228
specifications are considered as “features”. The study identifies the most informative features
from the input MWD data and develops a classification model that can precisely predict the
ore grade level of each hole. The recorded MWD data and the ore grade classes are the input
and output of the model, respectively.
2.1 Test site description
The MWD data and ore grade classes are collected from the Leveäniemi mine. The mine
is an open pit operation in the Svappavaara mining area in northern Sweden and is owned
and operated by Luossavaara-Kiirunavaara AB (LKAB). The open pit mining method allows
maximum ore extraction from the rock with a high degree of safety. The planned annual
production of the mine is 12 million tonnes of magnetite ore with an average grade of 44%
Fe. The bench height in the pit is 15 m. All production holes in the mine are drilled with
fully mechanized SmartROC D65 down-the-hole (DTH) drill rigs. All rigs are equipped with
MWD systems that monitor all relevant drill parameters at predefined length intervals during
the drilling. These recorded MWD data from six Epiroc D65 drill rigs are used for the analysis
in this paper. The rigs have the intelligence and power to drill production blast, pre-split and
buffer holes. The study’s samples come from 14340 holes; the measured drilling parameters in
each hole are penetration rate, percussive pressure, feed pressure and rotation pressure.
2.2 Classification method
After proposing a supervised feature selection method to identify the informative properties
of data, a classification model is trained by the data extracted from 90% of the whole data
(12906 holes) for two purposes:
• To evaluate and compare performance of the suggested feature selection method based on
train and test accuracies
• To predict the ore grade class of holes with recorded MWD data in future
Support vector machine (SVM) is the classification method used to model the input MWD
data (Cortes et al. 1995). An SVM model is a supervised classification method which divides
the points in space, so that the samples of the dispersed categories are classified by a clear gap
that is as wide as possible. Also, kernel functions are applied to analyze correlations between
input and output data in a higher dimension space which leads to more accurate results
(Shawe-Taylor & Cristianini 2004). In this study, Gaussian kernel method is used for high
dimensional analysis (Nadaraya 1964). This kernel method consider the input data in a space
with infinite dimension and gives the opportunity to identify complex dependencies.
2.3 Model evaluation
The k-fold cross validation method is used to avoid over or under fitting (Kohavi Ron et al.
1995). This model validation technique assesses how the results of a statistical analysis generalize to an independent dataset. Accuracy is the proportion of true predicted labels among
the whole number of test samples and it is used to evaluate the feature selection method.
3 DATA PRE-PROCESSING
Since the input data have significant effects on the accuracy of final model, applying pre-processing
methods leads to higher accuracy. Noisy samples and non-informative features are removed from
recorded raw data using noise elimination and feature selection methods respectively.
3.1 Noise elimination
Raw MWD data include noisy or faulty data (Ghosh et al. 2015, Van Eldert et al. 2018). For
more reliable results, it is necessary to use a noise reduction method and remove unreliable
specifications are considered as “features”. The study identifies the most informative features
from the input MWD data and develops a classification model that can precisely predict the
ore grade level of each hole. The recorded MWD data and the ore grade classes are the input
and output of the model, respectively.
2.1 Test site description
The MWD data and ore grade classes are collected from the Leveäniemi mine. The mine
is an open pit operation in the Svappavaara mining area in northern Sweden and is owned
and operated by Luossavaara-Kiirunavaara AB (LKAB). The open pit mining method allows
maximum ore extraction from the rock with a high degree of safety. The planned annual
production of the mine is 12 million tonnes of magnetite ore with an average grade of 44%
Fe. The bench height in the pit is 15 m. All production holes in the mine are drilled with
fully mechanized SmartROC D65 down-the-hole (DTH) drill rigs. All rigs are equipped with
MWD systems that monitor all relevant drill parameters at predefined length intervals during
the drilling. These recorded MWD data from six Epiroc D65 drill rigs are used for the analysis
in this paper. The rigs have the intelligence and power to drill production blast, pre-split and
buffer holes. The study’s samples come from 14340 holes; the measured drilling parameters in
each hole are penetration rate, percussive pressure, feed pressure and rotation pressure.
2.2 Classification method
After proposing a supervised feature selection method to identify the informative properties
of data, a classification model is trained by the data extracted from 90% of the whole data
(12906 holes) for two purposes:
• To evaluate and compare performance of the suggested feature selection method based on
train and test accuracies
• To predict the ore grade class of holes with recorded MWD data in future
Support vector machine (SVM) is the classification method used to model the input MWD
data (Cortes et al. 1995). An SVM model is a supervised classification method which divides
the points in space, so that the samples of the dispersed categories are classified by a clear gap
that is as wide as possible. Also, kernel functions are applied to analyze correlations between
input and output data in a higher dimension space which leads to more accurate results
(Shawe-Taylor & Cristianini 2004). In this study, Gaussian kernel method is used for high
dimensional analysis (Nadaraya 1964). This kernel method consider the input data in a space
with infinite dimension and gives the opportunity to identify complex dependencies.
2.3 Model evaluation
The k-fold cross validation method is used to avoid over or under fitting (Kohavi Ron et al.
1995). This model validation technique assesses how the results of a statistical analysis generalize to an independent dataset. Accuracy is the proportion of true predicted labels among
the whole number of test samples and it is used to evaluate the feature selection method.
3 DATA PRE-PROCESSING
Since the input data have significant effects on the accuracy of final model, applying pre-processing
methods leads to higher accuracy. Noisy samples and non-informative features are removed from
recorded raw data using noise elimination and feature selection methods respectively.
3.1 Noise elimination
Raw MWD data include noisy or faulty data (Ghosh et al. 2015, Van Eldert et al. 2018). For
more reliable results, it is necessary to use a noise reduction method and remove unreliable