Blind Source Separation And Its Application In Electrocardiography And Speech Signal Processing

In the circumstance where both the sources and the mixing system are unknown, recovering or estimating the sources based on a little prior information from observations is what called blind source separation (BSS). Normally this prior information refers to that the sources are mutually independent. Under the assumptions of linear instantaneous mixing and independent source components, the problem of BSS is what called independent component analysis (ICA). Due to the fact that independence of source components is not a strict condition, BSS has a wide range of applications in many fields, for example, speech processing, image processing, biomedical signal processing, communication and radar, etc..Research of blind source separation has been of two decades of history, during which a variety of methods and algorithms were proposed. Then what is the kernel of BSS theory? A few of literatures have outlined the main body of BSS theory and technology. However, it is far more than a refined core of BSS. Classifying as separability, principles of separation and optimization algorithms, kernel issues of ICA were first presented and discussed in this thesis.Negenttropy is a widely accepted criterion of BSS and it is used as a measure of nongaussianity. Changing the measure of independence to that of nongaussianity, however, is not rigorous and not convincing from the viewpoint of statistical mechanism of BSS. Regarding of this issue, detailed theoretical analysis from the aspects of statistical principle, approximation of negentropy and separation mechanism behind negentropy criterion, as well as simulation study, were presented in this thesis. Related arguments were also given in the paper.In the recent years, some researchers applied BSS technology to separate blindly atrial fibrillation (AF) wave from ECG. Most of these BSS algorithms were based on 12-lead signals. There exist some problems in the methods. First, many experiments of separation shows that there are two separated components which have the characteristics of AF wave. This may reveal that it possibly is not right to take AF as single independent source component in the model of 12-lead ECG for ICA. Secondly, six limb leads are not linearly independent. Taking these into account, a fast algorithm based on precordial 6-lead signals for the blind separation of atrial fibrillation was proposed in the thesis. On the hand, current BSS methods of AF neglected a fact that AF is quite random and its form and propagation might be quite different in different cycle of heart beats. To this point, another BSS approach of AF was proposed also.When separating signals with segmented data, there are no consistent component index mapping and no consistent signal amplitudes between two different segments because of the ambiguity of BSS. By changing the expression of ICA model in this thesis, a new concept of rotation angle of source component was introduced, a separability theorem for multiple-frame BSS was given, a general approach of determining the rotation angle was proposed, and therefore a preliminary theory framework for multiple-frame identifiable BSS was outlined. An approach for the underdetermined blind separation of meeting-type speech under this framework was presented.