Real-time prediction of free lime in cement clinker using support vector machine algorithm

Free lime content is an important quality parameter in the production of clinker, targeted between 0.5% to 1.5%. Existing studies have tried to predict absolute free lime content, using soft sensors data, with limited success due to the complexity of the clinker burning process. Subject matter experts believe that, instead of predicting the free lime absolute value, predicting free lime quality as good, over-burn, or under-burn is more practically beneficial. This study aims to predict the free lime clinker quality as good or under-burn by leveraging data mining methods and machine learning techniques. Seven months of hourly data pertinent to rotary kiln feed chemistry and operation parameters were collected from a real operational cement plant. Classification models were built using support vector machine (SVM). The SVM produced a sensitivity value of 0.998 for good clinker class and 0.865 for under-burn class with an accuracy of 96%. The availability of these predictions in real time can help plant operators to avoid under-burning and over-burning. Such insights will assist relevant cement plants to reduce off-specification products, coal usage, production cost, and carbon emissions.