Least costly identification experiment for control

Project members:dr.ir. X.J.A. Bombois (Xavier), M. Gevers (Université Catholique de Louvain, Belgium), G. Scorletti (Université de Caen, France), prof.dr.ir. P.M.J. Van den Hof (Paul), H. Hjalmarsson (KTH, Sweden), M. Barenthin (KTH, Sweden), M. Gilson (CRAN, France)
Keywords:Identification and estimation, Data driven and fault tolerant control, Process technology

Model-based control has nowadays reached many industrial
sectors (chemical industry, high-tech manufacturing industry,
...) as a crucial technology in realizing optimal process
operation. However, the development of model-based control is
still restrained by various open issues in control and system
theory. In most of these issues, the harmonious interaction
between system identification and robust control design is very
important. Since the beginning of the nineties, important steps
have been taken towards this harmonious interaction. Indeed, it is
now possible, given a model and its related uncertainty as
delivered by system identification, to check whether a controller
meets the robust performance requirements. However, until very
recently, there was still a complete lack of results allowing the
design of robust controllers based on cheap and plant-friendly
identification. Indeed, until then, very few attention had been
devoted on the choice of the data used to identify the model and
its uncertainty region and the lack of clear guidelines for the
choice of these data generally lead to identification experiments
that were more expensive than actually necessary, i.e. the
identified uncertainty region was either too large (and thus
unusable) or smaller than strictly necessary for the required
level of performance. Knowing that the generation of informative
data for an identification experiment is the most expensive step
of the whole robust control
design procedure, the related economic loss could be very important.

Based on these considerations, a brandnew line of research has
recently been defined. This research aims at developing tractable
techniques for an optimal design of the identification experiment.
In particular, the objective is to determine the least costly
identification experiment that delivers sufficient information on
the true system dynamics (i.e. that delivers a model with a
sufficiently small uncertainty region) for the design of a robust
controller with a satisfactory performance.

In particular, the influence of short data sets and the extension of the concepts
towards performance monitoring and robust filtering are currently investigated.

© Copyright Delft Center for Systems and Control, Delft University of Technology, 2017.