Dipartimento di Ingegneria informatica, automatica e gestionale

In cement plants, any unplanned stop or slowdown in industrial production leads to non-negligible sunk and maintenance costs. Cement production requires the use of expensive equipment to ensure continuous production. This study suggests a machine learning approach to predict and advise when a cement plant component will be at risk of failure. A predictive model was built based on data available for an industrial fan of a cement plant. This component has been chosen due to the maintenance criticality of rotating items. Component’s health was monitored using parameters such as vibration, temperature, and speed values of the fan support bearings. Classification and regression algorithms have been tested. For the classification, logistic regression and boosted decision trees were applied to label whether the plant was in critical status or not. The threshold value used to tag the data refers to fan vibrations. For the regression, linear regression and decision forest regression were used to determine the residual useful life (RUL) of the fan. This model has provided significant benefits such as reducing unplanned stops, losses of production, and production in critical conditions.