DERIVATION OF A DESIGNER GUIDE FOR INTERACTIVE GREY-BOX IDENTIFICATION OF NONLINEAR STOCHASTIC OBJECTS

The problem of using an 'identification tool box' for the design of 'grey-box' models for nonlinear dynamic objects is non-trivial. The design is an interactive process, and it is not given a priori in what order to execute the various subtasks that the tool box supports what design parameters to manipulate and how to interpret the intermediate results. The difficulties are enhanced when the uncertainty of the designer's a-priori information and the quality of the experiment are such that a model contains other stochastic elements than measurement error. This paper derives a systematic procedure for design of such models, assuming a generic tool box. The origin is the basic procedure commonly used in natural sciences, namely that of repeated refinement and falsification of hypotheses. The derivation is based on statistical decision theory and leads to the specification of a 'designer's guide' for grey-box identification. The procedure has been tested on two industrial processes, using the IdKit tool box, and a prototype of the guide has been implemented as a Unix shell. A simulated example illustrates the procedure.