文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > A Gauss-Newton method for blind source separation of convolutive mixtures
侵权投诉

A Gauss-Newton method for blind source separation of convolutive mixtures

In this paper we present several Gauss-Newton algorithms for blind source separation of convolutive mixtures. The algorithms can be interpreted as generalizations of two previous approaches due to Gerven-Compernolle [7] and Nguyen-Jutten [5]. Since they ar

3. SEPARATION ALGORITHM

The BSS problem can be solved relying only on the non-Gaussianity and statistical independence of the sources because,under these hypotheses, it has been shown in [2] that the sourcesare separated if and only if the components of y[n] become pair-wise independent. When this occurs, all the cross-cumulantsbetween different outputs vanish. This idea suggests that theseparating system be selected to minimize a weighted sum ofoutput cross-cumulants.

Let us denote Cxi1,yi2, ,ziq

α

1

,[

α2, ,αq

(k1, ,k(q]

1))

=Cum(xi1[n]×α1,yi2[n k1]×α2, ,ziqn k(q 1)×αq) the

(i1

, ,iq

)-order cross-cumulant of xi1

,yi2

, ,zi

q

. Similarly, let us

define

Cx,y,α

,z(1,α2, ,αq

(k1, ,k(q 1)) as the tensor whose elements at the

ixi,yi1

, ,iq

) position are C

12, ,ziqα1,α2, ,αq

(k1, ,k(q 1)).

Let us consider now the following set of cross-cumulant matricesof the output vector

Π={Cy,y

α,β[k]

: k= Lb1, ,Lb2 ; (α,β)∈ }

( is a set of n pairs of natural numbers). It is apparent that

these matrices become diagonal when y[n] is a vector ofstatistically independent signals. Therefore, we propose as acriterion to achieve source separation the joint diagonalization ofΠ. This diagonalization is equivalent to minimizing thefollowing dependence measure

Φ =∑

(α,β)∈

w(α,β)Φ(α,β)

(1)

Φ

=

∑

yy,y2(α,β)k

y,α,β

[k] Λ

α,β

[kF

(2)

being w(α,β) a set of weighting coefficients, F the Frobenius

norm and Λy,y

α,β[k] a temporal sequence of diagonal matricesdefined as Λy,y[k]=diag(diag(Cy,yα,βα,β[k]))

. The dependence

measure (1) is the weighted square distance between the set ofcross-cumulant matrices Π and its diagonal counterpart

Πy

diag={Λy,α,β[k]

: k= Lb1, ,Lb2 ; (α,β)∈ }

.3.1Minimization algorithm

Many different algorithms can be used to minimize Φ . Amongthem we choose a Gauss-Newton (GN) one since it will neverconverge to a local minimum and exhibits quadraticconvergence. In order to derive the GN iteration, let us rearrangethe variables in vector notation. If we denote vec( ) as thecorrespondence that assigns non-diagonal elements of aN×Qmatrix (with N≤Q) to a vector (N 1)Q×1 indexingfirst by columns and then by rows, we can define the followingvectors

z(α,β)

= vec(Cy,yTy,y

Tα,β[ Lb1]), ,vec(Cα,β[Lb2])

T

(3)z= T T

w(α,β)(α,β)) 1zT1, ,w(α,βn z(α,β)n

(4) b=[vec(B[ LTT

T

b1]), ,vec(B[Lb2])]

(5)

Substituting (3),(4) and (5) in (1), the dependence measure can

be written as the inner product Φ =zTZ. Denoting theJacobian matrix as J= bz, the GN iteration will consist inupdating the separation coefficients vector as

b(n+1)=b(n)+µ

(6)

where 0<µ<2 is the adaptation step and is the result ofsolving the following system of linear equations

JJT = J z

(7)

3.2Simplifications based on the Jacobian

matrix structure

The main limitation of the above algorithm is the computationalcomplexity required by the evaluation of the Jacobian matrix forarbitrary pairs (α,β), i.e., the order of the cross-cumulantinvolved in the dependence measure. Initially, every pair ofnatural numbers (α,β) could belong to . However, thealgorithm complexity can be considerably reduced if we chooseα=1 and β>1. In this case, the dependence measure isapproximately quadratic in the vicinity of the separationsolution.

Under the assumptions AS1 and AS2, for α=1, β>1 andstationary sources which are white sequences (or equivalently,when there are instantaneous mixtures) it is shown in [3] that theelements of the Jacobian matrix can be approximated by

Cy,y

j 1i,β[k]

x,y

j b≈δirδkmC 1s,β[0]

(8)

rsmwhere δij is the Kronecker operator (δij=1 if i=j, δij=0otherwise). Taking into account these approximations, the

proposed algorithm in (6) for two white sources and two sensorsreduces to the following adaptation rule

xj,yj

(1,β)C 1[0] Cyi,yj,β 1,β[k]b(n+1)[k(1,∑w

β)∈

i,j

]=b(n)i,j

[k] µ

∑ (9)

w

j,j

(1,β)(C 1,β

[0] )

2

(1,β)∈

with i,j

j≠i

=1,2; and k= Lb1, ,Lb2. Recall that iteration (9)

only requires 2n (Lb1+Lb2+1) cumulants!

上一页第2页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email