Multivariate Image Analysis in Mineral Processing - Advanced Control and Supervision of Mineral Processing Plants - page 116

features were strongly correlated with batch flotation kinetic curves and grade. The
reader is referred to Duchesne et al. [35] for more details.
3.5.2 Flotation Froth Health Monitoring using MR-MIA
This section presents a second case study of the froth flotation process, based on the
work of Liu et al. [37], aiming at illustrating an unsupervised image classification
problem using both color and textural features. Plant operators typically use their
visual appreciation of the general state of the flotation process, referred to as “froth
health” in this paper, to verify whether the dosage of the various reagents ( e.g. ,col-
lector, frother, activator, etc. ) is appropriate for the current state of the process ( e.g. ,
grade of ore feed), and to decide upon implementing control actions. For example,
overdosage of collector may lead to an overloaded froth and, eventually, to bubble
collapse and loss of separation capacity, generally requiring a significant amount of
time to recover normal operation. On the other hand, if insufficient collector is used,
the froth mineralization will be lower and loss of recovery will result. Operators
usually detect the onset of these two extreme situations by their assessment of froth
color and morphology, the later being essentially related with bubble size, distribu-
tion and shape [63, 64, 69, 70], also frequently called froth structure. Since froth
color extraction methods have been reviewed in the previous case study (Section
3.5.1), this section focuses on the extraction of froth structural features.
Over the years, several machine vision approaches were developed for moni-
toring froth state in order to overcome the lack of reproducibility in human visual
assessment and the limited time operators have to spend looking at the froth of any
given cell or column. Indeed, it is well known that qualitative visual assessment
varies from person-to-person and as a function of time, period of the day, state of
mind, etc. On the other hand, a machine vision solution provides on-line, consistent,
and quantitative measurements of froth characteristics, available to operators in the
control room for several flotation units.
The most frequently used approaches for quantifying froth structure were based
on image segmentation algorithms [64, 71-73] ( e.g. , watershed [74]) and textural
methods, such as gray Level co-occurrence matrix (GLCM) [70], [69, 75, 76] and
Fourier transform [61]. Image segmentation may lead to very precise results for
bubble size, but is computationally intensive, and sensitive to variations in lighting
conditions. A new approach was proposed by Liu et al. [37] combining both the
computational efficiency of textural methods and their robustness to lighting condi-
tions, and the ability to estimate a distribution of bubble sizes. The estimated distri-
butions are not as accurate as those provided by segmentation algorithms. However,
robust and reproducible estimates rather than accurate bubble size distributions are
required for process control purposes. The proposed approach uses both froth color
and textural features, called wavelet size signatures , for classifying froth state. These
features are obtained using the MR-MIA II algorithm discussed in Section 3.4.1.3.