| Summary: | Includes bibliographical references (p. 237-244). === Numerous, machine vision systems for froth flotation have been developed over the last ten years; however, there are many aspects of the systems, that still require further development before they become one of the standard instruments present on industrial flotation operations. This thesis aims to address these problems by developing improved measurement techniques and showing how these measurements can be used to model the concentrate grad e of the flotation cell being monitored in a manner which is, directly usable by plant personnel. This thesis, presents an improvement to the watershed algorithm for the measurement of bubble sixe distribution in flotation froths. Unlike the standard watershed algorithm, it is able to measure accurate bubble size distributions when both large and tiny bubbles are present in a flotation froth image. Flotation froths with “dynamic bubble size distribution s” are introduced and methods of reducing the high dimensional bubble size distribution data associated with them are discussed. A method of using characteristic histograms of frequently occurring bubble size distributions is introduced and shown to be an appropriate method to use. A number of standard texture measures are best suited to the classification of flotation froth images. Results show that the Fourier ring and texture spectrum based features, perform well whilst having a relatively small computational cost for classifying new images. Video footage from selected industrial operations has been used for the development of improved algorithms for the measurement of froth surface descriptors. Analyses of the relationship, between froth velocity, bubble size, froth class and concentrate grade are made. The results show that it possible to use a unified approach to model the concentrate grade, irrespective of the site on which the measurements are made. Results from three industrial case studies show that bubble size and texture measures can be used to identify froth classes. Furthermore the combination of froth classes and froth velocity information is shown to consistently account for the most variation in the data when the concentrate grade is modelled using a linear combination of these two measurements.
|
|---|
