Optimisation stochastique d'un melange de cru de ciment

One of the critical steps of the cement process is the production of the raw mix. A homogeneous raw mix will allow the production of a quality cement. It will also have a stabilizing effect on the clinkerisation process and reduce the energy consumption. The cement raw mix is the result of a mixing process of different materials. Some of these materials are found in the quarry adjacent to the cement plant (limestone, clay) while others (iron, for example) are brought to the plant. The present thesis is interested in making the raw mix more homogeneous by controlling the proportions of the different materials composing the mix. The control system developed in this thesis consists in estimating the materials and the mix using geostatistical estimation methods such as simple and ordinary cokriging. The covariance matrices used in such estimation methods are derived from the correlations and cross-correlations observed in the materials producing the raw mix, as well as the past proportions used to produce the mix. Different parameters were added in the control algorithm, such as gain and proportion variation control factors. These parameters were optimized so as to get the most homogeneous mix. The use of mobile means in the control system was also studied.;The various case studies helped determine that, with the right parameters, QMC allowed to produce the most homogeneous raw mix. However, the control algorithm with an estimation by cokriging presented advantages such as less variations in the material proportions.;The control algorithms developed using simple and ordinary cokriging were compared to a variant of the control algorithm used in Lafarge cement plants (a control algorithm using the last known composition of the mix to estimate its next composition, known as QMC) in various situations simulated using multivariate simulations. In the first situation, a rapid change of the mean on one of the materials was simulated. The control algorithm was then applied to simulations with different nugget effects and range. Finally, the reaction of the control algorithm was studied when the nugget effect and the range of the model given to the algorithm differed from the nugget effect and range of the simulated materials.