Sie sind auf Seite 1von 6

7th BALKANMINE CONGRESS – Proceedings, Prijedor 2017

DOI: 10.7251/BMC170701195A

ESTIMATION OF IRON CONTENT OF A DEPOSIT USING DEEP NEURAL


NETWORKS

Fırat ATALAY1, Güneş ERTUNÇ1

1
Hacettepe University, Beytepe Campus, Mining Engineering Department, Ankara, Turkey,
atalay@hacettepe.edu.tr, gertunc@hacettepe.edu.tr

ABSTRACT

Ore resource estimation is one of the primary step of pre-feasibility/feasibility studies of mine
projects. Feasibility of the projects directly related with the spatial distribution of the target content
which has economic interest. For this reason, spatial distribution of the ore content has to be
estimated. This paper aims to estimate the iron content of an iron deposit using deep neural networks.
42 drillholes are drilled at the deposit. In order to perform estimations solid model (wireframe model)
of the deposit is generated using sections method and block model of the deposit is also generated.
Using Matlab® a deep neural network is trained for drill hole data. This trained model is used to
estimate the iron content of the block model. Results of the estimates are compared with raw data and
trend analysis is performed to validate the estimations. Results shows that estimations are unbiased
and can be used for further for mine planning, pre-feasibility and feasibility studies.

Key words: Deep neural networks, Iron ore

1. INTRODUCTION

The estimation of the variability of the target variable of the ore resource plays vital role for
mine planning and other sequential steps at the life of a mine. At today’s technology and
economic conditions only limited number of data can be collected from a prospective ore
resource. Due to the limited availability of the data, variability of target variable where no
data is collected has to be estimated.
In ore resource estimation generally classical geostatistical methods are used such as kriging
methods. Unfortunately, in order to apply kriging methods, generally cumbersome steps have
to be followed such as trend removal, data transformation, variogram modelling and
generation of kriging plans. Also, stationarity is the main assumption of the kriging which is
hard to exists in real world. In order to by-pass these laborious steps machine learning
algorithms (ML) can be used.
ML algorithms are alternative geostatistical methods in ore resource estimation to traditional
methods. ML is sub-branch of artificial intelligence and can be defined as family of

195
7th BALKANMINE CONGRESS – Proceedings, Prijedor 2017

algorithms that estimates the outcomes of the specific events without being explicitly
programmed. Decision trees, Support vector machines, Ensemble methods, Neural networks
are just some examples of the ML [1-4]. Due to the simplicity of the algorithms they are used
in various branches of engineering such as electrical, automotive, computer and various other
fields [5-7]. Mining is not-exception to these areas of engineering. [8-10]. Specific to ore
resource estimation ML algorithms are used in various researches [11-14].
In this study, Deep Neural Networks (DNN)are used to estimate the iron grade distribution of
an iron ore deposit. For estimation purpose, firstly, raw data is composited in to equal lengths.
Ore body of the iron deposit is modelled using wire-frame modelling technique. Later block
model of the ore body is generated. In order to estimate the iron grade distribution on block
model, DNN model is trained and finally using this trained model iron grade distribution is
estimated.

2. RAW DATA AND COMPOSITING

47 drillholesare drilled at the deposit. All of the drillings are intersected with the iron ore. Due
to proprietary of data, location of the deposit is not exposed. Raw data is composited to equal
length samples and length weighted averages are calculated. Composites are shown in Fig. 1.

Figure 1. Oblique view of the composites

Summary statistics of the composites are given in Table 1.

Table 1. Summary statistics of the composites.


No of Minimum Maximum Average Variance Skewness
Data (%) (%) (%)
2701 0.01 56.28 17 200 1.16

196
7th BALKANMINE CONGRESS – Proceedings, Prijedor 2017

3. ORE BODY WIREFRAME AND BLOCK MODELLING

Ore body is modelled using wire-frame modelling technique. In order to create wire-frame model,
consecutive sections are used to create strings that represents the ore body. Later this strings are
triangulated to create volume of ore body. The created wireframe model is shown in Fig. 2.

Figure 2. Wireframe model of the deposit.

Total volume of the wireframe model is approximately 10 million m3.

Ore resource estimation cannot be performed directly on wireframes. Block model that honors
the wireframe model has to be created. For this reason, blocks having 20×20×10 m in X, Y
and Z directions respectively are used in block modelling. The block model is shown in Fig 3.

Figure 3. Block model of the iron deposit.

4. ESTIMATION USING DEEP NEURAL NETWORKS

DNNs are ML algorithms that learns the relationship between input and output variable(s).
Deep neural networks are specialized form of Artificial Neural Networks (ANN). Artificial
neural networks introduced in 1943which is inspired in animal brain. A node which stores the
information with 4 input is shown in Fig. 4

197
7th BALKANMINE CONGRESS – Proceedings, Prijedor 2017

Figure 4. An example node with weightings and bias parameter

The weighted signal collected at the node can be calculated as;

where b is bias term.


Later final value of the node is calculated using the activation function associated with the
neuron as;

Deep neural networks are ANN with more than two hidden layer and example of DNN is
shown in Fig. 5.

Figure 5.An example DNN.

Deep neural networks are approximately evolved after the ANN are first introduced. In 1986
back propagation algorithm is introduced which make deep neural networks calculation viable
and practical.

5. DEEP NEURAL NETWORKS FOR RESOURCE ESTIMATION

In resource estimation using DNN input and output variable(s) have to be determined. In
general, in resource estimation input variables are selected as X, Y and Z coordinates of the
composites and output is selected target variable measured at these locations.
Following the determination of the input(s) and output(s) the model have to be trained with
appropriate activation function with appropriate DNN architecture.
Finally, by training, DNN model becomes ready to make estimation at unsampled locations.
Simple feeding the coordinates of X, Y and Z as input where the grade of the target variable is
to be estimated. Trained model simply generates the output of grade by using weights
associated with all nodes of DNN.

198
7th BALKANMINE CONGRESS – Proceedings, Prijedor 2017

6. ESTIMATION OF THE IRON CONTENT USING DEEP NEURAL NETWORKS

6.1. Selection of the Activation Function

There is numerous activation function is available in use of DNN estimation. Linear,


Hyperbolic tangent sigmoid, log-sigmoid function are just examples s-among many. Ore
resources are defined on positive space which means that it is not possible to take negative
values. For this reason, every activation function cannot be used. The functions which
guarantiesthe positive outputs. Log-sigmoid function is one of appropriate candidate among
the alternative. Log-sigmoid function is special case of logistic function with formula;

where is defined in
Log-sigmoid function is defined in range which makes outputs of the DNN positive. For
this reason, log-sigmoid function is used in this study as the activation function.

6.2. Deep Neural Network Architecture Selection

There is no consensus over how to select the architecture of the DNN. For this reason, trial
and error methods seems to be best approaches. The [3 15 15 15 10 5 1] architecture is
selected in this study where 3 and 1 represents the input and output layer respectively.
Remaining integers shows the number of nodes at each hidden layer.

6.3. Training of Deep Neural Network Model

In training of the DNN sample set is subdivided in three categories which are training,
validation and testing purposes. Training data set is used to train the network and network is
adjusted respect to its error. Validation sub-set is used for halting of the calculations.
Calculations halts when generalization converges. Testing sub-set is used for network
performance monitoring. This data does not have direct effect on network training. 70 % ,15%
and 15% of the data is used for training, validation and testing respectively. In all calculations
MATLAB® is used which makes calculations and model tuning easy and error free.

7. RESULTS AND DISCUSSIONS

All blocks are estimated as described at previous section. Resulting block model is shown in
Fig. 6.

199
7th BALKANMINE CONGRESS – Proceedings, Prijedor 2017

Figure 6. Estimation results.

The summary statistics of the estimations are given in Table 2.


Table 2. Summary statistics of the estimates.
No of Minimum Maximum Average
Variance Skewness
Data (%) (%) (%)
2471 0.01 56.89 13.71 131 1.50

When Table 1 and 2 compared minimum and maximum of the composites and estimates are
nearly identical. While average and variance of the estimations are slightly lower than
composite values. Skewness of the estimates are higher than composites. Results shows that
estimation results are unbiased from the global statistical perspective. This means that
estimates can be used for further mine planning and ore reserve estimation purposes. This
study also shows that DNNs can be used to ore resource estimation for iron deposits.

REFERENCES

[1] Safavian, S. R., &Landgrebe, D. (1991). A survey of decision tree classifier methodology. IEEE
transactions on systems, man, and cybernetics, 21(3), 660-674.
[2] Hearst, M. A., Dumais, S. T., Osuna, E., Platt, J., &Scholkopf, B. (1998). Support vector machines. IEEE
Intelligent Systems and their applications, 13(4), 18-28.
[3] Dietterich, T. G. (2000). Ensemble methods in machine learning. Multiple classifier systems, 1857, 1-15.
[4] Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1988). Learning representations by back-propagating
errors. Cognitive modeling, 5(3), 1.
[5] Andrieu, C., De Freitas, N., Doucet, A., & Jordan, M. I. (2003). An introduction to MCMC for machine
learning. Machine learning, 50(1-2), 5-43.
[6] Guo, H., Crossman, J. A., Murphey, Y. L., & Coleman, M. (2000). Automotive signal diagnostics using
wavelets and machine learning. IEEE transactions on vehicular technology, 49(5), 1650-1662.
[7] Dash, M., & Liu, H. (2003). Consistency-based search in feature selection. Artificial intelligence, 151(1-2),
155-176.
[8] Porwal, A. K., &Kreuzer, O. P. (2010). Introduction to the special issue: mineral prospectivity analysis and
quantitative resource estimation.
[9] Aldrich, C., Moolman, D. W., Gouws, F. S., & Schmitz, G. P. J. (1997). Machine learning strategies for
control of flotation plants. Control Engineering Practice, 5(2), 263-269.
[10] Manoj, K., &Monjezi, M. (2013). Prediction of flyrock in open pit blasting operation using machine
learning method. International Journal of Mining Science and Technology, 23(3), 313-316.
[11] Zuo, R., & Carranza, E. J. M. (2011). Support vector machine: a tool for mapping mineral prospectivity.
Computers & Geosciences, 37(12), 1967-1975.
[12] Dutta, S., Bandopadhyay, S., Ganguli, R., &Misra, D. (2010). Machine learning algorithms and their
application to ore reserve estimation of sparse and imprecise data. Journal of Intelligent Learning Systems
and Applications, 2(02), 86.
[13] Porwal, A. K., &Kreuzer, O. P. (2010). Introduction to the special issue: mineral prospectivity analysis and
quantitative resource estimation.
[14] Chen, Y. (2015). Mineral potential mapping with a restricted Boltzmann machine. Ore Geology Reviews,
71, 749-760

200

Das könnte Ihnen auch gefallen