A One-Stop BSS Analysis Method of Integrated System Signals Based
on Wavelet Transform, FastICA and SOM
Hongyi Li
1,2
, Liantao Ma
1,2
, Bin Chen
2
and Di Zhao
1,*
1. LMIB, School of Mathematics and Systems Science, Beihang University, Beijing 100191, China
2. School of Software Engineering, Beihang University, Beijing 100191, China
*. Corresponding author (E-mail: zdhyl2010@163.com)
Keywords: blind source separation; wavelet transform; FastICA; SOM
Abstract.This paper proposes a new one-stop blind source separation analysis method by
combining wavelet transform, FastICA and self-organization mapping network. Firstly, we suppress
the Gauss noises by using wavelet transform. Secondly, we separate the mixed signals utilizing
FastICA to get the independent components. Finally, we cluster the signals with SOM to reveal the
latent relationship of independent signals. Experimental results have shown the validity and
effectiveness of the proposed method.
Introduction
Most of the measurement/detection method and equipment of a large-scale integrated system
work directly on the whole system. It is generally difficult to exactly acquire signals emitted by all
the electronic components. As a result, a one-stop method is desirable to suppress disturbing
elements, and extract useful information on the electronic component from signals of the whole
system, and finally could provide accurate math simulation model for further system design,
analysis, prediction and evaluation.
Aiming at the signals of integrated system, this paper uses the ICA to pretreat the mixed data
which is collected. Separate the mixed signals with no prior information to get the independent
signals produced by different parts of system.
To solve the BSS problem, there are some methods now, such as K-means clustering [1], EASI
[2], and ICA [3]. ICA is the main solution of BSS. And FastICA [4] is being widely used, because
of its good flexibility, robustness and fast convergence.
The performance of FastICA is sensitive to the initial value of the separating matrix input weight.
Now, researchers proposed many improvements about FastICA. Paper [5] suggests to improve the
FastICA in this aspect. First, compare the performance with the best functional simulation selected
from three different nonlinear functions, and adjust the parameter of best function to make further
improvement. Then, compute the initial value by the best simulating function. The experiment
result shows that, this method can solve the problem that the performance is sensitive to the initial
value. Besides, it can avoid the nonuniform convergence and improve the result of separation.
Paper [6] suggests that, when the FastICA is used in dealing with big data (such as image
processing), the fifth-order convergence Newton iteration method can be used to accelerate the
convergence speed. The simulation result shows that the improved FastICA has better separating
feature and less convergence counts.
Most of the improvement methods now are aiming at the separation method itself, trying to
improve the separating result and the convergence speed. There are little researches importing the
intelligence artificial algorithms such as artificial neural network, to build the one-stop processing
method for integrated system signals including noise.
The conventional ICA method is sensitive to noise. It cannot separate valuable signals correctly
when there is serious noise pollution. To solve this problem, this paper proposed the improved
method combing the wavelet transform and FastICA. Reduce the noise of collected signals with
wavelet threshold method. Improve the signal to noise ratio, then separate the signals using FastICA.
International Conference on Intelligent Systems Research and Mechatronics Engineering (ISRME 2015)
© 2015. The authors - Published by Atlantis Press