AI Case Study
Researchers investigate a cost-effective machine learning method for identifying gold from waste during ore sorting at an AngloGold Ashanti mine
Researchers identified an cost-effective machine learning method for identifying gold from waste material during ore sorting, using RGB camera images and hyperspectral measurements. The method was tested using data from an AngloGold Ashanti mine in Australia.
Industry
Basic Materials
Mining And Metals
Project Overview
The study concerns ore mined at AngloGold Ashanti's Sunrise Dam Gold Mine in Western Australia, and was conducted by researchers at The University of Western Australia and Anglo. It "aims to develop an effective machine learning method for the classification of waste and gold-bearing particles from SDGM. The study used RGB images and hyperspectral signatures of particles and tested the performances of three different machine learning algorithms, namely the naïve Bayes (NB) classifier, majority decision table (MDT), and support vector machine (SVM).
With their practical application to ore sorting for gold production in mind, the use of these algorithms is extended in three steps. The first step is feature selection, where various features are extracted from the hyperspectral and RGB image data and four different feature selection algorithms are applied to learn the optimal set of features for discriminating particles of waste from ore. This is achieved by extracting a large number of features from the particle images and combining them with the hyper-spectral reflectances measured by a spectrometer, and then removing redundant and uninformative features using a feature selection technique. Four different types of feature selection algorithms were tested, namely information gain, ReliefF, correlation-based feature selection, and Consistency.
After feature selection, the second step is to compare the performance of different machine learning and feature selection techniques for ore sorting by comparing the techniques based on the estimated impact on nominal profit lost due to sorting error. The third and final step is evaluating the increase in profitability gained using cost-sensitive classification methods against cost-blind methods. The three machine learning algorithms, each of which was trained using both cost-sensitive and cost-blind methods, were compared. Misclassification costs were calculated from the estimated nominal profit lost due to sorting error, where nominal profit is the revenue from the sale of refined ore minus the processing costs."
Reported Results
Nominal profit loss from sorting errors was used as the evaluation metric for the performance of each classifier. The support vector machine obtained the best results, based on nominal profit lost due to sorting error (sum of false negatives and positives). "Although the cost-blind SVM exhibited a greater discriminating ability, it rejected too many gold-bearing particles to be economically feasible. A cost-blind support vector machine achieved an ore acceptance rate of 84 % and a waste rejection rate of 87 %, which resulted in $0.98 nominal profit lost per tonne of crushed rock particles. Cost-sensitive training reduced the nominal profit lost to $0.34 per tonne, undercutting the costs associated with refining all particles by $0.24 per tonne."
Technology
"Feature selection was applied to groups of the representative features and resulting feature subsets were evaluated using three machine learning algorithms, namely a support vector machine, a naïve Bayes classifier, and a majority decision table, to identify a highly informative subset of features. Cost-sensitive training was used to minimise the nominal profit lost due to sorting error based on real cost values from the milling process, with the aim of economically balancing the ore acceptance rate with the waste rejection rate."
Function
R And D
Core Research And Development
Background
"Ore sorting increases the grade of an ore feed stream by separating very low- grade particles (‘waste’) from those containing higher concentrations of the desired mineral (‘ore’). Operating costs can be decreased by not further processing waste particles, thus the accuracy and efficiency at which an ore sorter operates has a significant economic impact for the mining industry. The Sunrise Dam Gold Mine is located in the eastern Goldfields in Western Australia. It contains the largest known gold deposit (12.5 Moz) within the Laverton Greenstone Belt."
Benefits
Data
"A total of 169 waste particles and 19 ore particles were sourced from the open pit and underground operations... RGB images captured with a commodity camera and hyperspectral measurements of 188 assayed particles from the Sunrise Dam Gold Mine. Advanced feature extraction methods were employed to capture visual cues such as texture and colour from the RGB images, which were combined with hyperspectral features to give nine types of representative features."