Identification of a rice drying model with an improved sequential optimal design of experiments

Getting relevant parameter estimation of a non-linear model is often a hard task from both an experimental and numerical point of view. The objective of optimally designed experiments procedure is to diminish the experimental effort needed such that the identification is within acceptable confidence ranges. After each experiment, the next experiment is optimally designed, taking into account all past experimental results. It allows quality information to be extracted from the experimental data with less experimental time and resource consumption.In this paper, we present an original approach and implementation of the classical A-, D- and E-optimality on the estimation of 5 unknown (transfer related) coefficients in a compartmental model used to describe the convective drying of rice. The originality of our method is that it uses reparameterization of both parameter and protocol vectors which permits to avoid using a global optimization algorithm. The presented method is implemented in Matlab as a Toolbox and fully tested on a pilot plant. The case study (drying of rice) is typical in the field of process engineering: the dynamic model is strongly non-linear in its parameters and cannot be analytically solved. In addition, the specific technical constrains (inertias, limits, etc.) on the pilot are explicitly taken into account for improved experimental feasibility.In this drying application, three experiments with non-constant drying conditions are shown to be quite as effective as a two-factor three-level grid of nine experiments at constant conditions, with only one third of the experimental effort.

► A, D & E-optimal design of real experiments is performed for a model of coupled ODEs.
► Reparameterization is used, Fisher information matrix is computed semi-analytically.
► The 2-factor 3-level design is far more time & product consuming than the A-optimal.
► Tens of local optimisation with random initial guesses replaced global optimisation.
► A-optimal design is better than D one on model composed of coupled non-linear ODEs.