息与控制工程学院,北京 100124;北京工业大学电子信息与控制工程学院,北京 100124; 计算智能与智能系统北京市重点实验室,北京 100124;北京工业大学生命科学与生物工程学院,北京 100124【正文语种】中文【中图分类】R318脑电信号(Electroencephalogram,EEG)是通过电极在头皮表面采集到的反映大脑内部状态的生物电信号,在神经科学、脑科学、临床医学与康复工程等领域具有十分重要的作用.脑电信号采集过程中极易受到眼电、肌电、心电等多种噪声干扰[1],而眼电信号由于幅值较大,严重影响脑电信号的分析和应用,如何有效去除脑电信号中的眼电伪迹尤为重要[2].盲源分离(Blind Source Separation,BSS)方法是在源信号信息及混合过程都未知的情况下,仅需对观测信号进行处理就可实现对源信号和系统的辨识,近年来在信息处理领域备受青睐而发挥着越来越重要的作用[3].目前,用于去除脑电信号中眼电伪迹的常用BSS方法有:(1)基于矩阵联合对角化的预白化算法(Joint Approximate Diagonalization of Eigenmatrixes,JADE).JADE算法引入多变量数据的四维累积量矩阵,通过特征分解使其得以简化,并提高了结果的稳健性[4,5];(2)PICA算法(Pearson Independent Component Analysis,PICA).该算法将ICA中的固定非线性对比函数方法与最大似然估计方法相结合,提高了对不同分布的源信号的分离能力,但计算量较大[6];(3)快速独立向量分析(Fast Independent Component Analysis,FastICA)算法.该算法是在传统ICA算法的基础上发展起来的一种快速寻优迭代算法,因采用了定点迭代优化算法,具有较快的收敛速度[7,8];(4)离散小波变换(DWT)和FastICA相结合的一种算法,记为DWICA算法[9].该算法的本质是在小波域上利用FastICA完成眼电去除,不仅收敛速度更快,而且增加了算法的稳健性.上述四种方法具有两个共同特点:(1)将分离或判断得到的眼电伪迹直接去除.因算法本身不能保证伪迹信号完全分离,而且噪声对分离效果也有一定影响,因而,分离出的眼电伪迹中依然包含有用的EEG成分,直接或全部去除会造成眼电伪迹的过估计[10];(2)实际采集的EEG信号符合线性混合模型的假设[11,12].由于EEG信号作为一种生物电信号十分微弱,且易受外界环境多种干扰耦合的影响,实际采集到的EEG不可避免的会发生非线性畸变,因此,脑电和眼电信号的非线性混合更加符合实际,这使得上述BSS算法在眼迹去除问题中具有一定局限性,导致分离能力和稳定性有所下降.核独立成分分析法(Kernel Independent Component Analysis,KICA)是近几年为解决BSS问题而提出的一种新方法,它利用核函数将信号从低维空间映射到高维空间,从而将非线性问题转换到高维空间的线性问题,对非线性混合问题有较好的分离能力和更好鲁棒性[13,14].FastKICA算法是在KICA的基础上发展而来的,具有更快的运算速度[15].本文提出一种基于FastKICA和DWT算法的眼电伪迹自动识别与去除方法,即FKIWT方法.采用FastKICA算法对脑电混合信号进行盲源分离,并以相关系数识别出眼迹成分,进而对分离出的眼电伪迹基于DWT进一步分析和处理,再利用FastKICA逆算法重建眼迹去除后的脑电信号.实验研究表明,FKIWT方法在有效改善眼电过估计问题、增强方法自身抗干扰能力和鲁棒性等方面均得到较好效果. 在源信号分量独立且最多只有一个分量服从高斯分布的前提下,KICA模型如式(1)所示.KICA方法首先利用核函数K(xi,xj)(i,j∈{1,2,…,n}且i≠j)来代替向量xi和xj间的内积,以实现将观测信号从低维空间映射到满足可再生核希尔伯特空间(Reproducing Kernel Hilbert Space,RKHS)特性的高维特征空间F;然后,以RKHS内的非线性函数作为对比函数,并运用对Gram矩阵的低阶近似等方法在RKHS内搜索对比函数的最小值[13],以求解分离矩阵W=H-1;进而,求得对源信号的估计Y=WX.对比函数最小值的求解等价于式(2)所示最小广义特征值的求解问题:FastKICA算法是在KICA的基础上发展而来的,该算法利用Hilbert-Schmidt独立性判决准则(Hilbert-Schmidt Independence Criterion,HSIC)作为衡量变量统计独立性的对比函数,用牛顿类法对对比函数进行优化,通过极小化对比函数,获取分离矩阵,并采用不完全Cholesky分解方法来提高计算性能[15]. FastKICA的基本工作过程如下:(1)确定观测信号X=[x1,x2,…,xn]T和核函数K(·,·).(2)对观测信号进行去中心化和白化处理.(3)基于HSIC准则定义对比函数C(W).(4)对对比函数C(W)的Hessian矩阵进行不完全的Cholesky估计.(5)用牛顿类法优化求解分离矩阵W.小波变换是一种有效的时频分析方法,具有多分辨分析和对信号的自适应特点,广泛应用于非平稳信号分析中.对任意离散函数f(t)∈L2(R),其离散小波变换定义为:相应的DWT逆变换定义为Mallat将计算机的多分辨率分析思想引入到小波分析中,统一了正交小波基的构造,提出了离散小波变换的快速分解和重构算法,即Mallat塔式分解算法,显著减少了DWT的运算数据量[16].图1给出了Mallat算法的信号分解过程示意图.图中,和分别表示低频和高频分解滤波器系数,↓2表示下采样过程.可见,信号分解过程可以看作是将信号通过一组高通和低通滤波器,然后将低通滤波得到的逼近分量再次通过高通和低通滤波器.随着空间尺度j由1逐级增大,完成信号的多分辨分解,最终得到信号的逼近分量和各阶细节分量,相应计算式如下[17]:图2展示了Mallat算法信号重构过程.图中,h和g分别是和的对偶形式,↑2表示上采样过程.Mallat算法信号重构满足下式:基于FKIWT进行眼电伪迹自动识别与去除,主要步骤如下:(1)基于FastKICA算法的眼电伪迹分离假设X(t)=[x1(t),x2(t),…,xn(t)]T∈Rn×M为n导观测信号,其中,M表示每导信号的样本点数,xi(t)(i=1,2,…,n-1)为n-1导脑电信号,xn(t)为眼电参考信号.利用FastKICA算法对X(t)进行核独立成分分析,得到分离矩阵W;进而,根据Y(t)=WX,(t)得到n个独立成分:(2)利用相关系数识别眼电伪迹相关系数用以描述两变量之间的相关性,其绝对值越大表明两个变量相似度越高.依式(10)计算每个独立成分yi(t)(i=1,2,…,n)与眼电参考信号xn(t)的相关系数:(3)基于DWT的眼电伪迹分析与重构采用Mallat算法对眼电伪迹成分yeog(t)进行L层离散小波分解,依式(7)计算逼近系数分量AL和细节系数分量{DL,DL-1,…,D1}.令逼近系数分量AL=0,细节系数分量保持不变,进而依式(8)所示Mallat塔式重构算法进行离散小波逆变换,实现信号重构,得到新的眼迹信号Dj.(4)利用FastKICA逆变换重建眼电去除后的脑电信号将替代式(11)中的yeog(t),有依式(13)对(t)进行FastKICA逆变换,将眼电伪迹去除后的脑电独立成分投影变换到各个电极上,以重建伪迹去除后的脑电信号.本实验纯净脑电信号采用BCI Competition II的Data set Ⅲ数据库,该数据库记录了一位25岁的女性受试者左右手运动想象的EEG信号.该实验包含140组训练数据,其中左右手运动想象各70组.每组实验持续9s,实验开始后0~2s休息,2~3s出现“+”准备提示符,3~9s出现箭头(向左或向右)运动想象提示,实验时序如图3所示.该实验利用G.tec脑电采集仪和Ag/AgCl电极采集了C3、Cz和C4三导的EEG 数据,采样电极位置如图4所示,采集频率为128Hz,并对数据经过0.5~30Hz 滤波.眼电数据来源于BCI Competition IV的Data sets 2b中采集的垂直眼电参考信号.由于采集的眼电信号中不可避免的会混有脑电信号,为了还原纯净的眼电信号,本文根据眼电信号的频率范围,对眼电数据采用FIR低通滤器,进行0~10Hz的低通滤波,去掉其中的高频信号后,作为纯净眼电信号.本部分将基于线性混合模型,利用FKIWT方法从眼电过估计、伪迹去除及鲁棒性等多方面开展眼迹去除实验研究,以展示本文提出方法的有效性.实验环境为matlab2014a.根据眼电和脑电间的双向传递特性,构建线性混合模型如下:均方误差(Mean Squared Error,MSE)是普遍采用的一种伪迹去除评价指标,其计算式如下:本节将从FastKICA算法和FKIWT方法的工作过程入手,通过计算每种方法去除掉的眼迹成分与纯净眼电信号的相关系数,检验本文方法改善眼电伪迹过估计能力和抗干扰能力.FastKICA算法中,核函数选用高斯径向基函数,核函数宽度取值为1,迭代精度设为0.0001,最大迭代次数设为10000.FKIWT方法中,FastKICA算法的参数选定同前;DWT选用coif小波基函数,小波分解层数为3.令X(t)=[C3real,Czreal,C4real,EOGreal]T为实际的脑电和眼电观测信号,由式(14)所示线性混合模型产生,影响因子k1至k6分别取值为0.2,0.2,0.2,0.05,0.1,0.15.采用FastKICA对X(t)进行核独立成分分析,并基于相关系数判断得到眼电伪迹成分yeog(t);FKIWT方法则在此基础上对yeog(t)进行离散小波变换,获得小波系数{A3,D3,D2,D1},在保留细节系数{D3,D2,D1}不变、逼近系数A3置为零的情况下,基于小波逆变换重构出眼迹成分Dj;计算eog(t),则ydel(t)体现了FKIWT方法真正去除掉的眼电伪迹成分.由于eog(t)仅保留了yeog(t)的细节部分,即ydel(t)实际上为yeog(t)的逼近分量,这与眼电信号的频率特性更吻合,因而,ydel(t)与纯净眼电信号EOGcle的相关系数会大于yeog(t)与纯净眼电信号EOGcle的相关系数,这意味着FKIWT方法眼迹去除过程中保留了更多有用的脑电信号,有效减少了眼电伪迹的过估计问题.图5给出任取一组脑电数据的眼电过估计实验结果.图中,EOGcle表示纯净眼电信号,yeog(t)为FastKICA算法分离和去除掉的眼电伪迹,ydel(t)则为基于FKIWT方法去除掉的眼电伪迹.由图可见,yeog(t)依然包含较多的脑电信号,而ydel(t)与EOGcle的波形更为相似,说明FKIWT方法相对FastKICA算法而言,去除掉的眼电伪迹包含更少的脑电成分,有效改善了眼电过估计问题.下面对140组脑电数据进行实验,并在线性混合模型中加入不同强度的白噪声,用来模拟脑电采集过程中受到的心电、肌电、出汗等其它干扰的影响,计算yeog(t)和ydel(t)分别与EOGcle间的相关系数reog和rdel.图6展现了从无噪声到有噪声,且噪声强度从-5dBw逐渐增强至10dBw时,基于140组数据所得平均相关系数的变化曲线.从图6可知,随着噪声的增强,相关系数rdel变化很小,reog变化较大且减小趋势明显,说明本文FKIWT方法相对FastKICA算法具有更好的眼迹去除效果,不仅有效保留了有用的脑电信号,有利于减弱眼电过估计影响,而且具有更强的抗干扰能力.本节将基于线性混合模型,以相关系数和均方误差为性能评价指标,将本文方法与其它常用方法进行比对实验研究,检验FKIWT方法的眼迹去除能力.图7呈现了某组纯净的脑电信号和眼电信号,图8则进一步给出了依据式(14)进行线性混合并加入-5dBw的高斯白噪声的脑电信号和眼电信号.这里,影响因子k1、k2和k3设定为0.2到0.4之间的随机数,k4、k5和k6设定为0.02到0.3之间的随机数,高斯白噪声用以模拟采集过程中眼迹之外的其它干扰.由图清晰可见,在采样点300和800附近,眼电信号对脑电信号产生非常强烈的扰动.利用本文FKIWT方法对图8所示混有眼迹和干扰的脑电信号进行处理,实验结果如图9.和图7对比可知,相应导联脑电信号波形非常接近,去噪效果良好.进而,随机产生20个线性混合矩阵H,采用多种方法基于140组脑电数据各进行2800次眼迹去除实验.表1显示了基于JADE、PICA、FastICA和DWICA四种常用方法及FKIWT法去除眼电伪迹后的脑电信号与纯净脑电信号的平均相关系数(r)和均方误差(MSE).这里,FastICA及DWICA方法中的ICA均选用基于负熵判据的FastICA算法,其迭代精度和最大迭代次数与PICA及FKIWT中的FastKICA设置相同,DWICA中DWT与FKIWT中DWT参数选择相同.由表1可知,FKIWT方法得到三导脑电信号的平均MSE较JADE、PICA、FastICA和DWICA分别减少了46.1%、28.8%、13.4%和8.3%;而平均相关系数则分别提高了0.0029、0.0013、0.0005和0.0003.显然,FKIWT法相对其它方法在C3、Cz和C4三导脑电信号上均取得了最大相关系数和最小均方误差,效果优势明显.方法的鲁棒性对于其能否获得稳定的实验结果及在线应用十分重要.为此,在4.2.4节实验的基础上,进一步计算眼电伪迹去除后的脑电信号与纯净脑电信号的相关系数的方差(var),以比较不同方法眼电分离效果波动的大小,从而评价各方法的鲁棒性.图10给出了五种方法在三导脑电信号上获得的平均实验结果.从图10可知,FKIWT方法在三导脑电上均得到了最小相关系数方差,特别是相对于JADE和PICA方法优势尤为明显,说明FKIWT方法具有较好的鲁棒性,更适合应用于实时脑机接口系统中.脑电信号采集过程中易受外界环境多种干扰耦合的影响,导致实际采集到的EEG 会不可避免地发生非线性畸变.为此,本节选用后置非线性混合模型作为脑电与眼电信号的混合模型进行实验研究,如图11所示.图中,xi,cle(t)表示第i导纯净的脑电信号或眼电信号,xi(t)为线性混合后的信号,i(t)为非线性畸变后的信号,i=1,2,3,4;f(·)为非线性畸变函数,H为式(14)中的线性混合矩阵.考虑生物电信号的自身特性及干扰耦合的特点,本节选择以下三种非线性函数模拟非线性畸变:(1)多项式函数f1(·)(2)三角函数f2(·)(3)双曲函数f3(·)下面将基于多种非线性畸变函数的非线性混合模型,针对FKIWT和其它常用方法的眼迹去除问题展开实验研究.为便于对比,实验环境及各方法中参数取值同4.2节.为了简化非线性混合模型并排除随机线性混合矩阵H对分离结果的影响,实验中将H设定为固定值,影响因子k1到k6分别取值为:0.2、0.2、0.2、0.05、0.1和0.15;根据生物信号非线性畸变的特点,有关非线性函数系数分别取值为a=0.01,b=0.001,c=30,d=π/120,g=π/120,p=30.将140组脑电数据进行后置非线性混合,获得含有眼迹的脑电数据,再利用FKIWT和其他四种方法完成眼迹去除,以均方误差和相关系数为评价指标,平均实验结果如图12和图13所示.从图可见,基于非线性混合模型,采用三种非线性畸变函数时,本文FKIWT法相对其它四种方法均取得了最佳实验效果.针对每种非线性函数情况,MSE均有所减小,相关系数都有所增大.其中,基于三种非线性函数在三导脑电信号上的平均MSE较JADE、PICA、FastICA和DWICA分别减少了36.1%、24.1%、20.4%和18.3%;而平均相关系数则分别提高了0.0051、0.0020、0.0008和0.0005.说明本文方法在非线性混合时的分离效果上较其他四种方法有较大的改善,且优于线性混合模型情况下的实验结果.本文提出一种FastKICA和DWT相结合的眼迹自动识别与去除方法,即FKIWT法.该方法的特色主要体现在从脑电与眼电信号时频特性的差别入手,利用DWT 的多分辨特性仅对眼电伪迹成分进行处理与分析,通过置近似小波系数为零、细节小波系数不变,使得去除的眼电伪迹成分更加逼近纯净的眼电信号.基于国际标准数据库的大量实验研究表明,FKIWT法不仅有效减少了眼电过估计问题,具有较强的抗干扰能力和鲁棒性,而且在线性混合模型和非线性混合模型下,相对其它常用方法而言,FKIWT法均取得了最佳的眼迹去除效果,特别在非线性混合模型下优势更为明显,这为FKIWT法真正应用于脑电信号的实际在线分析与处理奠定了基础,另外,本研究对于更好地提示脑电混合方式和脑电建模具有重要理论意义. 李明爱(通信作者) 女,2006年于北京工业大学获得博士学位,现为北京工业大学副教授,硕士生导师,主要研究方向为脑机接口技术与智能控制.郭硕达男,2012年于武汉纺织大学获得学士学位,现为北京工业大学控制科学与工程专业硕士研究生,主要研究方向为脑机接口技术、信息处理与模式识别.。
非线性盲源信号分离研究甘晓晔;李丽娜;宁晓霞【摘要】在工程应用中,常常遇到非线性混合的信号,而基于线性混合假设下提出的盲源分离算法在一般情况下对非线性混合问题可能失效或者导致错误的结果.鉴于此,文章对非线性盲源分离技术进行了初步的研究,为进一步扩展盲源信号分离的应用范围提供了理论依据.【期刊名称】《辽宁科技学院学报》【年(卷),期】2010(012)003【总页数】3页(P26-28)【关键词】盲源分离;非线性;混合信号;目标函数【作者】甘晓晔;李丽娜;宁晓霞【作者单位】【正文语种】中文【中图分类】TP301.6盲源信号分离(BSS)问题研究发展至今,解决线性盲源信号分离问题的算法发展较为迅速,一些成熟的盲源信号分离算法己经在许多领域得到了广泛的应用。
在工程实际中应用盲源信号分离时,绝大部分假设信号是线性混叠的,然而在很多情况下,如在机械设备故障诊断过程中,当信号通过传感器时很可能会发生非线性畸变或混合,信号常常是非线性或者弱非线性的,进行线性假设只不过是非线性的一种近似,用线性模型来描述也显得过于简单。
由于基于线性混合模型的算法没有对系统中的非线性特征进行消除或者补偿,因此基于线性混合假设下提出的分离算法在一般情况下对非线性混合问题可能失效或者导致完全错误的结果〔1〕。
这就要求去寻求适用非线性混叠情况的分离方法,非线性盲源信号分离的研究,可为进一步扩展盲源信号分离的应用范围提供理论依据。
目前,针对非线性盲源信号分离的研究方法主要有二种:一种是基于自组织映射(SOM)直接提取非线性分量的方法;另一种是基于假定非线性混合模型的方法。
前一种基于SOM非线性BSS算法是通过一种非线性映射实现输出向量的盲源解耦和分离,由于在实际应用中,SOM算法自身存在诸多缺点,如随着源信号数目的增加,SOM算法的网络复杂度呈指数增长,且对连续源信号分离时产生插值误差等等,因此限制了SOM算法的发展和应用范围。
后一种基于假定非线性混合模型的方法是先假定非线性混合模型,通过该模型来拟合实际混合中的非线性,然后通过神经网络等算法构建解混器,将解混后的向量通过线性BSS/ICA方法来实现分离,这是目前大部分学者积极研究的方法。
Statistics for complex variables and signals - Part I: VariablesAbstractThis paper is devoted to the study of higher-order statistics for complex random variables. We introduce a general framework allowing the direct manipulation of complex quantities: the separation between the real and the imaginary parts of a variable is avoided. We give the rules to integrate and derive probability density functions and characteristic functions, so that calculations may be carried out. In the case of multidimensional variables, we use the natural framework of tensors. The study of complex variables leads to the extension of the notion of complex circular random variables already known in the Gaussian case.1. IntroductionHigher-order statistics (HOS) are now an intensive field of research in Signal and Image Processing. This avenue of research is based on the use of a new characterization of variables and signals. Up to now this characterization was essentially based on second-order (energetic) measures: variance and covariance for variables, correlation and cross correlation for signals in the time domain, pectral power density and cross-spectral power density for signals in the frequency domain.After the pioneering papers of the potentialities of HOS are now used intensively. It would be very long to give a complete view of this domain in which new models are emerging that support the development of a large number of applications. A synthesis can be found in Furthermore, several special issues of journals have been devoted to this topic and a series of specialized workshops began in 1989.The essential features in research on HOS are found in modeling and in applications.In modeling, for random variables, HOS are essentially based on cumulants of order greater than 2. The higher-order description of signals is made through multicorrelations in the time domain, and multispectra in the frequency domain.Applications are being developed in a great number of fields.In nearly all theclassical domains of research in Signal and Image processing. HOS are introducing new methodologies. We can cite the blind source separation and blind deconvolution problems in a wide variety of situations: vibrations diagnostic, underwater acoustics, radar, satellite communications, seismic sounding, astronomy, etc. In nonlinear systems identification, HOS are a basic tool. Moreover, a close connexion exists between HOS and neuromimetic systems.This very active and fruitful field of research needs solid theoretical foundations. They were built a long time ago by mathematicians and statisticians who developed the theory of random variables and signals. The higher-order statistical properties of random variables are described in many classical textbooks . We found the development of atensorial approach particularly well fitted to the higher-order properties of multidimensional variables. The multicorrelations and multispectra are described.However, few authors have been concerned with the domain of complex random variables and signals, even if this situation appears in practical applications: in frequency domain processing after Fourier transformation, particularly in array processing, in single band systems of communications where analytic signals are commonly used, in time-frequency analysis by the Wigner-Ville distribution, etc.The Gaussian complex model, which is sufficient in the classical second-order approach. These authors have shown the algebraic simplifications brought by the use of a complex modeling. They have shown that new properties, like Gaussian complex circularity, are introduced by this complex modeling.More recently the lack of a general complex modeling was put in evidence: the authors noted that “paradoxically, one finds in the literature very few treatments of com plex random variables and processes”. They introduce the notion of “proper complex random processes” which is their denomination for circular processes. However, this approach is essentially limited to the second-order properties.This particular character of complex signals, when one is concerned with the bispectrum,has been exemplified.With the increasing use of higher-order statistics, it is now necessary to develop a general modeling for complex random variables and signals. It is the aim of this exposition, which is divided into two parts.In the first part, we are concerned with complex random variables. We begin by the definition of the probability laws using complex notations. We extend the results,which are already known for the Gaussian case to the general situations of monodimensional and multidimensional complex random variables, whether Gaussian or non-GausSian. Then, we extend the tensorial formalism developed in the real case to the multidimensional complex random variables. We show that, for a given order, different kinds of cumulants can be defined. This result is an extension of the pseudo-covariance introduced. With this modeling we can give a general definition of circularity, and we show that, in this specific case, many higher-order cumulants are null. We show the direct relation between the Fourier transform and circularity. Algorithms for the generation of complex circular non-Gaussian random variables are given and illustrated on simulations. An illustration of the new rules of calculation is given in Appendix B in the circula Gaussian case.Part II is devoted to the modeling and representation of complex random signals.For stationary signals, using the results given for the multidimensional random variables, we define the multicorrelations and multispectra for complex random stationary signals. We show that the complete characterization of complex signals at an order pdemands the introduction of different multicorrelations and multispectra. In the usual case of real valued signals these multicorrelations and spectra are identical. The situation is different for analytic signals for which some multicorrelations and spectra are null. Extending the concept of circularity to the signals, we can show that, for this kind of signal, the only nonnull multicorrelations and spectra possess the same number of conjugated and nonconjugated terms.Furthermore, we show that band limited signals are circular up to a certain order. We come back to the choice between moments and cumulants and show that, except for the classical interest presented by cumulants due to their additivity and to their characterization of the Gaussian property, they make it possible to distinguish clearly between the properties at each order and to eliminate singularities in the multispectrum. This modeling is then extended to digital signals and to digital time-limited signals used in the Discrete Fourier Transform.2. Starting pointThe purpose of this paper is to introduce a general model of complex random variables. The usefulness of this modeling will be illustrated with examples. We will show that it leads to new characteristics in the description of signals and allows a new insight into the developing field of higher-order statistics.Complex random variables (CRV) appear as the output of a great number of processingr such as:- Fourier transforms,- Array processing,-- Hilbert transforms.When dealing with CRV two approaches can be used:-to consider a CRV as a two-dimensional real random variable (RRV),-to develop algebraic tools directly with CRV.The second approach has two advantages:-it makes all the derivations simpler,-it preserves the physical sense related to the complex nature of the data.This approach has been developed in the Gaussian case leading to the theory ofcomplex Gaussian random variables (CGRV). In this situation the consideration of the CGRV has given rise to the important notion of complex random Gaussian circular variables.The principal aim of this paper is to generalize these notions to the general case of Gaussian and non-Gaussian random variables. The primary motivation is that new algorithms using higher-order statistics (HOS) are being developed, and it is clear that in this field, it is absolutely necessary to deal with non-Gaussian data. Furthermore, a theory will be developed using tensors which constitute he natural framework of higher-order statistics. Hence, the second main issue of this paper is the extension of the framework introduced by MacCullagh to the complex case.After a definition of the basic principles on which our modeling is built, we will present the technical realization of the principal tools. We will give a generaldefinition of CRV and illustrate the usefulness of this new formalism in the context of complex circular random variables.2.1. Complex random variablesThe definition of CRV is well-known. From two real random variables (RRV) x and y , we define the complex random variable z by z x jy =+.(1)where 21j =.The turning point is to associate a probability density function (pdf) with this CRV.In the Gaussian circular case, this is done by onsidering both z and its complex conjugate z * defined as z x jy *=-.The …formal‟ pdf of the Gaussian circular variable is then()2,1,zz z z P z z e σπσ**-*=Thus, it appears from the Gaussian example that we must consider both z and z*in the definitions in order to extract all the statistical information. The preceding definition for the complex Gaussian variable shows that E[''z] equals zero. Hence, the only nonnull second-order moment is E[z z*]. This means that both z and z*give statistical (and perhaps different) information. Therefore, a theory of higher-order statistics in a general case must consider the variable and its complex conjugate. The information is in the statistics of the two variables, but also in their cross-statistics. We now introduce our formalism to handle the complex random variables in a more general way.The main problem which arises from the preceding discussion is an algebraic one, since the variables z and z*are algebraically linked. In order to overcome this, we propose to include the real world of z and z*in a larger space in which z and z*are not algebraically dependent.One way to do this is to consider x and y(real and imaginary parts of z)as complex random variables.In this context we will continue to*=-,but despite the notations,z and z*are no longer write z x jy=+and z x jycomplex conjugates. In order to introduce a continuity between the classical notations and the new ones, using tensors, that will be presented shortly, we have chosen to use these ambiguous notations of z and z*.We will introduce in the following an alternative presentation that avoids this problem.This …trick‟ will allow us to treat z and z*as algebraically independent variables. We will see that this greatly facilitates all the calculations. However, these purely conceptual CRV are only used as means for easier calculations. When we want to come back to the real physical world, we have to restrict z and z*to belong to the subset generated by the real numbers x and y.2.2The rulesFor this, we must establish some rules in order to obtain definitions which make sense. We will propose two laws:1. All the functions used must be well-defined mathematically, and all the operators, like integrals, must converge.2. We want to be able to recover the classical formulae when we consider the particular case of RRV.This new point of view applies to both one-dimensional random variables as multidimensional random variables.We will now see how it works.3. PDF and characteristic functionsWe will consider successively the one- and the multi-dimensional cases of random complex variables.3. 1One-dimensional complex random variablesWe have seen in Section 2.1 that the pdf is a function (),,z z P z z **of z and z *.Let us try to define the first characteristic function. Let ωand ω*错误!未找到引用源。
ANSYS Workbench 中的几种载荷的含义1) 方向载荷对大多数有方向的载荷和支撑,其方向多可以在任意坐标系中定义:–坐标系必须在加载前定义而且只有在直角坐标系下才能定义载荷和支撑的方向.–在Details view中, 改变“Define By”到“Components”. 然后从下拉菜单中选择合适的直角坐标系.–在所选坐标系中指定x, y, 和z分量–不是所有的载荷和支撑支持使用坐标系。
2) 加速度(重力)–加速度以长度比上时间的平方为单位作用在整个模型上。
–用户通常对方向的符号感到迷惑。
假如加速度突然施加到系统上,惯性将阻止加速度所产生的变化,从而惯性力的方向与所施加的加速度的方向相反。
–加速度可以通过定义部件或者矢量进行施加。
标准的地球重力可以作为一个载荷施加。
–其值为9.80665 m/s2 (在国际单位制中)–标准的地球重力载荷方向可以沿总体坐标轴的任何一个轴。
–由于“标准的地球重力”是一个加速度载荷,因此,如上所述,需要定义与其实际相反的方向得到重力的作用力。
3) 旋转速度旋转速度是另一个可以实现的惯性载荷–整个模型围绕一根轴在给定的速度下旋转–可以通过定义一个矢量来实现,应用几何结构定义的轴以及定义的旋转速度–可以通过部件来定义,在总体坐标系下指定初始和其组成部分–由于模型绕着某根轴转动,因此要特别注意这个轴。
–缺省旋转速度需要输入每秒所转过的弧度值。
这个可以在路径“Tools > Control Panel >Miscellaneous > AngularVelocity”里改变成每分钟旋转的弧度(RPM)来代替。
4) 压力载荷:–压力只能施加在表面并且通常与表面的法向一致–正值代表进入表面(例如压缩) ;负值代表从表面出来(例如抽气等)–压力的单位为每个单位面积上力的大小5) 力载荷:–力可以施加在结构的最外面,边缘或者表面。
–力将分布到整个结构当中去。
观测噪声方差变化条件下系统状态估计方法宋婉娟; 张剑【期刊名称】《《探测与控制学报》》【年(卷),期】2019(041)005【总页数】6页(P96-100,105)【关键词】非线性状态估计; 容积卡尔曼滤波; 变分贝叶斯滤波; 噪声统计特性【作者】宋婉娟; 张剑【作者单位】湖北第二师范学院计算机学院湖北武汉 430205; 武汉体育学院体育工程与信息技术学院湖北武汉 430070【正文语种】中文【中图分类】TP202; TP3910 引言非线性系统状态估计问题是工程应用中的难点和瓶颈,备受研究人员关注[1-3],但是由于该类系统参量的未知性,很难建立精确的数学模型,目前常用的解决方法主要是采用滤波估计,数值逼近的思路解决。
例如,扩展卡尔曼滤波(Extended Kalman Filter,EKF)[4-7],通过对非线性系统进行一阶泰勒近似,将非线性系统进行局部线性近似处理,虽简单易行,但存在明显的截断误差,噪声敏感性强;无迹卡尔曼滤波(Unscented Kalman Filter, UKF)[8-10],通过对模型统计参量的非线性变换,实现系统的高阶近似,但是其估计精度严重依赖于系统初值和观测噪声,当噪声参量时变情况下跟踪精度会降低;粒子滤波(Particle Filter, PF)[11-15]通过粒子加权求和的方法对系统积分进行拟合,在粒子数足够多的情况下是可以无限逼近系统真实状态的,但是该方法的滤波精度和实时性之间的矛盾至今尚未有效解决,限制了其工程应用。
针对这些问题,文献[16]采用高阶球面积分传播容积点的思想提出了容积卡尔曼滤波(Cubature Kalman Filter, CKF)方法,对状态的均值和协方差进行非线性传播,同EKF、UKF以及PF等非线性滤波方法相比,不仅提升了滤波精度,计算的复杂性也大大降低,在目标跟踪、状态估计等领域得到了广泛的应用。
但是标准CKF方法严重依赖于精确的状态模型和噪声统计特性[17-19],而在实际的工程应用中很难满足,特别是观测噪声方差,具有较强的时变特性,很难满足噪声方差统计特性精确已知的要求[20],限制了其在工程中的推广应用。
盲信号分离作者:张贤达, 保铮作者单位:西安电子科技大学雷达信号处理重点实验室,陕西西安,710071刊名:电子学报英文刊名:ACTA ELECTRONICA SINICA年,卷(期):2001,29(z1)被引用次数:155次1.Amari S A theory of adaptive pattern classifiers 19672.The fast-ICA MATLAB package3.Zhang Q;Leung Y W A class of learning algorithms for principal component analysis and minor component analysis[外文期刊] 20004.Yang H H;Amari S;Cichocki A Information-theoretic approach to blind separation of sources in nonlinear mixture[外文期刊] 1998(3)5.Murata N;Muller K R;Ziehe A;Amari S Adaptive on-line learning in changing environments 19976.Makeig S;Bell A;Jung T P Independence component analysis in electro-encephalographic data 19967.Basak J;Amari S Blind separation of uniformly distributed signals:A general approach[外文期刊] 1999(5)8.Hyvarinen A Fast and robust fixed-point algorithms for independent component analysis[外文期刊] 1999(03)9.HERRMANN M;Yang H H Perspectives and limitations of self-organizing maps in blind separation of source signals 199610.Douglas S C;Cichocki A Adaptive step size techniques for decorrelation and blind source separation[外文会议] 199811.Amari S Natural gradient works efficiently in learning 199812.Cardoso J F;Laheld B Equivariant adaptive source separation[外文期刊] 199613.Cardoso J F Blind signal separation:Statistical principles[外文期刊] 1998(10)14.Cao X R;Liu R W A general approach to blind source separation 199615.Burel G Blind separation of .sources:A nonlinear neural algorithm[外文期刊] 199216.Bell A J;Sejnowski T J An information-maximization approach to blind separation and blind deconvolution 199517.张贤达;保铮通信信号处理 200018.Yang H H;Amari S Adaptive on-line learning algorithms for blind separation:Maximum entropy and minimum mutual information[外文期刊] 199719.Yang H H Series updating rule for blind separation derived from the method of scoring 199920.Yang B Projection approximation subspace tracking[外文期刊] 199521.Weingessel A;Homik K Local PCA algorithms[外文期刊] 200022.Veen A J;Talvar S;Paulraj A A subspace approach to btind spacetime signal processing for wireless communication systems[外文期刊] 1997(1)23.Torkkola K Blind separation for audio signalAre we there yet 199924.Tong L;Liu R W;Soon V C Indeterminacy and identifiability of blind identification 199125.Tan Y;Wang J;Zurada J M Nonlinear blind source separation using a radial basis function network [外文期刊] 200126.Taleb A;Jutten C;Olympieff S Source separation in post nonlinear mixtures:An entropy-based algorithm 199827.Papadias C B;Paulraj A A constant modulus algorithm for multi-user signal separation in presence of delay spread using antenna arrays[外文期刊] 1997(6)28.Pajunen P;Hyvarinen A;Karhunen J Nonlinear blind source separationby self-organizing maps 199629.Oja E The nonlinear PCA learning rule in independent component analysis 199730.Oja E Principal components,minor components,and linear neural networks[外文期刊] 199231.Amari S;Cichocki A Adaptive blind signal processing:Neural network approaches[外文期刊] 1998(10)32.Lin J K;Grier D G;Cowan J D Source separation and density estimation by faithful equivariant SOM [外文会议] 199733.Li Y;Wang J;Zurada J M Blind extraction of singularly mixed souree signals 200034.Leen T W;Girolami M;Sejnowski T J Independent component analysis using an extended infomax algorithm for mixed sub-Gaussian and superGaussian sources 1999INIOTIS D G;Papaparaskeva A partitional adaptive approach to nonlinear channel equalization[外文期刊] 1998(10)36.Kundur D;Hatzmakos D A novel blind deconvolution scheme for image restoration using recursive filtering[外文期刊] 199837.Karhunen J;Pajunen P;Oja E The nonlinear PCA criterion in blind source separation:Relations with other approaches[外文期刊] 199838.Karhunen J;Oja E;Wang L;Eigario R,Joutsensalo J A class of neural networks for independent component analysis[外文期刊] 199739.Karhunen J;Pajunen P Blind source separation using least-squares type adaptive algorithms[外文会议] 199740.Karhunen J;Joutsensalo J Representation and separation of signals using nonlinear PCA type learning[外文期刊] 199441.Jutten C;Heranlt J Blind separation of sourcesPart I:An adaptive algorithm based on neuromimetic architecture[外文期刊] 199142.Hyvarinen A;Pajunen P Nonlinear independent component analysis:Existence and uniqueness results 199943.Delfosse N;Loubaton P Adaptive blind separation of independent sources:A deflation approach 199544.Deco G;Brauer W Nonlinear higher-order statistics decorrelation by volume-censerving neural architectures 199545.Cruces-Alvarez S;Cichocki A;Castedo-Ribas L An iterative inversion approach to blind sourceon P Independent component analysis,a new concept? 199447.Cichocki A;Amari S I;Adachi M;Kasprzak W Self-adaptive neural networks for blind separation of sources[外文会议] 199648.Cichocki A;Unbehanen R;Rummert E Robust learning algorithm for blind separation of signals[外文期刊] 1994(17)49.Cichocki A;Unbehauen R;Moszczynski R;Rummaert E A new on-line adaptive learning algorithm for blind separation of source signals 199450.Choi S;Cichocki A;Amari S I Flexible independent component analysis[外文期刊] 200051.Cardoso J F;Souloumiac A Jacobi angles for simultaneous diagonalization[外文期刊] 199652.Cardoso J F;Souloumiac A Blind beamfomrming for non-Gaussian signals[外文期刊] 1993(6)1.夏淑芳.张天骐.李雪松.高清山基于变步长的混合图像自适应盲分离算法[期刊论文]-数据采集与处理 2011(2)2.杨晓梅基于稀疏表征的欠定盲源分离的研究[期刊论文]-科技信息 2011(1)3.林航.汤俊杰.谢玉川.钟臣一种带噪卫星测控信号分离算法[期刊论文]-指挥信息系统与技术 2010(6)4.邬诚变学习速率在线ICA算法在雷达信号分选中的应用[期刊论文]-现代雷达 2010(1)5.高辉.吴筱敏.高忠权.孟祥文.李根.黄佐华盲源分离法在火花塞离子电流信号分离中的应用[期刊论文]-西安交通大学学报 2010(1)6.李佰.刘辉粒子群算法用于盲信号抽取的研究[期刊论文]-电子器件 2010(1)7.艾延廷.费成巍.张凤玲.刘秀芳.孙晓倩.张宬ICA在航空发动机振动信号盲源分离中的应用[期刊论文]-振动、测试与诊断 2010(6)8.程娇.王晓凯.李锋独立分量分析可调速率相对梯度算法[期刊论文]-信息与电子工程 2010(2)9.张少刚基于数据挖掘的信号分离技术研究[期刊论文]-自动化与仪器仪表 2010(4)10.刘辉.李佰粒子群算法用于盲信号分离的研究[期刊论文]-现代电子技术 2010(17)11.夏淑芳.张天骐.苗圃.李雪松两种图像盲分离算法性能的比较与仿真[期刊论文]-电视技术 2010(4)12.吴亮.陈宗海基于独立分量分析的运动目标检测[期刊论文]-中国科学技术大学学报 2010(8)13.崔志涛.蹇科一种改进的自适应步长盲源分离算法[期刊论文]-科技信息 2010(18)14.徐欢基于独立分量分析的瞬时混合语音信号盲分离算法研究[期刊论文]-科技情报开发与经济 2010(11)15.王纪伟.高宝成一种盲信号分离算法的改进研究[期刊论文]-计算机与现代化 2010(2)16.廖旭晖一种基于负熵的快速定点盲源分离方法及其试验验证[期刊论文]-常州工学院学报 2010(1)17.朱艳萍.宋耀良.徐茂格一种基于改进粒子滤波算法的盲信号分离[期刊论文]-火力与指挥控制 2010(9)18.胡津津基于独立分量分析的瞬时混合语音信号盲分离算法研究[期刊论文]-安徽电子信息职业技术学院学报2010(4)19.周文.侯进勇基于独立分量分析的混沌信号盲分离[期刊论文]-现代电子技术 2009(21)20.余华.吴文全盲源信号分离EASI算法研究与改进[期刊论文]-计算技术与自动化 2009(4)21.余华.吴文全.刘忠自适应步长EASI算法研究及改进[期刊论文]-舰船电子工程 2009(10)22.乐剑.陈蓓一种雷达信号全脉冲数据特征提取方法[期刊论文]-中国电子科学研究院学报 2009(3)23.彭耿.黄知涛.姜文利.周一宇单通道盲信号分离研究进展与展望[期刊论文]-中国电子科学研究院学报 2009(3)24.由科军.冶继民超定盲信号分离RLS算法研究[期刊论文]-电子科技 2009(9)25.柴晓冬.尧辉明.柴靓采用基于独立分量分析的盲信号分离方法分离出FSK信号中的干扰信号[期刊论文]-中国铁道科学 2009(4)26.周治宇.陈豪空间同频电子侦察信号的盲分离[期刊论文]-中国空间科学技术 2009(3)27.杜元军.高勇ADS与Matlab协同仿真下的盲源分离实现[期刊论文]-现代电子技术 2009(13)28.王凌燕.侯文基于最大信噪比的潜艇振动信号盲分离算法[期刊论文]-机械工程与自动化 2009(3)29.吴建宁.王珏用独立分量分析去除大鼠模型脑电信号中的心电干扰[期刊论文]-测试技术学报 2009(4)30.邓兵.陶然.尹德强分数阶傅里叶域的阵列信号盲分离方法[期刊论文]-兵工学报 2009(11)31.马建仓.石庆斌.程存虎.赵述元航空发动机转子振动信号的分离测试技术[期刊论文]-振动、测试与诊断2009(1)32.杨杰.张家树基于聚类分析的实时语音信源数估计方法[期刊论文]-数据采集与处理 2009(2)33.王凌燕.黄公亮.侯文基于独立分量分析的潜艇振动信号盲分离技术[期刊论文]-计量与测试技术 2009(3)34.田其冲.郑卫国.孙大雷独立分量分析及其应用[期刊论文]-电脑知识与技术 2009(18)35.杨志聪基于矩阵对角化的盲源分离算法研究[期刊论文]-电脑知识与技术 2009(12)36.田其冲.郑卫国.孙大雷基于FastICA的语音分离与图像分离[期刊论文]-电脑编程技巧与维护 2009(16)37.贾崟.武俊义基于多层非负矩阵分解的工频干扰消除[期刊论文]-电力科学与工程 2009(4)38.周治宇.陈豪基于盲信号分离的同频信号的串行分离技术[期刊论文]-信息与电子工程 2009(4)39.王昆盲源分离问题的分析研究[期刊论文]-科技信息 2008(29)40.余华.刘忠盲信号分离的自适应算法研究[期刊论文]-计算技术与自动化 2008(4)41.杨柳绿.高勇基于训练序列的峭度盲源分离算法[期刊论文]-信息与电子工程 2008(5)42.邱天爽.毕晓辉稀疏分量分析在欠定盲源分离问题中的研究进展及应用[期刊论文]-信号处理 2008(6)43.刘秀芳.艾延廷.张宬基于独立分量分析的航空发动机振动信号盲源分离[期刊论文]-沈阳航空工业学院学报2008(5)44.李雪松.张天骐.杨柳飞.代少升基于图像单调特性的盲源分离算法[期刊论文]-计算机应用研究 2008(12)45.袁幸.段志善转子-轴承系统故障非线性振动响应识别方法的研究[期刊论文]-机床与液压 2008(4)46.余华.刘忠盲信号分离的自适应算法研究及比较[期刊论文]-化学工程与装备 2008(9)47.张玲基于PAST盲信号分离方法的仿真结果分析[期刊论文]-微计算机信息 2008(13)48.龚勤.张晓林.杨晓冬试论独立分量分析在民航空管系统上的应用[期刊论文]-科技创新导报 2008(5)49.马晓红.孙长富基于盲源分离的小波域多重音频水印方法[期刊论文]-电子与信息学报 2008(10)50.梁端丹.韩政.郝家甲独立分量分析及其应用研究[期刊论文]-现代电子技术 2008(3)51.聂凌烨.李雷.陈嘉明基于网络分量分析的盲源分离方法[期刊论文]-计算机应用 2008(z1)52.柴晓冬.叶学义.庄镇泉基于虹膜生物特征信息的图像加密方法[期刊论文]-计算机应用 2008(z1)53.纪祥鲲.王胜兵.艾小川线性独立成分分析的统一算法[期刊论文]-计算机与数字工程 2008(8)54.张晋东.秦贵和.崔月基于FastICA和SVM的EEG信号分类系统[期刊论文]-计算机研究与发展 2008(z1)55.徐桂芳.高勇基于训练序列的盲分离算法性能分析[期刊论文]-舰船电子工程 2008(8)56.石庆斌.马建仓盲源分离在机械振动信号分析中的应用[期刊论文]-测控技术 2008(5)57.骆鹿.樊可清快速定点独立分量分析在盲源分离中的应用[期刊论文]-科技信息(科学·教研) 2008(2)59.刘建强.冯大政.周祎基于多信道信号增强的卷积混迭语音信号盲分离的后处理方法[期刊论文]-电子学报2007(12)60.刘洋.邱天爽.毕晓辉基于独立分量分析和遗传算法的诱发电位提取新方法[期刊论文]-中国生物医学工程学报2007(3)61.孙云莲.罗卫华.李洪基于EMD的ICA方法在电力载波通信信号提取中的应用[期刊论文]-中国电机工程学报2007(16)62.李洪.孙云莲独立分量分析在多参数控制信号分离中的应用[期刊论文]-现代电子技术 2007(2)63.刘新艳.毋丹芳.李维勤基于NPCA的盲源分离算法[期刊论文]-无线电通信技术 2007(2)64.袁幸.段志善机械耦合振动响应双谱解卷积识别方法[期刊论文]-煤矿机械 2007(12)65.周成鹏.柴晓冬一种新的复波恢复方法——FastICA[期刊论文]-计算机工程与应用 2007(27)66.王宏志.贺利良.宋春鹏基于高阶统计量的MIMO系统辨识与信号分离算法[期刊论文]-吉林大学学报(工学版)2007(6)67.赵振兵.苑津莎.高强.罗广孝盲源分离在电气设备温度在线监测中的应用[期刊论文]-高电压技术 2007(8)68.邓智恒.罗伟栋实际语音盲分离客观评价指标研究[期刊论文]-电声技术 2007(9)69.曾燕芳.阔永红高噪声环境下的麦克风阵列VAD算法研究[期刊论文]-噪声与振动控制 2007(2)70.吴军彪.陶国良.陈进基于时频独立分量分析的Wigner-Ville分布交叉项消除法[期刊论文]-浙江大学学报(工学版) 2007(2)71.邬诚.李宏伟.王国庆一种新的变学习速率自适应独立分量分析算法[期刊论文]-计算机工程与应用 2007(1)72.王咏平.高俊未知数量稀疏源盲分离的一种新方法[期刊论文]-海军工程大学学报 2007(1)73.Xiaodong Chai.Chengpeng Zhou.Zhaoyan Feng.Yinhua Wang.Yansheng Zuo Complex-wave retrieval based on blind signal separation[期刊论文]-中国光学快报(英文版) 2006(1)74.蒋淑敏.宋瑞.高洪涛.胡育筑非负矩阵因子分解算法解析手性药物重叠峰的HPLC-DAD数据[期刊论文]-中国药科大学学报 2006(5)75.柴晓冬.左言胜基于盲信号分离的复波恢复方法[期刊论文]-仪器仪表学报 2006(10)76.蒋淑敏.戴冬梅.胡育筑手性药物联用色谱重叠峰的定性定量解析方法[期刊论文]-计算机与应用化学 2006(12)77.蒋淑敏.戴冬梅.胡育筑手性药物联用色谱重叠峰的定性定量解析方法[期刊论文]-计算机与应用化学 2006(12)78.刘海林基于广义特征值的病态混叠盲源分离算法[期刊论文]-电子学报 2006(11)79.王乐.顾学迈.王永建基于小波分析的卫星测控信号盲识别算法[期刊论文]-南京理工大学学报(自然科学版)2006(6)80.张永祥.明廷涛.王洪磊基于最小互信息准则的盲源分离在齿轮箱故障诊断中的应用[期刊论文]-机械设计与制造 2006(6)81.冶继民.张贤达.金海红超定盲信号分离的半参数统计方法[期刊论文]-电波科学学报 2006(3)82.李婷.邱天爽基于带参考信号的ICA算法的脑电信号眨眼伪差的分离研究[期刊论文]-中国生物医学工程学报2006(3)83.孙洪.安黄彬一种基于盲源分离的雷达信号分选方法[期刊论文]-现代雷达 2006(3)84.冶继民.金海红.楼顺天.张贤达未知源信号数目投影自然梯度盲信号分离算法[期刊论文]-西安电子科技大学学报(自然科学版) 2006(2)85.魏广芬.唐祯安.陈正豪.余隽.王立鼎.闫桂珍气体传感器阵列信号的盲分离研究[期刊论文]-高等学校化学学报2006(1)86.吴景田实时语音盲信号分离及声源定位研究[学位论文]硕士 200687.马晓红传声器阵列语音增强中关键技术的研究[学位论文]博士 200688.冯丹凤基于盲源分离技术的多目标辨识与定向技术研究[学位论文]硕士 200689.陈楚楚非协作情况下直接扩频信号扩频码盲估计研究[学位论文]硕士 200690.张旭秀盲源分离及其在脑电信号处理中应用的研究[学位论文]博士 200691.张华非平稳宽带有色信号盲卷积分离算法的研究[学位论文]硕士 200692.武建盲源分离研究及其在脑电信号分析中的应用[学位论文]硕士 200693.刘瑾基于盲源分离的语音增强方法研究[学位论文]硕士 200694.张鹏基于信号处理的压电自感知执行器解耦研究[学位论文]硕士 200695.丁小燕基于盲源分离的数字音频水印技术研究[学位论文]硕士 200696.梁忠杰数字音频鲁棒水印技术的研究[学位论文]硕士 200697.黄灿辉独立分量分析在盲信号分离中的应用研究[学位论文]硕士 200698.王毅.牛奕龙.陈海洋独立分量分析的基本问题与研究进展[期刊论文]-计算机工程与应用 2005(27)99.冯兆艳.柴晓冬基于密钥图像的加密技术[期刊论文]-信息安全与通信保密 2005(10)100.李小军.楼顺天.张贤达基于估计概率密度函数的独立分量分析方法[期刊论文]-西安电子科技大学学报(自然科学版) 2005(4)101.李舜酩转子振动故障信号的盲分离[期刊论文]-航空动力学报 2005(5)102.张梅.金施群ICA不确定性问题在圆度误差分离中的解决措施[期刊论文]-工具技术 2005(10)103.李舜酩.杨涛盲源信号分离及其发展[期刊论文]-传感器技术 2005(4)104.季忠.金涛.杨炯明.秦树人虚拟噪声通道在基于ICA消噪过程中的应用[期刊论文]-中国机械工程 2005(4) 105.李舜酩机械振动信号盲源分离的时域方法[期刊论文]-应用力学学报 2005(4)106.李小军.楼顺天.张贤达一种针对复值信号的独立分量分析方法[期刊论文]-西安电子科技大学学报(自然科学版) 2005(3)107.刘伟群.阳春华.易叶青JADE盲信号分离算法失效问题研究[期刊论文]-计算技术与自动化 2005(3)108.李婷.邱天爽.牛杰独立分量分析在脑电信号处理中的应用及研究进展[期刊论文]-北京生物医学工程 2005(3) 109.申永军齿轮系统的非线性动力学与故障诊断研究[学位论文]博士 2005110.梅铁民盲源信号分离时域与频域算法研究[学位论文]博士 2005111.林秋华基于盲源分离的图像与语音加密新方法研究[学位论文]博士 2005112.沈希忠盲信号处理的二阶和高阶分离算法研究[学位论文]博士 2005113.孔薇盲分离算法的研究及其在声信号处理中的应用[学位论文]博士 2005114.钟振茂基于盲源分离的声学故障特征提取方法研究[学位论文]博士 2005115.何培宇面向实时处理的自适应声学语音信号处理[学位论文]博士 2005116.易叶青基于进化算法的盲源信号分离[学位论文]硕士 2005117.谢朋朋车辆声频信号盲分离及遗传算法研究[学位论文]硕士 2005118.魏广芬基于微热板式气体传感器的混合气体检测及分析[学位论文]博士 2005119.刘洋独立分量分析及其在脑电信号提取中的应用[学位论文]硕士 2005120.王瑜空域处理语音增强方法的研究[学位论文]硕士 2005121.王翀数字音频水印技术的研究[学位论文]硕士 2005122.冶继民信源数目未知与变化时的盲信号分离方法研究[学位论文]博士 2005123.王明祥独立分量分析方法及在图像处理中的应用研究[学位论文]博士 2005124.张兼OFDM系统盲信道估计算法研究[学位论文]硕士 2005125.梁越卷积混合信号盲分离技术及应用研究[学位论文]硕士 2005126.孙战里盲源分离算法及其应用的研究[学位论文]博士 2005127.孙战里盲源分离算法及其应用的研究[学位论文]博士 2005128.施剑基于麦克风阵列的盲语音分离算法研究[学位论文]博士 2005129.康浔盲信号分离及在非线性混叠环境下的分离问题[学位论文]硕士 2005130.李先伟混叠卷积信号的盲反卷积研究[学位论文]硕士 2005131.梁丽丽麦克风阵列语音增强方法的研究[学位论文]硕士 2005132.潘丽丽独立成分分析及其在脑功能磁共振成像中的应用[学位论文]硕士 2005133.牛奕龙盲源分离算法研究[学位论文]硕士 2005134.李维勤基于混合信号概率密度函数估计的盲信号分离[学位论文]硕士 2005135.李婷独立分量分析及其在脑电信号噪声分离中的应用[学位论文]硕士 2005136.张荣强说话人识别中特征提取的方法研究[学位论文]硕士 2005137.何昭水瞬时线性混叠与卷积混叠盲分离研究[学位论文]博士 2005138.张发启盲源信号分离技术[期刊论文]-现代电子技术 2004(20)139.郑华.张粒子.谢莉.申景娜关于电力市场下系统边际价格概率模型的新研究[期刊论文]-中国电机工程学报2004(11)140.解静.李艳斌.陈建峰最大似然估计的独立分量分析法[期刊论文]-无线电工程 2004(7)141.李广彪.张剑云.毛云祥盲源分离的发展及研究现状[期刊论文]-航天电子对抗 2004(6)142.张梅.金施群圆度测量误差分离技术解决ICA的不确定性问题[期刊论文]-黑龙江科技学院学报 2004(5) 143.曹薇.高隽.胡学友一种有效的图像盲水印算法[期刊论文]-合肥工业大学学报(自然科学版) 2004(8) 144.刘盛鹏.方勇基于TMS320C64x DSP/BlOS Ⅱ的嵌入式语音采集与盲分离系统设计[期刊论文]-电子技术2004(4)145.李小军.朱孝龙.张贤达盲信号分离研究分类与展望[期刊论文]-西安电子科技大学学报(自然科学版) 2004(3) 146.张玲.何培宇.刘开文关于噪声环境中语音信号盲分离的研究[期刊论文]-四川大学学报(自然科学版) 2004(1) 147.申永军.杨绍普.刘献栋齿轮故障诊断中的信号处理技术研究与展望[期刊论文]-机械传动 2004(3)148.魏乐基于非负矩阵分解算法进行盲信号分离[期刊论文]-电光与控制 2004(2)149.张明键盲分离算法的研究[学位论文]博士 2004150.岳顺民配电市场价格分析及价格监管模式研究[学位论文]博士 2004151.王进军基于小波分析和人工神经网络对ECG信号滤噪和分类的研究[学位论文]硕士 2004152.李小军独立分量分析及其在阵列信号处理中的应用[学位论文]博士 2004153.李合生.赵明生.吴佑寿基于小波与傅里叶变换的一类短波盲信道自适应分离算法研究[期刊论文]-信息与控制 2003(4)154.金海红.冶继民未知源信号个数盲信号分离的半参数统计方法[期刊论文]-西安电子科技大学学报(自然科学版) 2003(6)155.李合生.赵明生.吴佑寿一种短波通信盲信道快速分离算法[期刊论文]-通信学报 2003(7) 156.刘洋.樊可清二阶盲辨识(SOBI)在模态分析中的应用[期刊论文]-微计算机信息 2010(31)本文链接:/Periodical_dianzixb2001z1010.aspx。
语音信号分离技术研究随着人工智能技术的不断发展和应用,语音信号处理技术也得到了迅速发展。
其中,语音信号分离技术更是备受瞩目。
随着语音信号处理技术的不断发展,语音信号分离技术在各个领域得到越来越广泛的应用。
本文将探讨语音信号分离技术的发展现状、研究方向和应用前景。
一、语音信号分离技术的发展现状语音信号分离技术是指将混合在一起的多个语音信号分离并恢复出单独的每一个信号。
在语音信号分离技术出现之前,大多数语音信号处理技术都是基于对单一语音信号的处理。
而随着语音信号分离技术的出现,我们能够在混音语音中分离出不同的语音信号,从而实现一些以前难以想象的任务。
目前,语音信号分离技术已经被广泛应用在很多领域,比如语音识别、听力康复、通信和娱乐等。
在语音识别方面,语音信号分离技术的应用可以提高识别率。
而在听力康复方面,可以在嘈杂环境中帮助听力受损者更好地听到说话者的声音。
在通信和娱乐方面,语音信号分离技术可以更好地实现语音和视频通信,以及音乐和视频制作等。
二、语音信号分离技术的研究方向目前,语音信号分离技术主要有线性盲源分离(Linear Blind Source Separation, LBSS)和非线性盲源分离(Nonlinear Blind Source Separation, NLBSS)两种方法。
LBSS是指从混叠的信号中提取独立的源信号,而不需要任何有关源信号的先验信息。
LBSS算法通常基于统计学理论(如独立成分分析、主成分分析等),对混叠信号进行分离。
这种方法在实际应用中被证明是有效的,但是在源信号高度相关时会存在困难。
NLBSS则是指在混叠信号中进行非线性分析,通过对信号的高阶统计特性进行建模,实现对独立源信号的提取。
NLBSS方法通常使用基于神经网络或非线性统计模型的算法。
与LBSS相比,NLBSS方法可以有效地处理信号的高度相关性,但是需要较多的计算资源。
现在,基于深度学习的语音信号分离技术正在不断发展,成为语音信号分离技术的热点研究方向。
An Approach for Nonlinear Blind Source Separation of Signals with Noise Using Neural Networks and Higher-Order CumulantsNuo Zhang,Xiaowei Zhang,Jianming Lu and Takashi YahagiGraduate School of Science and TechnologyChiba University,Chiba-shi,263-8522JapanEmail:zhang@graduate.chiba-u.jpAbstract—In this paper,we propose a robust approach for blind source separation when observations are contaminated with Gaussian noise and nonlinear distortion.A radial basis function networks(RBFN)is employed to estimate the inverse of nonlinear mixing matrx.We utilize an novel cost function which consists of mutual information and higher-order cumulants of signals. Compared with moments,higher-order cumulants can provide a clearer form and more information of signals.Thus the proposed method has not only the capacity of recovering the nonlinearly mixed signals,but also removing high-level Gaussian noise from transmitted signals.Through the simulation and analysis of artificially synthesized signals,we illustrate the efficacy of this proposed approach.I.I NTRODUCTIONBlind source separation is a class of methods that recover unaccessible independent original signals from nonlinear mix-tures.It has received a great deal of attention recently.So far there are a number of algorithms have been proposed for blind source separation problem and have been applied to the area of speech signal processing,communication and medical signal processing(MEG,ECG and EEG).The main study of blind source separation considers about the instantaneous linear mixture by now[1].Nonlinear blind source separation is a much more recent research topic.Gen-erally,a nonlinear mixing model is more realistic and accurate than linear model and suitable for many practical situations. Moreover,there are also many Gaussian noise are present in the real-world and it has been studied within traditional independent component analysis method by researchers[2].It is of important significance to address the topic of nonlinear blind separation with noise in depth.However blind separation of signals in nonlinear mixtures has been very sparsely studied because of non-uniqueness of separation.Neural networks are valuable tools to deal with a variety of nonlinear problems occurred in many practical domains and have been studied extensively.Hence several blind source separation methods have been derived from neural networks naturally.This can be found in following litera-tures:Kohonen’s self-organizing map(SOM)is utilized to extract sources from nonlinear mixture(but the computational complexity of this kind of approaches grow exponentially) [3],an information BP algorithm for training of separating system[4].One drawback is slow convergence rate yield from the highly nonlinear relationship between the output and learning weights of the network.In contrast,an RBF network has several advantages over a multilayer perceptron in terms of natural learning manner.An approach using RBF network proposed by Tan et al.[5]can recover source signals better,and the convergence rate is very fast.However,this method degrades greatly when high-level noise is present with nonlinear distortion.The disadvantage of this approach is that it does not consider the influence of noise in the cost function. The purpose of this paper is to utilize higher-order statistics to obtain a robust method with RBF network when both nonlinear distortion and Gaussian noise are present.Note that,cumulants usually present in a clear form,where more information can be provided by higher-order cumulants.In addition,we used some useful properties of cumulants,which can not be provided by moments.(They will be described in the following sections).We introduced three-and four-order cumulants together with mutual information into a novel cost function.The minimization of the proposed cost function can recover source signals as good as original method. Furthermore,it outperforms the original method when the high-level Gaussian noise is present in nonlinear mixtures. Thus,the proposed approach can also be used to reduce noise from data.According to the theoretical analysis and computer simulation results,it is shown that the proposed algorithm has better performance than original approach when facing with Gaussian noise in nonlinear mixtures.II.N ONLINEAR M IXING S YSTEM M ODEL WITH N OISE The noise introduced in this paper is Gaussian noise.Denoted x(t)=[x1(t),x2(t),···x n(t)]T the vector of observed random variables,and s(t)= [s1(t),s2(t),···s n(t)]T the independent source vector, the noisy nonlinear mixing model can be expressed asx(t)=f[s(t)+G](1)where G is Gaussian noise,f is an unknown multiple-input and multiple-output(MIMO)mapping function.We further assume the noise is independent with the source components.For simplicity and convenience,we also assume the dimension between original signals and mixtures is equal to each other,without the loss of generality.Fig.1.System model.Figure.1.shows the mixing model and blind source sep-aration system expressed in Eq.(1).The purpose of the separation system is to recover the original signals only from their observations x (t )and remove the Gaussian noise.Obviously,this problem is difficult to solve.Hence we assume f (·)is componentwise invertible and its inverse f −1(·)= f −11(·),f −12(·),···,f −1n (·) exists.This paper will employ an RBF network to solve the problem due to its capacity of approximating an arbitrary function [6].It is also difficult that how to recover the mixing matrix when the mixtures are degraded by Gaussian noise.A class of noise reduction method based on higher-order cumulants has been proposed [7].Considered the better properties of higher-order cumulants,we applied it into the cost function in this paper,what will be introduced in the following sections.III.R ADIAL B ASIS F UNCTION N ETWORKSThe RBF network is very different from multilayer percep-trons with sigmoidal activation function in which it utilizes neurons with radial basis function that are locally responsive to input stimulus in the hidden layer.Fig.2.The RBF network.As shown in Fig.2.,an RBF neural network consists of two layers:hidden layer and output layer with linear neurons.A hidden RBF neuron is usually implemented using a Gaussian kernel function,which has two parameters,what are center,and width.The whole relationship between input and output of an RBF neural network can be given by:y (x )=M j =0w j φj (x ),(2)andφj (x )=exp−1σ2j(x −µj )2(3)where w j is the connecting weight between the j -th hidden RBF neuron and linear output neuron,φj (x )is the response function of the j -th hidden RBF neuron,µj and φj are the center and width of the j -th hidden RBF neuron,respectively,M indicates the number of hidden RBF neurons in the network.There are a number of learning strategies for RBF networks such as randomly selecting the radial basis function centers and employing unsupervised or supervised procedures for selecting the radial basis function centers.In this paper we will use the RBF network to construct the inverse mapping of the nonlinear mixing function in Fig.1.In the following,we will focus on the learning rules of the parameters in our proposed algorithm.IV.T HE P ROPOSED C OST F UNCTION AND H IGHER -O RDERC UMULANT Recently,a few unsupervised RBF based methods have been proposed [5].They make use of the characteristics of neural networks to solve the nonlinear mixing problem.This paper will study the problem by adopting the higher-order cumulants in the cost function when high-level noise is present.In the original method,the higher-order moments of sources and estimations are required to be the same,which yields the drawback that the moment information of sources have to be known,which is too rigorous in real world.The proposed higher-order cumulants based method is not constrained to the severe condition.Accordingly,the proposed algorithm can separate more general kinds of signals with high-level noise and without prior information of source signals.Thus it is more flexible and applicable.A.The Definition of Cumulant and its PropertiesSince the datum we used in this paper are assumed to be zero mean random vector,the definition of cumulants up to order four,can be given as followscum i (y )=0,cum ij (y )=E {y i y j },cum ijk (y )=E {y i y j y k },cum ijkl (y )=E {y i y j y k y l }−E {y i y j }E {y k y l }−E {y i y k }E {y j y l }−E {y i y l }E {y j y k }.(4)In addition,cumulants have two properties. 1.If thesets of random variables (x 1,x 2,...,x n )and (y 1,y 2,...,y n )are independent,then cum {x 1+y 1,x 2+y 2,···x n +y n }=cum {x 1,x 2,···x n }+cum {y 1,y 2,···y n }.2.If the sets of random variables (x 1,x 2,...,x n )are Gaussian distributed,then cum {x 1,x 2,···x n }=0.Obviously,utilized the conclusion the two properties above to the output of the separation system,we can draw the following relation in theory:cum (y +G )=cum (y ).(5)Therefore,by applying such property of higher-order cumu-lants into our cost function to estimate RBF networks’param-eters,we can conclude that there is no effect on estimations from Gaussian noise,such that better separation results can be obtained.Higher-order cumulants also have another property that it is helpful to measure the dependence among random vector. For example,four-order cumulant is a four-dimensional array whose entries are given by the fourth-order cross-cumulants of the data.This can be considered as a four-dimensional matrix,since it has four different indices.The four-order cumulants contain all the fourth-order information of the data. The diagonal elements characterize the distribution of single components.The off-diagonal elements of cross-cumulants(all cumulants with ijkl=iiii)characterize the statistical depen-dencies among components.If and only if,all components are statistically independent,the off-diagonal elements vanish,and the cumulants(of all orders)are diagonal.In the next section, we will derive the cost function with higher-order cumulants.B.Cost Function with Higher-Order CumulantsThe general measure for the degree of the dependence among random variables is mutual information.Several mutual information based algorithms have been presented for linear mixtures separation in literature.However it is well known that only mutual information itself is not sufficient to recover source signals from nonlinear mixtures.In order to solve this problem we introduce higher-order cumulants into the cost function,by minimizing which we can make the output of separation system as independent as possible.Moreover, utilizing the properties described above,this approach is robust to high-level Gaussian noise.Three-and four-order cumulants are introduced into our cost function,in order to recover the original signals from their nonlinear mixtures when Gaussian noise is present,given as follows:J(W)=I(y)+ijk=iiicum ijk(y)2+ijkl=iiiicum ijkl(y)2(6)where I(y)is mutual information of the separation system out-puts,ijk=iii cum ijk(y)2andijkl=iiiicum ijkl(y)2are squaresum of non-diagonal three-order and four-order cumulant of outputs,respectively.Mutual information I(y)in Eq.(6)is defined asI(y)=ni=1H(y i)−H(y)(7)where H(y)is joint entropy of random vector y;and H(y i) is entropy of random variable y i,the i-th component of vector y.This cost function consists of two parts:mutual information and square of cumulants of system outputs.Note that,these two items are always nonnegative,and vanish when both of them are minimized.Therefore,the proposed algorithm can recover original signals through minimizing the cost function.The proposed algorithm contains two steps:first,it uses traditional k-means algorithm for the selection of the centers and widths of RBF neurons;then updates the weights in output layer of neural network by utilizing gradient decent of cost function.After estimating the centers and widths of hidden layer using k-means algorithm,the weights of output layer can be updated by minimizing the cost function.In this paper,we utilize gradient descent manner to obtain the unsupervised learning rule of output layer weights of RBF network.The gradient of the cost function of Eq.(6)with respect to output layer weights is shown as follows:∂J(w)∂w=∂I(y)∂w+ijk=iii2cum ijk(y)∂cum ijk(y)∂y∂y∂w+ijkl=iiii2cum ijkl(y)∂cum ijkl(y)∂y∂y∂w.(8)By using the Gram-Charlier expansion method to approxi-mate the pdf of y i,we obtain the gradient decent of mutual information as follows∂I(y)∂w=ni=1∂H(y i)∂y i∂y i∂w−∂g(x,w)∂w−1∂∂w∂g(x,w)∂w.(9)From definitions in(4)and(8),we can simply derive gradient decent of the two latter items in Eq.(6).From Eq.(4),(6)-(9)wefinally derive the updating rule of output layer weights:w k+1=w k−λ∂J(w k)∂w k(10) where0<λ<1is learning rate.C.Description of Learning RuleIn this paper,the simulation results are analyzed based on the root-mean-squares(RMS)error,which also is employed to stop iteration of the unsupervised learning:RMSE=1NNt=1(s(t)−ˆs(t))2(11)whereˆs is recovered source signals,N denotes the number of signals.The proposed algorithm can be summarized as follows: a.Initialize the centers,widths and weights of RBFnetwork by using small random number.The ter-mination condition,a small positive numberε,oflearning is also determined.b.Update the centers and widths of RBF network’shidden layer by using k-means algorithm.e Eq.(8),Eq.(9)and Eq.(10)to update outputlayer weights of RBF network.d.Repeat step b until the RMSE between s andˆssatisfies the condition(RMSE<ε).V.S IMULATION AND R ESULT A NALYSISIn this section we present and discuss simulation results of the proposed algorithm.The proposed algorithm is performed with artificial signals:sinusoid and modulated sinusoid signal corrupted by additive Gaussian noise.In order to compare the performance of proposed algorithm with the traditional method,we use the same RBFN with 2inputs,6hidden neurons and 2outputs,and 5000samples of signals.The mixing process is described as follows:x 1x 2 =A 2 (·)3(·)3A 1 s 1s 2 (12)where A 1=0.250.86−0.860.25 ,A 2= 0.50.9−0.90.5.We first performed a simulation without influence of noise to verify the separation ability of proposed algorithm.Since the proposed algorithm has the same good result as that of the original method,we do not show the results here for brief.Then we used mixtures obtained from Eq.(12)in the following simulation.For a more general problem,we considered sources corrupted by Gaussian noise with 10dB SNR (Eq.(1)).Fig.3.shows two source signals.They are nonlinearly mixed with noise into mixtures,illustrated in Fig.4.From Fig.5.and Fig.6.,we can see that the proposed algorithm can obtain better separation results compared with originalmethod.Fig.3.Sourcesignals.Fig.4.Mixtures.Furthermore,we used RMSE to analyze the separation results of the proposed and original algorithms.The RMSE curves is shown in Figure.7.The proposed method has the same convergence rate with original method.Although both of them are converged at 1000iterations,we obtained better estimations from the proposed algorithm due to the robustness to Gaussian noise of higher-order cumulants.VI.C ONCLUSION A nonlinear separation method using higher-order cumu-lants,which is robust to noise,has been proposed in this paper.The comparisons of performance of the proposed and originalalgorithm has been studied in the presence of strongly nonlin-ear distortions and Gaussian noise.Simulation results showed that the proposed algorithm can obtain clearer estimations of sources corrupted by noise from nonlinear mixtures.By using higher-order cumulants,the proposed algorithm can separate mixtures without the requirement of prior infor-mation of sources.Therefore,it is more applicable to high-level noise in real problems.Fig.5.Separation results by originalmethod.R EFERENCES[1] C.Jutten and J.Herault,“Blind separation of sources,part I:An adap-tive algorithm based on neuromimetic architecture,”Signal Processing ,vol.24,pp.1–20,1991.[2] A.Hyv ¨a rinen,“Independent component analysis in the presence ofgaussian noise by maximizing joint likelihood,”Neurocomputing ,vol.22,pp.49–67,1998.[3]M.Herrmann and H.H.Yang,“Perspectives and limitations of self-organizing maps in blind separation of source signals,”Progress in Neural Information Processing:Proceedings of ICONIP’96,pp.1211–1216,1996.[4]S.A.H.H.Yang and A.Cichocki,“Information backpropagation forblind separation of sources from nonlinear mixture,”Proc.IEEE ICNN,TX ,pp.2141–2146,1997.[5]J.Ying Tan;Jun Wang;Zurada,“Nonlinear blind source separation usinga radial basis function network,”Neural Networks,IEEE Transactions on ,vol.12,no.1,pp.124–134,Jan 2001.[6]W.A.Light and E.W.Cheney,“Interpolation by piecewise-linear radialbasis functions,”Journal of Approximation Theory ,vol.64,no.1,pp.38–54,Jan 1991.[7]N. C.R.S.Bhattacharya and S.Sinha,“2-D signal modelling andreconstruction using third-order cumulants,”Signal Processing ,vol.62,no.1,pp.61–72,Oct 1997.。