Application of nonlinear adaptive signal processing techniques to blind source separation and interference suppression

Abstract

In active noise and vibration control it is important to identify independent noise generation mechanism and sources. In all large physical systems only a mixture of noise sources can be observed. The mixture model can be either linear or nonlinear. Recently there has been an explosion of significant research literature on blind deconvolution. In blind deconvolution, the actual mixture model and source signals are not known; however, in the theoretical development it is assumed that sources are statistically independent. In practice, good results have been obtained even if sources are not independent. If the mixture model is linear, statistically independent noise sources can be separated by blind deconvolution techniques that are based on the theory of the independent component analysis (ICA). We implement ICA by a cascade of the linear adaptive algorithm that is based on singular value decomposition (SVD) and a nonlinear algorithm that is based on a radial-basis function (RBF) neural network. A RBF neural network can also approximate the inverse of the nonlinear mixture model. We use a fast RBF updating algorithm for adaptation of the nonlinear network. Theoretical results are illustrated by computer simulations. Results are compared with linear adaptive signal processing algorithms. [This research has been supported by ONR, Dr. Kam Ng, Program Officer.]