Data Modeling Techniques For A Cement Plant

Abstract

Many industries, among which the cement industry, have showed growing interest in the exploitation of its gathered data to optimize its production line. In this work, typical problems occuring in cement plant are addressed. The first one concerns the prediction of cyclones cloggings phenomena. Several methods are discussed in an attempt to solve this predictive maintenance problem. Whilst one method relies on operating points clustering via K-Means, the other one consists in modeling the problem as a binary classification task where samples close to cloggings get a value 1 and the normal samples get a value 0. After some processing to counteract the imbalanced dataset problem and a feature space reduction, the Random Forest, SVM and One-Class SVM algorithms are evaluated to conduct the classification. The second task was the prediction of the clinker quality based on some measurements inside the production line. Through the collection of raw meal quality, fuels flows and clinker quality measurements, a multivariate time series problem is established and an autoregressive model (VAR) is used in this forecasting task. In any case, the prediction performance is relatively low. Even if some alternative methods could improve the predictions, the main reasons explaining poor forecast can be found in the available dataset in which the sampling period of some key data was too low. Ultimately, the understanding of monitoring data obtained from industrial plants could result in efficiency improvements and cost reductions.