The purpose of the work is to enable design and manufacture of cement with emphasis on the quality and the properties of cement. Data used in the design and manufacture were collected from predictions of properties and characteristics of cement. The properties of cement were predicted from its characteristics and from the production conditions in cement kiln and mill. The cement characteristics were in some investigations predicted from the production conditions. The design was based on sensitivity analysis focusing the influence of the characteristics and production variables, x, on the properties, y. The influence was analyzed by predictions of y from the simulated variation in x. In cases there were correlation within observation matrix X, the simulations were constrained by the latent structure of X. The simulations were in the form of optimization of the function for prediction of y. The optimal solutions of the production variables were then implemented in the manufacture of cement.
The prediction models were evaluated by multivariate data analysis by using partial least square regression (PLS). PLS is a member of the bilinear class of methods. The method compresses the observation matrix X to its most relevant factors and using these compressed variables as regressors for y. In the thesis the compressed variables were sometimes called PLS-components and sometimes latent variables.
Two types of sensitivity analysis for the examination of influences of the x-variables on y were applied. The first is based on comparison of the size and certainty of the regression coefficient from PLS on scaled and weighted x. The second was based on predictions of y from simulated variation in x. The influence of a single x-variable or a latent variable was valued or ranged from the variation of predicted y relative to the confidence intervals of y. The development of a program for optimization was an important part of work presented in the thesis. The optimization was in the form of linear programming where the regression function of y was optimized but constrained by the latent variables and upper and lower limits of at least one of the x-variables.
The purpose was not to focus only on the methods but also to apply the methods on real data from production and characterization of cement for prediction of the quality and the properties of cement. Two investigations were performed on pure observations of X and Y. In the first investigation three PLS-models were evaluated for prediction of the three properties; amount of water required to achieve standard consistency, setting time and compressive strength at 1 day from production conditions in a cement mill. In the second investigation, a PLS-model was evaluated for predictions of compressive strength up to 28 days from characteristics of the cement. The observation X was made up by four submatrices representing three different types of characteristics, the first one representing the mineralogy of the clinker part of the cement, the second the particle size distribution and the third and fourth superficial microstructure of cement. The mineralogy characterized by X-ray diffraction (XRD) analysis could be related to the production conditions in the cement kiln. The particle size distribution and superficial microstructure, the latter characterized by thermogravimetric analysis, could be related to the production conditions in a cement mill.
From the observation X matrix an artificial observation matrix was made for predicting potential compressive strength of clinker from the mineralogy. The original variation in the mineralogy in the new artificial matrix was maintained, while the other variables were kept constant and equal to their mean values. The mineralogy was represented by XRDcurves, which are characterized as spectral data, in two selected 2 _ ranges. Further, the spectral data were optimized to min and max potential compressive strength. In addition, they were interpreted qualitatively with respect to the variation in the mineralogy. The spectral data from thermogravimetric analysis included in PLS, were the differential form of a mass loss curve recorded during the analysis.
In addition to examining the influence of submatrices or blocks to on the properties by sensitivity analysis, multi-block regression methods were applied. By application of multi-block methods, the part of the characteristics or the microstructure that influences the properties could be found directly from the regression analysis.
Finally, the production conditions in a cement mill and a cement kiln were optimized to achieve optimal cement properties. Amount of water required to achieve standard consistency and setting time were predicted from production conditions in a cement mill, and potential compressive strength of clinker up to 28 days were predicted from production conditions in a cement kiln. The production conditions were optimized to achieve min and max values of the properties. The microstructure or the characteristics were predicted from the optimal production conditions to explain the influence of the production conditions on the properties mechanistically and chemically.
The main contributions in the form of papers in this thesis are (the roman numerals refer to the list of papers at the end to this chapter):
1. Developing methods for model-based optimization based on PLS, sensitivity analysis, prediction and linear programming. A case of demonstration was optimizing compressive strength of cement from variation of particle size distribution of cement [I-III]
2. Modelling compressive strength up to 28 days of cement on the characteristics of cement, predicting and optimizing potential compressive strength from the mineralogy of clinker [IV-V]
3. Presenting the principles and application of multi-block methods in cement production [VI-VII]
4. Optimizing production conditions to achieve optimal cement properties [VIII-IX]
The use of multivariate data analysis, sensitivity analysis and model-based optimization is very useful in the design and manufacture of cement. The methods enabled tailoring of cement aiming at target values of properties like compressive strength, setting time and initial flow properties. The tailoring can be based on the variables representing the characteristics of cement as well as the variables representing the production conditions in the cement kiln and the cement mill. The max values of the compressive strength at 1 and 28 days were achieved by optimizing the production conditions in the kiln giving the optimal mineralogy for achieving max strengths. Optimal values of cement properties at early ages were achieved by optimizing the production conditions in the cement mill giving the optimal superficial microstructure.