Unbiased parameter estimation of linear systems with colored noises (Q1360493)

This paper deals with the problem of consistent parameter estimation for the linear system modeled by \(A(z^{-1}) y(k)= B(z^{-1}) u(k)+ \nu(k)\) in the general case where the measurement noise \(\nu(k)\) is unknown and correlated not only with the system output \(y(k)\) but also with the system input \(u(k)\). A new method is presented for eliminating the noise-induced biases in the least-squares (LS) estimates. Two stable filters \(f_1(z^{-1})= \prod^{n+ 1}_{j=1} (1-\lambda_i z^{-1})\) and \(f_2(z^{-1}) =\prod^{n+1}_{i=1} (1-s_i z^{-1})\) are introduced to filter respectively \(u(k)\) and \(y(k)\) so as to get an augmented system \(\overline A(z^{-1}) \overline y(k)= \overline B(z^{-1}) \overline u(k) +\nu(k)\) with \(n+1\) known zeros \(s_i\) and poles \(\lambda_i\), where \[ \overline A(z^{-1}) =A(z^{-1}) f_1(z^{-1}), \quad \overline B(z^{-1})= B(z^{-1}) f_2(z^{-1}) \] \[ \overline y(k)= y(k)/f_1(z^{-1}), \quad \overline u(k)= u(k)/f_1(z^{-1}). \] After eliminating the noise-induced biases in the ordinary LS estimate \[ \widehat \theta_{LS} (N)= R^{-1}_{\overline \varphi\overline \varphi} (N) R_{\overline \varphi \overline y} (N), \] the consistent estimates of the parameters in \(\overline A(z^{-1})\) and \(\overline B(z^{-1})\) are: \[ {\overset \Delta \theta}_{BELS} (N)= {\overset\Delta \theta}_{LS} (N)-R^{-1}_ {\overline \varphi\overline \varphi} (N)Q\widehat M_N^{-1} \left[ \begin{matrix} H_1 {\overset\Delta \theta}_{LS} (N)-{\mathbf p}_1\\ 0 \end{matrix} \right] \] where \(Q=[0,I_{4n+2}]\in R^{(4n+2) \times (4n+4)}\), \(R_{ \overline\varphi \overline\varphi} (N)= {1\over N} \sum^N_{k=1} \overline\varphi_k \overline \varphi^\tau_k\), \(R_{\overline \varphi \overline y} (N)= {1\over N} \sum^N_{k=1} \overline \varphi_k \overline y(k)\) \[ \overline\varphi_k= \bigl[-\overline y(k-1), \dots, -\overline y(k-2n-1),\;\overline u(k),\dots, \overline u(k-2n) \bigr]^\tau \] \[ M_N= \left[ \begin{matrix} 0 & 0 & \alpha_1 & \cdots & \alpha_{2n-1} & \alpha_{2n} & \alpha_{2n+1} & \cdots & \alpha_{4n+2}\\ -H_1\lim_{N \to \infty} {\overset \Delta \theta}_{LS} (N)+ {\mathbf p}_2 & \beta_1 & \beta_2 & \cdots & \beta_{2n} & 0& \beta_{2n+1} & \cdots & \beta_{4n+2} \end{matrix} \right] \] \vskip 2mm \[ {\mathbf p}_1 =[-\lambda_1^{ 2n +1}, \dots, -\lambda_{n+1}^{2n+1}, 0,\dots, 0]^\tau\in R^{(2n+2) \times 1},\quad {\mathbf p}_2 =[-1, \dots, -1,0, \dots,0]^\tau \in R^{(2n+2) \times 1} \] \vskip 2mm \[ H_1= \left[ \begin{matrix} \Lambda_1 & 0\\ 0 & S \end{matrix} \right],\;H_2= \left[\begin{matrix} \Lambda_2 & 0 \\ 0 & S\end{matrix} \right],\;\begin{cases} H_1R^{-1}_{\overline \varphi\overline \varphi}= [\alpha_1, \dots, \alpha_{4n+2}]\\ H_2R^{-1}_{\overline \varphi \overline \varphi}= [\beta_1, \dots, \beta_{4n+2}] \end{cases} R_{\overline \varphi\overline \varphi}= \lim_{N\to \infty} R_{\overline \varphi\overline \varphi} (N) \] \vskip 2mm \[ \Lambda_1=\left[\begin{matrix} \lambda_1^{2n} & \lambda_1^{2n-1} & \dots & 1 \\ \vdots & \vdots & & \vdots\\ \lambda_{n+1}^{2n} & \lambda^{2n-1}_{n+1} & \dots & 1 \end{matrix} \right],\qquad \Lambda_2= \left[ \begin{matrix} \lambda_1^{2n+1} & \lambda^{2n}_1 & \dots & \lambda_1\\ \vdots & \vdots & & \vdots \\ \lambda^{2n+1}_{n+1} & \lambda^{2n}_{n+1} & \dots & \lambda_{n+1} \end{matrix} \right], \] \[ S=\left[ \begin{matrix} s^{2n}_1 & s_1^{2n-1} & \dots & 1 \\ \vdots & \vdots&& \vdots \\ s^{2n}_{n+1} & s^{2n-1}_{n+1} & \dots & 1 \end{matrix}\right]. \] A recursive algorithm for \({\overset\Delta \theta}_{B ELS}\) is also developed, and the analysis shows that the results of the proposed method are independent of the noise model used.