TVAR Parametric Model Identification Based On RBFNN And Its Application

In recent years, the time-varying parametric model, especially time-varying auto-regressive (TVAR) parametric model is used to describe a nonstationary random signal widely for its parsimony and high resolution. Estimation of TVAR parametric model can be done by representing the whole system behaviour in terms of some fixed basis sequences, which was called coordinate approach. Then the identification task was equivalent to the estimation the parameters in this expansion. Comparing to the adaptive algorithm assuming that the signal was locally stationary over a relatively short time, the coordinate approach can estimate the time-varying parameters more effectively. And the time-varying parametric model had higher resolution when processing big product of time and bandwidth signal.This paper firstly introduced nonstationary random signal and its research direction. Then described TVAR parametric model and investigated two major problems of the model: the selection of basis sequences and the determination of model order. Based on radial basis function neural network (RBFNN)' good performance on function approximation, applied RBF to the TVAR parametric model.Then two kinds of reducing redundancy approach were discussed and modified, one was KL transform of coefficient matrix, and the other was time-varying optimal parameter search (TV-OPS) algorithm. New order selection criterion was also presented in terms of the normalized order projection distance based on decline factor (NOPD). The effect of basis functions degree to estimate time varying parameters were investigated, and the selection of optimal degree was also discussed. Then we present a new reducing redundancy approach based on 2D-DCT of basis functions vector matrix. Experiment results on synthesized data and real speech signal verified the effectiveness of the novel approach.As an application of TVAR parametric model, multi-component chirp signal was modelled with the modified model, and by combining the Radon transformation and the CLEAN method, a new technique for multi-component chirp signal detection and parameter estimation in terms of Radon-TVAR was presented. The results proved that the new algorithm can detect the parameters of multi-component chirp signal more effectively than the traditional methodsFinally, summarized the work of this paper and some problems on the TVAR model to be solved and pointed out the direction of the development.