An IsaMilla™ Soft Sensor based on Random Forests and Principal Component Analysis

Abstract

Online measurement of particle size is vital to the development of advanced control systems for comminution processes. Horizontal stirred mills, such as the IsaMill, are designed for more efficient ultrafine grinding and have made significant inroads in the mineral processing industries since their introduction more than a decade ago. Despite their energy efficiency, significant improvement is possible via more efficient control of these mills. Advanced control generally requires online information on the key performance variables of the mill. In this regard, measurement of the particle size in the mill is problematic. However, this problem can be addressed by use of soft sensors, whereby the particle size can be estimated from the measurements of other process variables. In this investigation, such a soft sensor is developed for online estimation of particle size on an industrial IsaMill in Western Australia. The sensor consists of a random forest model that uses operational variables measured online as predictors to estimate the P 80 particle size of the mill. Principal component analysis is used in conjunction with the random forest to enable it to assess the similarity of new process measurements to the data in its training data base. When the new data exceed a Hotelling's T 2 or a prediction error or Q-index threshold, recalibration of the model is automatically performed.