Machine Learning Strategies for Control of Flotation Plants

Abstract Although flotation processes are notoriously difficult to model from first principles, knowledge-based systems can be used to great advantage to monitor and control plants, provided that process knowledge can be captured effectively on the plant. By making use of machine learning techniques the features of the surface froths of flotation cells can be used to construct representations of the behaviour of a plant. Two probabilistic decision tree methods and a backpropagation neural net were all equally capable of classifying the different froths at least as well as a human expert. Explicit decision trees were derived, relating froth characteristics to froth surface structures. Relatively sharply clustered Sammon maps of froth structures were obtained, allowing good visualisation of multidimensional flotation data.