Multivariate Image Analysis in Mineral Processing - Advanced Control and Supervision of Mineral Processing Plants - page 137

tion problem, further investigations of these issues are required. Overall, the results
were found to be very promising and support future work on this subject.
3.6 Conclusions
Multivariate imaging techniques were presented in this chapter and applied to the
development of three novel on-line vision sensors for use in the mineral process-
ing field. These sensors aim at: (1) estimating flotation froth mineral grade; (2)
froth health monitoring for appropriate reagent dosage; and (3) on-line estimation
of ROM ore lithotype composition on conveyor belts. Both MIA, based on well
known PCA and PLS, and MRA, relying on wavelet decomposition, have been pre-
sented and used for spectral/textural analysis of process images ( i.e. , flotation froth
and ROM rock mixtures). These techniques were shown to be very efficient for
extracting the stochastic features of process images relevant for monitoring and/or
predicting process conditions or product quality. As opposed to traditional auto-
matic inspection problems in the manufacturing industries, the relevant features of
images collected in process industries are usually unknown apriori . A machine
vision framework based on multivariate methods was also presented for identify-
ing these relevant image features, and for using them in mineral process modeling,
monitoring, and advanced control.
The first case study served to illustrate a multivariate image regression problem
using spectral image information only. It was shown that predictions of zinc min-
eral grade obtained using RGB froth images were in a good agreement with the
trends of the laboratory assays. The long-term robustness of such a sensor to light-
ing conditions, mineral feed and other environmental conditions still needs to be
demonstrated, but the results obtained at the Laronde plant were found encouraging
and granted further investigation by the research consortium COREM (unpublished
work). Among other issues to be investigated are (1) the optimal positioning of the
camera for a given flotation cell or column to make sure that images are representa-
tive of the entire froth surface, (2) selecting the best flotation unit(s) within a bank
on which to install the camera. Fast on-line froth grade measurements could be in-
tegrated into a grade control loop manipulating reagent dosage when mineral feed
and other disturbances occur. It could also be used to decide how to distribute the
reagents across the flotation bank for optimal use of chemicals.
An unsupervised classification problem using both spectral and textural image
features was proposed as the second case study, again using the froth flotation pro-
cess. The concept of on-line “froth health” was presented for detecting inappropri-
ate reagent dosage for a given ore feed composition. A new bubble sizing algorithm
based on wavelet decomposition ( i.e. , wavelet size signatures) was discussed. It is
robust to noise and to variations in lighting conditions, and does not require direct
image segmentation. After combining with some froth color features ( i.e. , clear win-
dows and black holes), a PCA monitoring scheme was developed and was shown to
be able to detect the onset of froth collapse, related to over-dosage of reagents with