Study On Target Perception-Oriented Blind Signal Processing Algorithms

Blind signal processing (BSP) is a very hot topic in the signal processing society. The advantage is that BSP does not require any prior knowledge except the independence among different sources, so BSP has been applied in many disciplines. This dissertation is targeted to the challenging problems in BSP. To improve the perception capacity is the final goal in the application with BSP enforcing the target in perception systems. Two frameworks are mainly studied. One is the blind separation of convolutive mixtures in heavy reverberation, which is the first hard problem the dissertation resolves. The complex-valued source separation and the permutation are studied for the blind separation of convolutive mixtures in the frequency domain. The second is the particle filtering based noisy source separation, which provides a novel approach for the post-nonlinear and underdetermined blind source separation under the noisy environment. In the real-word application, it is mainly discussed that how blind separation of convolutive mixtures are adopted in the speech source separation and the active sonar target detection.The dissertation first deals with the blind separation of complex-valued sources. After the improper complex-valued vector and the characteristic of pseudo-autocorrelation matrix are explored, we find that JADE, Complex FastICA, Complex ICA, SUT and Equivariant SUT do not make full use of the good characteristics of the improper complex-valued vector. Then, based on the pseudo-autocorrelation matrix, we construct a second order statistics (SOS) cost function. With the decent gradient algorithm, a new SOS based blind separation of complex-valued sources algorithm is inferred. We name it as Strong SOS. Along the same idea, we construct the cost function based on the minimum of the mutual information between different independent sources. In the process to infer the new algorithm, we perform the conjugate transpose and transpose operation to the complex-valued vector and matrix, and the new algorithm is obtained based on the pseudo-crosscorrelation matrix. We name it as Strong HOS compared to Strong SOS. Through simulations, it is proved that Strong HOS and Strong SOS perform better than JADE, Complex FastICA, Complex ICA, SUT and Equivariant SUT. It is the reason that Strong HOS and Strong SOS explore the characteristics of improper complex-valued vectors entirely. Blind separation of complex-valued sources is the first step to the blind separation of convolutive mixtures in the frequency domain. Strong HOS and Strong SOS make much solider foundation to correct the permutation in the second step.And then, the permutation indeterminacy of frequency domain approach to blind separation of convolutive mixtures is studied. Based on the pseudo-parameter of improper vectors, the traditional beamforming method——MVDR is revised. Simulation validates that the extended MVDR functions better than MVDR in low SNR situation.After the mutual parameters are deeply analyzed, it is natural to find that the computation method of DFT introduces the correlation to the mutual parameters method, hence, the mutual parameters can be extended to other signals except speech. By different simulation, different current algorithms are compared. KL distance behaves better than correlation coefficient. However, in some frequency bins, mutual parameters still possibly can not correct the permutation indeterminacy. At last, Strong SOS+ExMVDR, Strong HOS+ExMVDR, and Strong HOS+KL are composed together for the frequency domain method.To test the performance of algorithms we put forward, the experiment simulates the situation in the ordinary office. The reverberation time is 130 milliseconds and the sampling frequency is 8000 Hz. Sources are the Chinese speech of 17 words and the light music. The impulse filter is generated by the famous Image algorithm. Two standard algorithms are adopted, and they are the Polar ICA+MVDR of NTT in Japan and FDICA of ICA Center. The improved SNR from the Strong SOS+ExMVDR and Strong HOS+ExMVDR are over 1 dB than Polar ICA+MVDR. IBM Viavoice can recognize all the 17 words estimated by Strong SOS+ExMVDR and Strong HOS+ExMVDR, but only 14 by Polar ICA+MVDR and 13 by FDICA.This experiment validates the effectiveness of the methods we put forward in this dissertation. Our work pushes the development of blind separation of convolutive mixtures in the frequency domain.Particle filtering (PF) is an optimal solution to the nonlinear and nongaussian problem, and it has been a hot topic in the past ten years. With the development of BSP, noisy blind source separation and post nonlinear blind source separation gradually become more and more attractive. This dissertation revolves the noisy post nonlinear and underdetermined blind source separation through the application of PF. First, the noisy blind source separation model is decomposed to two parts which are the noise free model and the received signal contaminated by noises; second, in theory, the feasibility of the dynamic state space equations and noisy blind source separation is explored. Exactly, the Time Varying AR model of the noise free mixtures describes the state equation, and the relation of noisy mixtures and noise free mixtures constructs the observed equation in the state space problem; third, along this way, the noisy model can be converted to be the noise free model, and the post nonlinear model can turn into the linear model provided that the post nonlinear function is a priori or can be estimated. Thus, the current blind source separation algorithms can perform on the estimated noise free mixtures directly. Under this framework, the noisy linear determined problem, noisy nonlinear determined problem, noisy linear underdetermined problem and the noisy nonlinear underdetermined problem are resolved in the theory.As the solutions to noisy nonlinear problem are still open, the linear determined experiment is done first. The denoising source separation (DSS) of J. Sarela and H.Valpola in HUT is selected for the comparison to PF+FastICA. The mixing is instantaneous, and the noise is additive. The input SNR is between 0 to 12dB. No matter in gaussian noises or gamma noises, PF+FastICA performs better than DSS. This experiment proves our approach is effective. Then, under 10dB SNR noisy environment, the noisy nonlinear determined, noisy linear underdetermined and the noisy nonlinear underdetermined experiments are done respectively. The benefit of PF is between 5 and 6dB. The performance in the three experiments is close to the linear determined one. All the experiments prove the validation of our approach. The framework of PF+ICA provides a novel approach to nonlinear and underdetermined blind source separation under the noisy environment.In the application of blind separation of convolutive mixtures, we mainly explore the feasibility in the active sonar target detection. When the Doppler Effect is produced, both BNMF and whitening method are useful; or in the case that the target power has been already known, PCI may cancel the reverberation. However, these methods are no use provided that prior knowledge is not available. In this dissertation, the target echo is regarded as the first source, and the reverberation in shallow water or the background interference are the second source. After beamforming, the main beam and the adjacent beam are the observed signals. Strong SOS+Corr is adopted. The first experiment is about the simulated target, and the function of our approach is close to PCI. The second experiment is the true target. After Strong SOS+Corr is performed, the benefit of the matched filter output is about 2.5dB, and the characteristic of the time-frequency analysis is much better. The two experiments prove our approach and idea are helpful in the active sonar target detection, target feature extraction and target recognition.This work is supported by National Science Fund under grant 60372075, and also supported by Shanghai No: 05JC14026 and opening fund of VSN-2006-04.