Research Of Feature Extraction Based On Independent Component Analysis And Its Applications

Independent Component Analysis (ICA) is a kind of powerful method for Blind Signal Processing (BSP). The principle of ICA algorithm is to find the potential mutual independent components, to remove higher-order redundance between components and to extract the independent original signals source by analyzing the high-order statistical correlation of the multidimensional data. This property leads to a promising prospect of ICA in applied fields such as telecommunications, audio signal separation, biomedical signal processing, and image processing. And ICA becomes one of the most exciting new topics both in the fields of signal processing and pattern recognition. The main work of this thesis is to discuss the ICA as a method of feature extraction.Support Vector Machine (SVM) is a new kind of machine learning algorithm proposed recently which is based on VC dimension theory and structural risk minimization of statistical learning theory. SVM can obtain the optimum result from the gained information which is not the optimum result only when the samples are infinite. SVM has much stronger theory foundation and better generalization than neural network which is based on empirical risk minimization. So SVM is popular in recognition field.ICA is still staying at the developing stage, and the investigation of its theory and application should be enhanced and improved further. The basic principles of ICA and some algorithms are introduced in this thesis. Aiming at the problem that Fast ICA feature extraction has the shortcoming of long computing time, a modified ICA algorithm is suggested. And we proposed a scheme to integrate ICA feature extraction and SVM for gas mixture quantitative analysis and electrocardiogram (ECG) diagnosis. The main works in this thesis are as follows:(1) The basic principles of ICA and some algorithms based on information theory are introduced. Two kinds of effectual algorithms in common use are discussed in detail, FastICA and InfoMax ICA.(2) Kernel Independent Component Analysis (KICA) is a new developed nonlinear ICA algorithm. We proposed a pattern recognition method for gas mixture quantitative analysis by combined use of KICA and least squares support vector regression (LSSVR) in this thesis. In the proposed method, the KICA algorithm based on an entire function space of nonlinear subspace is firstly used for preprocessing gas sensor data, which can reduce the data correlation. And then a LSSVR carries out the gas mixture recognition, The measuring data was obtained from a gas mixture of butane and ethanol for experiments. The results indicate that the KICA method is efficient.(3) Modifying the kernel iterative procedure of FastICA with accelerating aitken method, an improved FastICA (I-FastICA) algorithm is given. The I-FastICA algorithm can cut down iteration times. Also, the convergence process of the algorithm can be accelerated. We proposed a scheme to integrate improved fast independent component analysis I-FastICA and SVM for ECG diagnosis. In the proposed method, the I-FastICA algorithm is firstly used for feature extraction of ECG data, and then a SVM carries out the ECG signal classification. The ECG samples attributing to seven different beats types were sampled from the MIT-BIH arrhythmia database for experiments. The results show that the I-FastICA algorithm reduces computation time with the correspondent separation performance.