Study Of The Blind Separation Based On Time-Frequency Ratio

The problem of blind signal separation is to extract source signals when only the observed data which obtained from the receiving antenna with unknown source signals and communicatory channel. Namely, the blind signal separation problem is to recover a set of source signals when only their mixtures with unknown coefficients are observed. Because the problem of blind signal separation arises in diverse fields of science and engineering like speech analysis and recognition, array processing, multi-user detection, EEG data communication, image recovery, feature extraction, earthquake wave detection, etc, the algorithm research about that problem has become the most important problems in signal processing field since 1990s.Blind signal separation has been researched for several years. But it still has many problems which are waiting to be solved. And most methods are based on ICA, especially, many researchers regard the BSS as ICA, in fact, this point of view reduced the area of BSS conception. However, a new idea to solve the BSS is do it in Time-Frequency (TF) area, and several investigations based on time-frequency analysis have been reported. At present TF BSS methods can be divided to about two classes. The first one is WVD two step methods, this method broke through the limit of stationarity, but still significantly related to classical ICA principle. The second class includes several methods based on TF transforms of observed signals. Some of these methods, i.e. DUET and its modified versions, require the sources to have no overlap in the TF domain, which is quite restrictive. This paper considers a new TF method which is based on the Time-Frequency ratio method given by Deville in 2005. Then we propose an improved TIFROM approach with outstanding feature.First of all, thesis introduces the investigation background, history and development of the BSS, presents several common mixing models. Especially for the instantaneous mixing model, introduces a lot of theoretic correlated to classical BSS methods and this paper method, which include statistical theory, information theory and Time-Frequency analysis knowledge. Then this paper explains several classical BSS methods as INFORMAX, Extended-ICA, Fast-ICA based on Newton iteration, and WVD method based on Wigner-Ville distribution.Time-Frequency Ratio based BSS as a novel BSS concept is discussed in detail. Through computing the ratio of mixed signals TF transformation in one TF window, this method can estimate the elements of the mixing matrix. The key is how to find the area where only one source occurs. Algorithms to find the single source zone with TIFCORR, TIFCOHERE and TIFROM are compared in this paper. The computer simulation experiment results indicate that the TIFROM method have the best Signal-to-Interference Ratios.However, with the number of the sources increasing, the problems about separating lost and too long operating time exist when using TIFROM. This paper consider the feature of signals TF transformation, first select analysis zone properly instead of the whole TF domain, then analysis the TF ratio relationship between one of the mixing signals and all the others. Using the minimal variance method, single source zones exist precisely. This method increases the Signal-to- Interference Ratios obviously. In common it can complete the separating program in one time if the number of mixing signals is not too large, the efficiency of the algorithm has been improved evidently. The performances of improved TIFROM, TIFCORR, TIFCOHERE and TIFROM are compared using computer simulation to approve that the new method is more excellent.An well-known technique exist for introducing coherence between the phases of frequency-adjacent FFT or STFT points, by unwrapping these phases with respect to frequency. This paper discusses this phenomenon, divides the analysis zones according to constant time and adjacent frequency. Phase unwrapping will be a straight line in single source zone. Resolving the phase unwrapping of the mixing signals TF transformation also using the minimal variance method can find the single source zone precisely. Therefore, this method resoles the time-delayed problem effectively.As a new method of the area of BSS, the time-frequency ratio method has many special feature compared with those classical ICA methods. This paper give several computer simulations to compare these method in some complex condition include non-stationary signals, two gauss signals, dependant signals, and the model with adding noise. The algorithms include improved TIFROM, TIFROM, FastICA, Extended-ICA, INFORMAX and WVD. Experiment results approve that the new method can solve the difficulties better than others.A useful signal processing method must have the ability of practical application. BSS as a new technology have a wide application area and foreground. The method based on time-frequency ratio is a new member in this area, its ability of practical application is more important. This paper discusses its capability with speech signal and EEG signal. The experiment results present that the method separated the speech signal successively and subtract the EOG interference in EEG signal effectively.The methods based on Time-Frequency Ratio blaze a way in BSS area. However, as a new approach, it still has many deficiencies to be investigated, e.g. the condition with low signal-noise ratio, convolution mixing model, undetermined mixing, etc.