The Use of Machine Learning Techniques for the Extraction of Process Knowledge from Industrial Flotation Plants

In flotation processes the structure of the froth phase contains a wealth of information regarding the behaviour of the plant. Various structural features of the froth phase provide an indication of the froth viscosity, froth stability, mineral content, bubble size, etc., which can all in turn be related to the performance of the plant. In this paper the application of machine learning techniques to exploit information from digital images of the froth phase of industrial flotation plants is discussed. This includes both connectionist techniques to identify control decisions necessary to maintain optimal operation of the plant, as well as symbolic methods, such as induction techniques. Both approaches are used to classify froth structures based on statistical features derived from digitized images of the froth surfaces. It was found that backpropagation algorithms perform significantly better than either non-incremental or incremental induction. Backpropagation algorithms requires significant user input in order to optimally train the neural network as opposed to induction, which requires little user input. In contrast to back propagation, decision tree induction produces easily understandable decision trees which can be incorporated into a rule base or an expert system. Finally, incremental induction provides a simple means to smoothly and continuously adapt induced rules to follow changing process conditions.