In this paper, the problem of time-varying parametric system identification
by wavelets is discussed. Employing wavelet operator matrix representation
, we propose a new multiresolution least squares (MLS) algorithm for time-v
arying AR (ARX) system identification and a multiresolution least mean squa
res (MLMS) algorithm for the refinement of parameter estimation. These tech
niques can achieve the optimal trade-off between the over-fitted solution a
nd the poorly represented identification. The main features of time-varying
model parameters are extracted in a multiresolution way, which can be used
to represent the smooth trends as well as track the rapidly changing compo
nents of time-varying parameters simultaneously and adaptively. Further, a
noisy time-varying AR (ARX) model can also be identified by combining the t
otal least squares algorithm with the MLS algorithm. Based on the proposed
AR (ARX) model parameter estimation algorithm, a novel identification schem
e for time-varying ARMA (ARMAX) system is presented. A higher-order time-va
rying AR (ARX) model is used to approximate the time-varying ARMA (ARMAX) s
ystem and thus obtain an initial parameter estimation. Then an iterative al
gorithm is applied to obtain the consistent and efficient estimates of the
ARMA (ARMAX) system parameters. This ARMA (ARMAX) identification algorithm
requires linear operations only and thus greatly saves the computational lo
ad. In order to determine the time-varying model order, some modified AIC a
nd MDL criterions are developed based on the proposed wavelet identificatio
n schemes. Simulation results verify that our methods can track the rapidly
changing of time-varying system parameters and attain the best balance bet
ween parsimonious modelling and accurate identification.