Mining Goes Digital – Mueller et al. (Eds)
© 2019 Taylor & Francis Group, London, ISBN 978-0-367-33604-2
226
Ore grade prediction using informative features of MWD data
S. Liaghat, A. Gustafson, D. Johansson & H. Schunnesson
Division of Mining and Geotechnical Engineering, Luleå University of Technology, Sweden
ABSTRACT: Detailed knowledge of the content and geometrical variation of ore grade is
essential in mining operations for production planning and economic analysis. Common ore
grade specification methods, sampling and analysis are costly and time consuming. Measurement While Drilling (MWD) technique can directly extract grade information from the
drilling process increasing data resolution and reducing cost.
This study introduces a supervised feature selection method based on the Hilbert-Schmidt
independence criterion to increase the accuracy of the results and decrease processing time.
Potential of the method for recognizing the most effective and non-repetitive dimensions of
input data has also been investigated. By exploiting the lower dimension data, a classification
model is developed to map the parameter values to ore grade levels.
Evaluation of the model using MWD data from LKAB’s Leveäniemi mine proved the
effectiveness of the proposed feature selection and classification method.
1 INTRODUCTION
The economic evaluation of possible mining operation for exploring new ore bodies is critical
but requires accurate ore grade information (Wellmer et al. 2010). Two common methods for
specifying the ore grade are core drilling and drill cutting analysis. The results of core drilling form the basis for investments and long-term production planning. Since the sampling
holes are quite far apart, low-resolution data and uncertainty about un-cored areas between
holes will be observed (Starr & Ingleton, 1992). Drill cutting analysis extracts information
from all production bore holes, but it only provides information on the average grade of a
sampled hole (Neff 2003). Both methods provide low resolution data and are costly or time
consuming.
Measurement While Drilling (MWD) is a technique of gathering drilling data, which can
be employed to characterize the mechanical properties of the penetrated rock mass (Dowell
et al. 2006). A holistic analysis of parameters from MWD data gives helpful information on
the penetrated rock mass and can be used for several purposes (e.g. to delineate ore from
waste). If the parameters include the mechanical properties of the rock, the geology and ore
grade levels can be estimated. (Gao et al. 2006). It is also possible to predict other important
properties like fragmentation and blast ability based on drilling parameters (Barton et al.
1974, Bieniawski 1995, Deere & Miller 1966).
To facilitate the extraction of different properties from the recorded MWD data, it is necessary to select a proper calculation technique. Machine learning techniques (MLTs) such as
neural network, boosting and feature selection are precise and cost efficient for this purpose.
MLTs are less time consuming than other experimental methods used for parameter estimation (Kadkhodaie-Ilkhchi et al. 2010). Kadkhodaei et al. (2010) have developed a model to
predict rock mass properties by applying pattern recognition (PR) and MLTs to raw data
recorded during hole drilling. As the quality of the input data has a significant effect on the
model performance, this study aims to improve the results by implementing pre-processing
techniques like noise elimination and feature selection on the raw input data. With feature
selection, certain features (parameters) that are either redundant or only relevant to the other
© 2019 Taylor & Francis Group, London, ISBN 978-0-367-33604-2
226
Ore grade prediction using informative features of MWD data
S. Liaghat, A. Gustafson, D. Johansson & H. Schunnesson
Division of Mining and Geotechnical Engineering, Luleå University of Technology, Sweden
ABSTRACT: Detailed knowledge of the content and geometrical variation of ore grade is
essential in mining operations for production planning and economic analysis. Common ore
grade specification methods, sampling and analysis are costly and time consuming. Measurement While Drilling (MWD) technique can directly extract grade information from the
drilling process increasing data resolution and reducing cost.
This study introduces a supervised feature selection method based on the Hilbert-Schmidt
independence criterion to increase the accuracy of the results and decrease processing time.
Potential of the method for recognizing the most effective and non-repetitive dimensions of
input data has also been investigated. By exploiting the lower dimension data, a classification
model is developed to map the parameter values to ore grade levels.
Evaluation of the model using MWD data from LKAB’s Leveäniemi mine proved the
effectiveness of the proposed feature selection and classification method.
1 INTRODUCTION
The economic evaluation of possible mining operation for exploring new ore bodies is critical
but requires accurate ore grade information (Wellmer et al. 2010). Two common methods for
specifying the ore grade are core drilling and drill cutting analysis. The results of core drilling form the basis for investments and long-term production planning. Since the sampling
holes are quite far apart, low-resolution data and uncertainty about un-cored areas between
holes will be observed (Starr & Ingleton, 1992). Drill cutting analysis extracts information
from all production bore holes, but it only provides information on the average grade of a
sampled hole (Neff 2003). Both methods provide low resolution data and are costly or time
consuming.
Measurement While Drilling (MWD) is a technique of gathering drilling data, which can
be employed to characterize the mechanical properties of the penetrated rock mass (Dowell
et al. 2006). A holistic analysis of parameters from MWD data gives helpful information on
the penetrated rock mass and can be used for several purposes (e.g. to delineate ore from
waste). If the parameters include the mechanical properties of the rock, the geology and ore
grade levels can be estimated. (Gao et al. 2006). It is also possible to predict other important
properties like fragmentation and blast ability based on drilling parameters (Barton et al.
1974, Bieniawski 1995, Deere & Miller 1966).
To facilitate the extraction of different properties from the recorded MWD data, it is necessary to select a proper calculation technique. Machine learning techniques (MLTs) such as
neural network, boosting and feature selection are precise and cost efficient for this purpose.
MLTs are less time consuming than other experimental methods used for parameter estimation (Kadkhodaie-Ilkhchi et al. 2010). Kadkhodaei et al. (2010) have developed a model to
predict rock mass properties by applying pattern recognition (PR) and MLTs to raw data
recorded during hole drilling. As the quality of the input data has a significant effect on the
model performance, this study aims to improve the results by implementing pre-processing
techniques like noise elimination and feature selection on the raw input data. With feature
selection, certain features (parameters) that are either redundant or only relevant to the other