Environmental Engineering Reference
In-Depth Information
order to minimize rock type recognition errors. Hence, the proposed methodology
requires a greater number of steps compared with the previous case studies due to
the higher level of complexity of this machine vision problem. This section sum-
marizes the work published by Tessier et al. [26, 78], and presents a supervised
classification problem of image features using PLS discriminant analysis or PLS-
DA.
It is well known that variations in ROM ore properties such as size, composition,
and grindability strongly affect the performance of autogeneous (AG) and semi-
autogeneous (SAG) grinding mills, such as power consumption, circulating load,
and product size distribution [79]. The Julius Kruttschnitt Mineral Research Centre
(JKMRC) has pioneered the Mine-to-Mill TM 4 optimization projects to address these
issues since 1998. In the past, most efforts to track and control these variations have
focused on developing on-line size analysis using vision systems for troubleshooting
and/or for feedforward/feedback control of crushers and grinding mills, for reducing
power consumption and avoiding mill overload. A variety of commercial vision
systems are currently available for that purpose, such as WipWare's WipFrag [80]
and Metso's VisioRock [81]. These commercial packages use some segmentation
algorithms and other traditional image processing techniques to compute ore size
distributions on-line from gray scale images.
Monitoring other ROM ore properties, such as mineral species ( i.e. , lithotype)
and grindability, is a more difficult problem and most approaches developed in the
past are either limited in the location where the measurement can be taken within the
plant or in the size and type of the ore fragments to be analyzed. Furthermore, these
approaches are generally specific to one ore property, and imaging approaches do
not consider the impact of rock surface moisture. For example, tracking of ore grind-
ability was proposed by Simkus and Dance [82] using data collected from drilling
machines to infer grindability, then following truck loads using GPS technology,
and finally modeling of stock piles, bins and hoppers to estimate the work index
of the ore incoming the concentrator plant. State estimation techniques were also
proposed for grindability [83, 84], involving detailed mechanistic models as well as
input-output process data. Machine vision approaches to mineral type recognition
have also been proposed for a high color contrast binary mixture of fine chalcopyrite
and molybdenite particles [62] as well as for ternary mixtures of manganese, iron
and alumina [85], again having very high color contrasts. A method for gold ores
having different textural characteristics was proposed by Petersen et al. [86]. Cop-
per ores were classified into seven lithotypes using several image features ( i.e. , 667
color, textural and geometrical features) and neural networks . Methods for rock tex-
ture retrieval in large databases were also found in the pattern recognition literature
[88],[89]. Finally, grading of crushed aggregates ( i.e. , civil engineering materials)
was proposed by Murtagh et al. [90] using WTA and PCA decomposition. A more
extensive literature review is available in Tessier et al. [26, 78].
The objective of this study is to develop a machine vision approach that would
classify rock types on a conveyor belt in order to estimate their proportions in the
4 Mine-to-Mill is a registered trademark of the Julius Kruttschnitt Mineral Research Centre Isles
Road, Indooroopilly, Qld Australia 4068, www.jkmrc.com.au
Search WWH ::




Custom Search