Theory Of Adaptive Blind Separation Based On High Order Statistics

Blind signal processing is a new orientation in the field of signal processing, it also has broad range of application in daily life. Blind source separation is to separate sources from mixed signals, without knowing about the characters of the sources and how they are mixed, only assuming that the sources are independent with each other. Blind source separation is a part of blind signal processing. This paper only does research on adaptive blind source separation algorithm in two mixture models梚nstantaneous mixture and convolutive mixture.Blind source separation needs high order statistics. The algorithm based on cross cumulants uses high order statistics directly. The basic idea of the algorithm is shown by the characteristics of cumulants; the performance of the algorithm is discussed; the shortcomings of it are pointed out.Infomax algorithm is another adaptive blind source separation. Infomax uses high order statistics by non-linear function and the idea from Information Theory. The original Infomax is only applied in the instantaneous mixture and in the single channel blind deconvolution. This thesis applies the idea to blind sources separation in multiple channel convolutive mixture. Through simulation, we can show that the algorithm has good performance.Natural gradient is a new optimization way that is proposed in a special space?Riemannian Space. The parameter space in blind source separation is Riemannian Space. Natural gradient has many better characters compared with normal gradient. This thesis applies natural gradient to Infomax algorithm. And with the special character of the natural gradient, the performance of the algorithm is discussed.Applying blind source separation to the real situation is the basic goal. This paper uses the blind source separation algorithms to process the data received form an experiment in water tank.