Estimation of constant parameters in a class of time varying plants by means of smoothing

The problem addressed in this paper is as follows. Consider a nonlinear discrete time system with parameters, which can be a priori classified into two groups: constant parameters and time varying ones. It is also assumed that loose attributes like "slow" or "fast" express the only prior information concerning the dynamics of varying parameters. The question of interest is how to estimate the constant parameters of the process (on the basis of a recorded batch of input and output data) so that possible alterations of the parameters that are expected to vary with time are taken into account. To deal with this kind of nonlinear fixed-interval smoothing problem a two-level procedure based on rectangular window technique and Gauss-Newton method is presented. All the computations are performed offline. The procedure is applicable in a pre-processing stage of adaptive parameter estimation where the parameters that are evidently time invariant are identified from past records. Thus the number of parameters that have to be tracked (i.e. estimated) on-line is reduced. The efficiency of the scheme is illustrated on a simulated industrial process.