233
relations, and it cannot extract non-linear and complicated correlations. The low accuracy of
the ReliefF method can be explained in the same way. It does not work based on kernel trick
so the complicated dependencies cannot be detected.
Finally, BAHSIC uses the kernel trick. This improves the accuracy of the model, but
because HSIC 1 is an estimation of the HSIC 0 with no bias and high variance, it cannot
extract relations precisely. Hence, the resulted accuracy is slightly lower than that obtained
by S-BAHSIC. By using the proposed feature selection method on MWD data, the most
informative attributes are identified and selected which leads to development of an accurate
model. This model predicts mineral content level of a hole based on its MWD data which is
available in large quantity and at lower cost as compared to conventional ore grade estimation methods like core drilling and drill cuttings analysis. This model is time-efficient as well,
which reduces the time required to estimate the ore grade.
6 CONCLUSION
In this study, the ore grade range of a hole is predicted based on MWD data. To prepare
the input data, the raw MWD data and ore grade percentage of each hole are extracted.
Since the input data affect the accuracy and performance of a model, MLTs are used for
pre-processing the recorded raw data. Applied noise elimination method removes faulty
and noisy samples. Drill parameters at various depth ranges are considered features and
S-BAHSIC is proposed to identify the most informative features supervisedly. A kernel SVM
model is developed based on the reduced MWD data and the 5-fold cross validation method
is used to obtain the classification accuracy. The evaluations imply that S-BAHSIC leads to
more classification accuracy in comparison with the other common feature selection methods (e.g. mRMR, ReliefF, BAHSIC). Hence, the mineral content level of a borehole could
be predicted using the MWD data which are produced in enormous quantity and at a lower
cost. Also, the proposed model is time efficient that makes it advantageous over traditional
ore grade estimation methods.
ACKNOWLEDGEMENT
The authors are grateful for valuable inputs and support from the staff and management
of the Leveäniemi mine. Epiroc Rock Drills AB are also acknowledged for their input to
the project. Vinnova, the Swedish Energy Agency and Formas are acknowledged for partly
financing this project through the SIP-STRIM program. The authors would also like to
thank the support from the project of SLIM funded by the European Union’s Horizon
2020 research and innovation program under grant agreement nº 730294. Finally, CAMM is
acknowledged for financing parts of this study.
REFERENCES
Barton, N., Lien, R., Lunde, J., 1974. Engineering classification of rock masses for the design of tunnel
support. Rock mechanics, pp. 189–236.
Bieniawski, Z.T., 1995. Classification of rock masses for engineering: the RMR system and future
trends. Rock Testing and Site Characterization, pp. 553–573.
Corinna, C., Vapnik, Vladimir, N., 1995. Support-vector networks. Machine Learning., p. 273–297.
Deere, D., Miller, R.D., 1966. Engineering classification and index properties for intact rock. Univ. of
Illinois, Tech. Rept., pp. 65–116.
Dowell, L., Mills, A., Matt, L., 2006. Drilling Data Acquisition. In: Drilling Engineering. s.l.: Society of
Petroleum Engineers, p. 647–685.
Gao, T., Cao, J., Zhang, M.Q.J., 2006. Lithology recognition during oil well drilling based on fuzzyadaptive hamming network. s.l., s.n., p. 574–578.
relations, and it cannot extract non-linear and complicated correlations. The low accuracy of
the ReliefF method can be explained in the same way. It does not work based on kernel trick
so the complicated dependencies cannot be detected.
Finally, BAHSIC uses the kernel trick. This improves the accuracy of the model, but
because HSIC 1 is an estimation of the HSIC 0 with no bias and high variance, it cannot
extract relations precisely. Hence, the resulted accuracy is slightly lower than that obtained
by S-BAHSIC. By using the proposed feature selection method on MWD data, the most
informative attributes are identified and selected which leads to development of an accurate
model. This model predicts mineral content level of a hole based on its MWD data which is
available in large quantity and at lower cost as compared to conventional ore grade estimation methods like core drilling and drill cuttings analysis. This model is time-efficient as well,
which reduces the time required to estimate the ore grade.
6 CONCLUSION
In this study, the ore grade range of a hole is predicted based on MWD data. To prepare
the input data, the raw MWD data and ore grade percentage of each hole are extracted.
Since the input data affect the accuracy and performance of a model, MLTs are used for
pre-processing the recorded raw data. Applied noise elimination method removes faulty
and noisy samples. Drill parameters at various depth ranges are considered features and
S-BAHSIC is proposed to identify the most informative features supervisedly. A kernel SVM
model is developed based on the reduced MWD data and the 5-fold cross validation method
is used to obtain the classification accuracy. The evaluations imply that S-BAHSIC leads to
more classification accuracy in comparison with the other common feature selection methods (e.g. mRMR, ReliefF, BAHSIC). Hence, the mineral content level of a borehole could
be predicted using the MWD data which are produced in enormous quantity and at a lower
cost. Also, the proposed model is time efficient that makes it advantageous over traditional
ore grade estimation methods.
ACKNOWLEDGEMENT
The authors are grateful for valuable inputs and support from the staff and management
of the Leveäniemi mine. Epiroc Rock Drills AB are also acknowledged for their input to
the project. Vinnova, the Swedish Energy Agency and Formas are acknowledged for partly
financing this project through the SIP-STRIM program. The authors would also like to
thank the support from the project of SLIM funded by the European Union’s Horizon
2020 research and innovation program under grant agreement nº 730294. Finally, CAMM is
acknowledged for financing parts of this study.
REFERENCES
Barton, N., Lien, R., Lunde, J., 1974. Engineering classification of rock masses for the design of tunnel
support. Rock mechanics, pp. 189–236.
Bieniawski, Z.T., 1995. Classification of rock masses for engineering: the RMR system and future
trends. Rock Testing and Site Characterization, pp. 553–573.
Corinna, C., Vapnik, Vladimir, N., 1995. Support-vector networks. Machine Learning., p. 273–297.
Deere, D., Miller, R.D., 1966. Engineering classification and index properties for intact rock. Univ. of
Illinois, Tech. Rept., pp. 65–116.
Dowell, L., Mills, A., Matt, L., 2006. Drilling Data Acquisition. In: Drilling Engineering. s.l.: Society of
Petroleum Engineers, p. 647–685.
Gao, T., Cao, J., Zhang, M.Q.J., 2006. Lithology recognition during oil well drilling based on fuzzyadaptive hamming network. s.l., s.n., p. 574–578.