In the minerals industry numerous problems cannot be solved by
conventional mathematical models owing to their complexity or a lack of
phenomenological understanding. Neural networks provide one way of
mapping the ill-defined relations between process variables and functions
for such ill-defined problems. Consequently, processes such as leaching
and froth flotation are mostly controlled in an empirical way by using
rules of thumb. In addition, these processes involve so many independent
and dependent variables that the plant operator finds it difficult to
visualise or even observe a change in process conditions. The structure of
froths developed on the surfaces of industrial scale froth flotation cells has
a significant effect on both the grade and recovery of valuable minerals in
the concentrate. Although these effects are well known at the process
operational level, where considerable heuristic knowledge is available,
little work has been reported on a detailed characterisation of the
mechanisms and the visual characteristics of the surface froth. Recent
results from an on-line observation of froths in several plants have proved
the relationships between image features representing froth behaviour and
metallurgical results. It will be shown how supervised and unsupervised
neural nets are being used on operating plants to interpret computer vision
data. In froth flotation the operator is supposed to visually observe
process changes from the appearance of the froth, which is an
unreasonable demand under industrial conditions. The system described
here determines textural parameters on-line, and tracks the changes in
process conditions via a Self-Organising Map (SOM) neural net. This
monitoring system warns the operator about fluctuations in reagent
addition, and gives an idea of the type of froth encountered. In another
example, changes in the mineralogical characteristics of gold ores are
represented on an SOM map, based on the diagnostic leaching behaviour
of such ores.