Home > Workflow collections > Public records > Zerlegung von Signalen in unabhängige Komponenten: Ein informationstheoretischer Zugang |
Dissertation / PhD Thesis | FZJ-2014-03619 |
2005
John von Neumann Institute for Computing
Jülich
ISBN: 3-00-013620-7
Please use a persistent id in citations: http://hdl.handle.net/2128/6765
Abstract: The main object of this dissertation is to decompose multivariate signals into components as independent as possible. For this purpose a new approach based on information theory is developed that allows not only to solve the minimization problem but also to further analyze the components obtained. In the first part of this dissertation a precise estimator of mutual information (MI) is introduced which is used as starting point for further developments. In numerical tests it is proven that for a variety of distributions this new estimator shows minimal systematical and statistical errors in comparison to other estimators. Furthermore, by means of MI a novel cluster method is formulated which defines distances of higher dimensional objects without having to average over properties of the single objects.In the second part the MI estimator is used to formulate a new independent component analysis (ICA) algorithm. Numerically, it is found that the algorithm outperforms previous algorithms in regard to accuracy. In addition, the exact estimation of the absolute dependency allows to express a reliability test for any ICA component. Subsequently, an extended ICA algorithm is formulated which achieves improved results with noisy data.In the third part of this dissertation these new methods are applied to experimental data. First, the ECG of a pregnant woman is separated into signals of the mother and fetus by an extended version of the standard ICA model. In the next application the infrared spectra of the pure substances are identified from the spectra of organic mixtures, despite the fact that the first ones are not really independent. Subsequently, the new algorithms are used to eliminate the bias in a synchronization analysis due to an overlap of two signals. As a last application it is shown that multivariate methods are also able to perform a sleep stage classification on intracranial EEG data.
Keyword(s): Dissertation
The record appears in these collections: |