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

Figure 2. Feed-backward separation model.
For more sources and sensors, we observe from (8) that theJacobian J is a sparse matrix and that there exists efficientmethods [4] to solve linear sparse systems such as (7).
3.3Feed-backward algorithm
The algorithm (6) was obtained for the feed-forward separation
system represented in Figure 1. A different version of thisalgorithm can be obtained for the feed-backward separationstructure plotted in Figure 2. It has been shown in [3] thatchanging the sign of the adaptation coefficients [b and replacingthe feed-forward outputs yn] by the feed-backward ones s[n]the simplified adaptation rule (9) is converted into the followingone
ws i,s j(1,β)Cx1,β1,β[k]c(n+1)n
∑j,s j[0] C (10)
i,j
[k]
=c()i,j
[k]+µ
(1,β)∈
∑
wj,s j1,[0] )2
((1,β)∈
β)(Cx1,β
where i,j
j≠i
=1,2; and k= Lb1, ,Lb2. When n =1, this
expression reduces to
C si,s j,j
c(n+1)
1,β
[k]j≠i
=1,2;
i,j
[k]=
c(n)i,j
[k]+µ
;
iCxj,s j
k1,β
[0] = Lb1, ,L (11)
b2;
This feed-backward simplified algorithm allows us to establish alink with existing BSS methods. When ={(1,1)}, i.e. β=1,the algorithm (11) reduces to the Symmetric AdaptiveDecorrelation (SAD) algorithm proposed in [7]. On the otherhand, for β=3 the algorithm (11) is very similar to theNguyen-Jutten (NJ) method proposed in [5]. Moreover, ourapproach presents several advantages with respect to the NJalgorithm: it exhibits greater speed of convergence since itcorresponds to a Gauss-Newton method and it is asymptoticallystable (as shown in the following section). Finally, it isinteresting to note that, in contrast to the existing approaches[5][7], the new algorithm performs adequately for non-causalmixtures.
4.CONVERGENCE ANALYSIS
The dependence measure Φ to be minimized in our method isnot a quadratic form of the separating system coefficients andthus may contain undesired minima that will impair theconvergence of the adaptive algorithms. In the sequel we presenta convergence analysis to show that, even though these minimamay exist (depending on the chosen set ), they are not stablepoints. Since a general proof is rather involved, we will considerthe simpler case of two sources, two sensors and instantaneousmixtures. In this case, the feed-forward algorithm (9) reduces to
yb(n+1)=b(n)
Ci,yj 1,βi,j
i,j
µC,j
j,1,βj
;ij≠i
=1,2; (12)
In the sequel we will demonstrate that when β is an odd numbergreater than one, the adaptation rule (12) may exhibit severalnon-separating stationary points but the only stable stationarypoints correspond to a separating solution.
Let H=BA be the global system transfer matrix (y=H s). Thestationary points of Φ occur when the updating term in (12)vanishes, i.e., when
Cyi,yj
1,β
=hCss
iihβjiβiβj
+1+hijhjjCβ+1=0; i,j
j≠i
=1,2;
Assuming that β is chosen so that the cumulants of the sourcesCs
s
βi+1 and Cjβ+1 are non-zero, it is straightforward to show thatthe stationary points in (12) can be classified into the followinggroups:
1. His a (possibly scaled) permutation matrix.2. His singular with at least one null row.s
3. h1
ij=±η hii and hjj=η hji with η
β+=±
Cβi+1Cs
j.
β+1
The first case corresponds to the separation solution. Thesestationary points are asymptotically stable (if 0<µ<2) becausethe approximation (8) for the Jacobian matrix becomes exact andthe algorithm coincides with a Gauss-Newton minimizationmethod. This circumstance does not hold for the non-separatingstationary points. Note that in the second case, one of the outputsis zero, the denominator of the updating term in (12) vanishesand the algorithm becomes unstable. For the third case, thestability ODE analysis has been done leading to the followingcharacteristic polynomial in λ
y,yx,y,yI µ J=λ Cxi,ijjji2
2(1 µ)λ+(1 µ2
) µ2
β
2
1,1,β 1C1,1,β 1
Cxj,y
jxi,y
1,βC 1,βi
However, since β was chosen to be odd, this expression is
simplified to
I µ J=λ2 2(1 µ)λ+(1 µ2) µ2β2
Now, we can use the Anderson-Jury inequalities [1] to check ifthe roots are within the stability region. These are β2<1 and
0<µ<2
1+β
. Then, since by hypothesis β>1, we have
reached a contradiction and the algorithm will also be unstable inthe third case.
5.COMPUTER SIMULATIONS
In this section we present the results of several computerexperiments carried out to illustrate the performance of theproposed algorithms.
A Gauss-Newton method fo... 暂无评价 4页 免费 A generalization of b... Blind source separation of convolutive mixtures Shoji Makino This paper ...
Blind source separation ... 8页 免费 A Gauss-Newton method fo... 暂无...Dynamic Algorithm for Blind Separation of Convolutive Sound Mixtures Jie Liu?...
Blind source separation ... 8页 免费 A Gauss-Newton method fo... 暂无...“The coherence function in blind source separation of convolutive mixtures of...
Blind source separation ... 8页 免费 A Gauss-Newton method fo... 暂无...mixturesFrequency_domain infomax for blind separation of convolutive mixtures隐藏...
免费 A Gauss-Newton method fo... 暂无评价 4页 免费喜欢此文档的还喜欢 ... Amplitude modulation decorrelation for convolutive blind source separation The ...
A Gauss-Newton method fo... 暂无评价 4页 免费 Blind Separation of Inst...generalized to the context of the source separation of convolutive mixtures. ...
BLIND SEPARATION OF CONVOLUTIVE MIXTURES BASED ON SECOND ORDER AND THIRD ORDER STATISTICSBLIND SEPARATION OF CONVOLUTIVE MIXTURES BASED ON SECOND ORDER AND ...
A Gauss-Newton method fo... 暂无评价 4页 ... Blind source separation ... 6页 免费 Blind ...These applications deal with convolutive mixtures [...
5页 免费 A Gauss-Newton method fo... 暂无评价 4页 免费... An adaptive beamforming perspective on convolutive blind source separation The...
5页 免费 A Gauss-Newton method fo... 暂无评价 4页 免费... An adaptive beamforming perspective on convolutive blind source separation The...
热门文档