Research Of Independent Component Analysis Algorithm And Its Application To Biomedical Signal Processing

Independent component analysis (ICA) is a recently developed signal processing technique. Its basic task is separating or extracting independent source signals that are linearly combined in observations. Recently, there is a trend to develop blind source separation (BSS) or blind source extraction (BSE) algorithms based on ICA, due to its potential applications in a lot of fields, such as biomedical signal processing, telecommunication system, image processing and speech processing. In recent years, there are some progress in ICA theories and algorithms. However, there are still many unsolved problems exist, which could restrict the development of many ICA applications. Many organizations have done lots of work to promote the applications of ICA in many fields. As a result, ICA has become one of the most exciting hot topics both in the fields of neural networks and signal processing.This dissertation mainly focuses on basic ICA algorithms, constrained ICA algorithms, extraction algorithms with noise model, and their applications to biomedical signals extraction. Especially, the main contents arc as follows:1. This dissertation proposes a fast and adaptive algorithm based on fully-multiplicative orthogonal-group, which solves some problems of current basic ICA algorithms, such as slow convergence speed. The algorithm adopts a density model that combines the t-distribution density model, the light-tailed distribution density model, and the Pearson system model so that it not only can separate mixtures of sub-Gaussian and super-Gaussian source signals, but also can separate skewed and near Gaussian signals. Therefore the algorithm was successfully applied to obtain a clear Fetal Electrocardiogram (FECG) signal with better separation performance and faster convergence speed, compared with some famous basic ICA algorithms.2. The dissertation develops a robust extraction algorithm based on approximate negentropy to overcome some drawbacks of current constrained ICA algorithms, such as bad robustness to outliers. The algorithm is very robust to outliers because of using an approximate negentropy. And it only needs to estimate the coarse kurtosis value range of a desired signal, not need the additional priori information. Moreover, the algorithm can work well in some adverse situations when the kurtosis values of some source signals are very close to each other. All these make that the algorithm is an appealing method which directly extracts an accurate and reliable FECG.3. To solve the current problem about obtaining Atrial Fibrillation (AF) signal, the dissertation presents a two-stage based algorithm, which can successfully and directly extract a desired AF signal by two BSE methods. Compare with current BSS or basic ICA methods of separating all source signals, the algorithm based BSE is simple on operation, and can save a lots of times and resources. Extensive experiments on real-world data of patients suffering from AF have showed that it can rapidly and efficiently extract a clear AF signal and greatly reduce lots of noise. Therefore, the algorithm is expected to have great potential in clinical diagnosis.4. Many existing ICA or BSE methods are limited to noise-free mixtures, which are not realistic. So the dissertation proposes a novel cost function from that the effect of noise is removed. Maximizing the cost function, it can obtain a BSE algorithm, which caters for the effects of noise. The algorithm is robust to the estimation errors of the time delay as long as the errors are not too large. Compared with some classical algorithms, the proposed algorithm has better extraction performance in the presence of noise, as confirmed by simulations and experiments on real-world data.