Efficient machine learning model to predict fineness, in a vertical raw meal of Morocco cement plant

Abstract

Soft sensor enables computing parameters that can be physically impossible to measure. This work aims to develop a soft sensor for raw meal fineness in a vertical roller mill of a cement plant. In previous research, some key indicators of the process were ignored. It is well known that vertical raw meal is feeded with heterogenous material that impacts raw meal fineness, and therefore, can deteriorates prediction accuracy of the soft sensor. To deal with this constraint, one-year history of industrial data was exploited and processed by a robust multivariate outlier identification method, followed by process expert outlier verification. Then a particular attention was given to process and material quality parameters using a statistical method to test parameters significance on fineness based on data variance calculation. Eight process and four quality parameters were retained to develop robust nonlinear model using artificial neural network: Bayesian regularization and Levenberg Marquart, then compared with Support vector regression with different acquisition function and kernels to tune and select optimum model hyperparameters. The best model performance was obtained with the Bayesian regularization algorithm.