Algorithms And Applications Of Blind Source Separation

Blind source separation (BSS), or source signal separation, as a branch of blindsignal processing, has become one of the hotted research area in signal processing. Ithas many potential applications especially in communication system, biomedical signalprocessing, speech signal processing, signal analysis and process control, etc. The goalof BSS is to recover the original signals from a set of mixed (observed) signals with noor little prior knowledge about the sources and mixture way. In this dissertation, wemainly discuss the BSS algorithms and applications for linear instaneous mixturemodel.First, the theories and methodologies of BSS are discussed. The necessaryconditions and ambiguities of BSS are analyzed. The preprocessing techniques ofcentering and whitening for BSS are illustrated. Some knowledges about informationtheory, say entropy, negentropy, mutual information, kurtosis and high-order cumulantare given. The popular BSS algorithms are explained. The main contributions of thisdissertation are:1. The BSS for stationary source using its auto-correlation is discussed. Theinfluence of the linear prediction parameters on the extraction result is analyzed.However, the parameters are randomly generated. To improve the extractionperformance, this dissertation proposes to estimate the linear predictor parameters byKalman filter. Simulation verifies the effectiveness of the approach.2. A new objective function and the corresponding algorithm are proposed toextract the source signal of auto-regressive model. The noisy mixture is taken intoconsideration as well as the noise-free case. The theoretic analysis is given to supportthe proposed approach. Since the algorithm is a second-order statistics method, it saveslots of computation time and resources and has fast extraction speed.3. The fetal electrocardiogram (FECG) extraction is a typical BSS problem. Thisdissertation extends the work of Barros to extract the FECG based on the noise-freemodel. The improved algorithm speeds up the extraction. However, the extracted signalstill has some noise, which is the result of extraction by ignoring the noise. Then the noisy case is considered, and the new objective function and the correspondingalgorithm are presented. The algorithm is proved to be monotone decrease andeventually lead to the FECG. The simulation results show the effectiveness of theproposed method and indicate BSS has application value in FECG extraction.4. The problem of separating the nonnegative sources can be deduced to find anorthogonal demixing matrix to make the outputs nonnegative. This dissertationimproves the nonnegative ICA algorithm to increase the separation performance. Itrevised the existing objective function by adding a term to guarantee the orthonomalityof the demixing matrix. The improved nonnegative ICA algorithm is used for facerecognition. The experimental results illustrate the ICA based feature extraction is aneffective approach in the face recognition field.