A new look at the statistical identification of nonstationary systems (Q2188279)

Consider a discrete time nonstationary linear system governed by equation \(y(t)= \varphi^T(t) \theta(t)+e(t)\), \(t=\dots,-1,0,1,\dots\), where \(\theta(t)=(\theta_1(t),\dots,\theta_n(t))^T\) is a vector of time-varying system coefficients (parameters), \(\varphi(t)=(u(t-1),\dots,u(t-n))^T\) is a vector of past values of input signal \(u(t)\) and \(e(t)\) is a measurement noise. It is assumed that \(\{u(t)\}\) is a zero-mean wide-sense Gaussian stationary process with exponentially decaying autocorrelation function \(r_u(i)\), \(\{e(t)\}\) is independent of \(\{u(t)\}\) zero-mean sequence of i.i.d. random variables with finite variance \(\sigma_e^2\) and \(\{\theta(t)\}\) is a uniformly bounded sequence. Given \(r_u(i)\) and \(\sigma_e^2\), the goal is to estimate sequence \(\{\theta(t)\}\) based on the observations in sequence \(\{y(t)\}\). In the article a two-stage identification procedure is proposed. At the first stage a preestimation of parameter trajectories is implemented which results in unbiased but very noisy estimates. At the second stage preestimated trajectories are filtered (so, this is a postfiltering stage) using the basis function framework. Method of postfilter settings is based on parallel estimation and cross-validatory analysis. It is shown that the proposed two-stage procedure is not worse than the classical method of parameter tracking but provides significant computational savings. Simulation results for nonstationary two-tap FIR system, illustrating the proposed technique, are presented.