H. Zhang et al. / Journal of Computational and Applied Mathematics 220 (2008) 548 – 558 549
considered a more efficient one-pass IC ordering procedure under the mean square error (MSE) criterion. And in [19]
a post-procedure step which based on canonical correlation analysis and the prior information of medical signals was
added to provide a better ICs ordering. Another ordering technique of constrained ICA was firstly proposed in [11,13].
They introduced constraints into the classical ICA to order the resulted ICs and normalize the demixing matrix in the
signal separation procedure. In this approach the separation of ICs and the elimination of indeterminacy are performed
simultaneously. Meanwhile it has been shown that this constrained ICA learning algorithm converged approximately
three times faster than the two-stage approaches [13]. At present, constrained ICA has been successfully used for some
fields such as speech analysis [10], functional MRI data extracting [12–15] and so on. However, a resulting trade off
of this method is that the convergence depends on a good choice of the learning rate. A bad choice of the learning rate
can, in practice, destroy convergence. Therefore, some ways to make the learning radically fast and reliable may be
needed.
In this paper, a simple yet efficient method of constrain ICA is firstly proposed for recovering and ranking the resultant
ICs simultaneously. Different from constrained ICA model proposed in [13] which incorporated some statistical measure
into the minimizing mutual information criterion, we present a new constrained ICA model which composed of
three parts: maximum likelihood criteria as objective function, statistical measure as inequality constraint and the
normalization of demixing matrix as equality constraint. Next, we construct a new algorithm which incorporating
newFP algorithm into this constrained ICA model. The new algorithm need not be a choice of learning rate which
is cumbersome in practical application. It is the new algorithm that overcomes this drawback of the existing learning
used in [13]. It is worth mentioning that Lagrange multiplier method is adopted to provide an adaptive solution to this
problem. Finally, in order to verify the efficiency of this new algorithm, we analyze its stability of convergence and
statistical accuracy, and make comparison with the existing algorithm introduced in [13] for separation of synthesized
signals and speech signals. Experiment results show the validity of the proposed algorithm. And comparison results
indicate the new algorithm has better performance, and show that it is more simple to implement than the existing
algorithm as the new algorithm does not depend on the choice of learning rate. Meanwhile, this new algorithm is
applied for the real-world fetal ECG data, these experiment results further demonstrate the efficiency of our new
algorithm.
The next section summarizes the classical ICA method and the newFP algorithm. Section 3 introduces the technique
of constrained ICA and drives the new constrained fixed-point algorithm. Section 4 demonstrates the present technique
with experiments using synthetic signals and speech signals. And in Section 5, we apply the new technique for real-world
fetal ECG data. The final section provides discussions and conclusions.
2. Classical ICA
Suppose that there exist M independent source signals s(t) = (s
1
(t),...,s
M
(t))
T
and N observed mixtures x(t) =
(x
1
(t),...,x
N
(t))
T
of the sources signals (usually N M). A typical ICA model is
x(t) = As(t ), (1)
where A is an unknown N × M mixing matrix. The task of classical ICA is to identify an M × N demixing matrix W
such that the M output signals
u(t) = Wx(t) = WA s (t ) = PDs(t), (2)
where P ∈ R
M×M
is a permutation matrix, D ∈ R
M×M
is a diagonal scaling matrix, and u(t) = (u
1
(t),...,u
M
(t))
T
.
Consequently, the source signals are recovered up to scaling and permutation.
2.1. The Likelihood of the ICA model
From (2), the probability density function of the observations x can be expressed as [9]:
p(x)=|det W | p(u), (3)