Mining Goes Digital – Mueller et al. (Eds)
© 2019 Taylor & Francis Group, London, ISBN 978-0-367-33604-2
94
Multivariate Gaussian process for distinguishing geological units
using measure while drilling data
K.L. Silversides & A. Melkumyan
Australian Centre for Field Robotics, The University of Sydney, Sydney, NSW, Australia
ABSTRACT: Banded iron formation hosted iron ore deposits are typically stratigraphically
modelled using exploration drilling. This drilling has a large horizontal spacing, resulting in a
coarse model resolution. Production blast holes are drilled at a higher horizontal resolution,
but rarely include detailed information. Measure While Drilling (MWD) data is available,
however it only provides information about the relative hardness or strength. There is often
a large overlap in the MWD data for adjacent rock units, making manual classification difficult. A Gaussian Processes model was used to automatically label MWD points from two
adjacent geological units. For a shale to ore contact, the trained GP had an accuracy of 79%.
When distinguishing between two ore units, the GP had an accuracy of ∼96% for the library
and ∼83% for the cross-validation data. Therefore this method can provide additional detail
about the location of the contact between geological units for modelling.
1 INTRODUCTION
The banded iron formation hosted iron ore deposits of the Hammersley Region of Western
Australia are typically stratigraphically modelled using exploration drilling (De-Vitry et al.
2010; Jones et al. 1973). While these data sources are dense down hole, ∼0.1–2 m resolution,
the exploration holes are typically drilled with a horizontal spacing of ∼50 m. This results in
a coarse resolution in the modelling of the contacts between different geological units, with
the accuracy decreasing away from the exploration holes. This presents an opportunity to
increase the local accuracy of these models using production data.
Production blast holes are drilled at a much higher horizontal resolution (∼5 m), but the
detailed information collected on the exploration drill holes is rarely obtained. Measure
while drilling (MWD) data is available, however it can only provide information about the
relative hardness or strength of different geological units. MWD parameters include penetration rate (PR), force on bit (FOB) and torque. The data is collected at 10 cm intervals. PR
alone cannot be used to distinguish units, as it is also dependent on the energy inputs. An
increase in PR can be due to either a softer rock or an increase in the energy provided by the
drill, i.e. through increasing torque or FOB. Several measures, such as adjusted penetration
rate (APR) (Zhou et al. 2011) and specific energy of drilling (SED) (Teale 1965), have been
created to combine MWD parameters. However these reduce the dimensionality of the data
to a single quantity, causing a loss of information. When the natural logarithms of the MWD
parameters are plotted for two adjacent rock units, it is difficult to manually classify the
regions due to the large overlap between the units. However, a machine learning algorithm
can distinguish regions that are dominated by a single unit.
This study uses data from stratified BIF-hosted iron ore mines that contain banded layers of shale and mineralised BIF (Thorne et al. 2008; Dalstra & Rosiere 2008). The ore is
consistently harder than the larger shale bands, however the differences in the MWD values
between different ore units is much harder to identify. Our MWD based GP identification
method is tested in two cases. The first is the shale to ore contact at a typical Marra Mamba
style deposit containing the West Angelas Shale and Mount Newman Member iron ore. The
© 2019 Taylor & Francis Group, London, ISBN 978-0-367-33604-2
94
Multivariate Gaussian process for distinguishing geological units
using measure while drilling data
K.L. Silversides & A. Melkumyan
Australian Centre for Field Robotics, The University of Sydney, Sydney, NSW, Australia
ABSTRACT: Banded iron formation hosted iron ore deposits are typically stratigraphically
modelled using exploration drilling. This drilling has a large horizontal spacing, resulting in a
coarse model resolution. Production blast holes are drilled at a higher horizontal resolution,
but rarely include detailed information. Measure While Drilling (MWD) data is available,
however it only provides information about the relative hardness or strength. There is often
a large overlap in the MWD data for adjacent rock units, making manual classification difficult. A Gaussian Processes model was used to automatically label MWD points from two
adjacent geological units. For a shale to ore contact, the trained GP had an accuracy of 79%.
When distinguishing between two ore units, the GP had an accuracy of ∼96% for the library
and ∼83% for the cross-validation data. Therefore this method can provide additional detail
about the location of the contact between geological units for modelling.
1 INTRODUCTION
The banded iron formation hosted iron ore deposits of the Hammersley Region of Western
Australia are typically stratigraphically modelled using exploration drilling (De-Vitry et al.
2010; Jones et al. 1973). While these data sources are dense down hole, ∼0.1–2 m resolution,
the exploration holes are typically drilled with a horizontal spacing of ∼50 m. This results in
a coarse resolution in the modelling of the contacts between different geological units, with
the accuracy decreasing away from the exploration holes. This presents an opportunity to
increase the local accuracy of these models using production data.
Production blast holes are drilled at a much higher horizontal resolution (∼5 m), but the
detailed information collected on the exploration drill holes is rarely obtained. Measure
while drilling (MWD) data is available, however it can only provide information about the
relative hardness or strength of different geological units. MWD parameters include penetration rate (PR), force on bit (FOB) and torque. The data is collected at 10 cm intervals. PR
alone cannot be used to distinguish units, as it is also dependent on the energy inputs. An
increase in PR can be due to either a softer rock or an increase in the energy provided by the
drill, i.e. through increasing torque or FOB. Several measures, such as adjusted penetration
rate (APR) (Zhou et al. 2011) and specific energy of drilling (SED) (Teale 1965), have been
created to combine MWD parameters. However these reduce the dimensionality of the data
to a single quantity, causing a loss of information. When the natural logarithms of the MWD
parameters are plotted for two adjacent rock units, it is difficult to manually classify the
regions due to the large overlap between the units. However, a machine learning algorithm
can distinguish regions that are dominated by a single unit.
This study uses data from stratified BIF-hosted iron ore mines that contain banded layers of shale and mineralised BIF (Thorne et al. 2008; Dalstra & Rosiere 2008). The ore is
consistently harder than the larger shale bands, however the differences in the MWD values
between different ore units is much harder to identify. Our MWD based GP identification
method is tested in two cases. The first is the shale to ore contact at a typical Marra Mamba
style deposit containing the West Angelas Shale and Mount Newman Member iron ore. The