With matrices A e 9\n.n , Be 9\n.1 and C e 9\1.n. We denote x(/)e 9\n the state vector describing the system, u(t) the input signal to be restored, and Ym(t) is a noisy signal. The signals v](t) and V2(t) are random signals supposed to be white, zero mean uncorrelated Gaussian processes with covariance (Jv] and (Jv2' And the associated Kalman filter is given by:The gain of the stationary Kalman filter is given by:the positive symmetric matrix solution of the Algebraic Riccati Equation (ARE): AP+PA t -PCta:iCP+Bav1B t =0 (4) Let the augmented state vector be defined as: X (I) = [xp (I)]Xe (/) we then obtain the following augmented state equations:The deconvolution procedure developed by [2] is depicted by figurel.Figure 1. Deconvolution structure used by [2}.We have to find the signal u(t) minimizing:
ABSTRACTWhere h(t) is the impulse response of the process and u(t) is the excitation. The inverse operation of recovering the signal u(t) fromy(t) and h(t), is called deconvolution. As one may find in the literature, this problem is known to be an ill-posed one [1]. To be more precise, the ideal solution is the inverse system but the presence of measurement noise does not allow a direct inversion. Meanwhile, an approximate inversion through the well-known regularization method can be easily obtained. We propose a deconvolution procedure based on the one developed by Sekko & al.[2] for time-invariant systems, which was later improved by Neveux [3,4] in order to adapt it for time-varying processes. In the deconvolution method [2], the signal of measurement is corrupted by a system represented by a transfer function or a state equation. In the proposed procedure the deconvolution is undertaken by means of a trajectory tracking scheme. We use the following state equations to represent the system under study: .Ym (I) = C.xp (I) + V2 (I)In control theory, the response of a process to an exciting signal is described by the following convolution product: +00 y(t) = fh(t-t') u(r) dt'(1)