You are on page 1of 6

Application of Neural Network Model for Ore Boundary

Delineation Based on Geophysical Logging Data

Yi Huang & Stefan Wanstedt


Division of Mining Engineering
Lulel University of Technology
S-97187, Luleg, Sweden
e-mail, Yi@anl. luth. se, Stefan@anl. luth. se

Abstract

In a mining operation, knowledge regarding the ore boundary is extremely important. Mining cost and ore
quality largely depend on this information. The conventional technique to get this information is diamond
core drilling. The disadvantages of this technique are that it is very expensive and time consuming. In recent
years, geophysical logging has been introduced to the mining industry to get this ore boundary information.
However, effective interpretation to delineate the ore boundary from the geophysical logging data is still a
problem. In this paper, a back propagation neural network model is applied to delineation of the ore
boundary based on borehole 4 geophysical parameters logging data in a Swedish underground mine. Three
boreholes geophysical logging data was tested for ore boundary delineation purpose. The result from the
neural network model about the ore boundary delineation is encouraging and much better than the existing
geophysical logging data interpretation techniques.

1. Introduction of the problem

In mining industry, geological information regarding the ore boundary is very important, and mining cost
and ore quality largely depend on this information. The traditional technique to get this information is
diamond core drilling with which small pieces of rock is extracted from the rock mass. However, this
technique is very expensive and time consuming. As a result, the number of diamond core drilling holes is
limited and the distance between the diamond core drilling holes is usually 50 meters. In 50 meters, the ore
body boundary can vary significantly. For mining purposes, this ore boundary information is often not
enough. In a production mine, many boreholes are drilled for blasting purposes or for reinforcing the rock
mass. None or a little use is made of the opportunity to explore these holes. In recent years, geophysical
logging technique was introduced to mining industry (Aymler,etc., 1976, Elkington, 1983, and Wanstedt,
1991). This geophysical logging technique with the help of statistical analysis (Urbancic and Bailey, 1988,
Carlsten, etc. 1989) and correlation analysis ( Wanstedt, 1992) can provide the ore boundary information
roughly and partly solve the problem of precise ore boundary information shortage. The prediction accuracy
about ore boundary is, however, not good enough as ambiguity can occur when differentiating between ore
and waste sulphides. We think it can be improved. In this paper, the neural network technique has been used
to interpret the geophysical logging data for ore boundary delineation.

0-7803-3210-5/96 $4.0001996 IEEE 2148


2. Neural networks model

A neural network is a nonlinear dynamic system and has some important features such as self-leanling,
adaptive recognition, nonlinear dynamic processing and associative;memory etc. It has the ability to learn
knowledge from historical data and bring forth new knowledge and generalisation. Therefore, neural
networks have been widely used in industry. For examples, neural networks have been used to diagnose the
faults of chemical processes (Fan, Nikolaou and White, 1993), to identify abnormal event in nuclear power
plants (Ohga and Seki, 1992), to extract shoreline features (Ryan, Yuen and Hunt, 1991), to analyse
financial health of a company (Barker, 1990), to predict reliability (Karunanithi, Whitley and Mallaiya,
1992), etc. Neural network models are good at solving problems, in which the historical data is rich, inany
parameters influence the process and results, and one doesn't have fuill understanding of this process.

Considering OUT problem's characteristics, a back-propagation neural network model (see Schalkoff, 1992
and Patrick 1990) was used. After several tries, the back propagation neural network model shown in figure
1 was chosen for the ore boundary delineation problem:

Rook type

a nodler

6 node-

,- nodes

4 nodes

Qeophysicrl Logging data

Figure 1. The structure of the neural network model

It has 4 nodes the in input layer, 115 nodes in the first hidden layer, 5 nodes in the second hidden layer and 3
nodes in the output layer. The transfer function between the input layer and the first hidden layer is "tansig",
between the first hidden layer and second hidden layer is "tansigl' and the second hidden layer and the output
layer is "logsig" , The training data is shown in table 1. After training, the neural network model reaches;to a
very good convergence.

3. Data handling

The geophysical logging data includes gamma-ray, density, neutron and resistivity. A technical description
and measurement procedure of the logs can be seen in Whstedt (1992).
We have used data from 3 boreholes (B33, B34, B36) in a Swedish mine. The data in each hole comprises 4
geophysical logs (1) gamma-ray, (2) density, (3) Neutron and (4)resistivity, with corresponding core logs
and assays of miineralized zones of the core. Every 10 cm, a geophysical measurement was recorded by each
probe. The data from borehole B33 was taken as training data for the neural network model. Rock types
according to the core log are reduced into 3 groups, where group I 11swaste rock, group 2 is semi-ore, and
group 3 is ore. In order to overcome problems caused by possible operating error, it is necessary to have
enough training samples for the neural network model. This also improves possibility for convergence and

2149
successful application on field data. For this purpose, borehole B33 is divided into 20 sections and for each
section, the 4 geophysical logging data are averaged.

In neural computation, a non-linear function, called sigmoid function is used to constrain the output value
from a node to a range of 0 and 1. Therefore, it is essential to normalise the target signals in the same range
before training. In practice, the input signals are also required to be normalised in the same range. After
several tries, finally, these 4 geophysical logging data are normalized with the following formulas.

For density, the normalized data di is calculated by the following formula

d, =(D.
- Dh)/(Dm
- D-)
where Di is the measured density parameter, D m b is minimum value of density and Dmax is maximum
value of density.

For resistivity, the normalized data pi is calculated by the following formula

where Pi is the measured resistivity, P,in is minimum value of point resistivity and Pmax is maximum
value of point resistivity

For neutron, the normalized data ni is calculated by the following formula

ni = ( N ,- N m, ) I ( " - N m. )
where Ni is the measured neutron parameter, "in is minimum value of neutron and "ax is maximum
value of neutron

For gamma-ray, the normalized data gi is calculated by the following formula

gi = lOg(Gi)
where Gi is the measured gamma-ray parameter

After above procedure, the training data from borehole b33 is formed and shown in table 1.

Table 1. Training Data

input Data Output Data


Depth (m) lensity Resistivity Neutron Gamma-ray Rock class
1.8--3.3 0.0298 0.952 0.5382 0.6317 1 0 0 1
3 -4-9.6 0.0217 0.8554 0.6784 1.0834 1 0 0 1
9.8--10.0 0.6071 0.2817 0.233 0.6199 0 0 1 3
15.5--15.7 0.5179 0.165 0.2604 0.2315 0 1 0 2
15.8--16.4 0.6973 0.2142 0.2299 0.2469 0 0 1 3
16.5--17.0 0.4355 0.2242 0.2921 0.5504 0 1 0 2
17.5--17.9 0.431 0.1497 0.2556 0.6836 0 1 0 2
18.0--18.8 0.6144 0.1381 0.193 0.86 0 0 1 3
19.0--20.1 0.6954 0.1064 0.2037 0.4484 0 0 1 3
20.3--20.5 0.3849 0.2778 0.2162 0.4192 1 0 0 1

2150
4
0.6--20.8 0.5224 0.2017 0.444 0.717 0 I 0 2
0.9--21.3 0.1464 0.2612 0.2784 0.9451 0 1 0
1.4--22.1 0.2679 0.2959 0.296 1.022 0 1 0
3.0--23.8 0.3743 0.3404 0.2492 0.773 0 1 0
3.9--24.1 0.5357 0.2331 0.1623 0.282 0 0 1
4.5--26.8 0.033 0.8553 0.6614 0.988 1 0 0
7.5--29.2 0.0985 0.6832 0.6238 0.720 1 0 0
9.8--33.1 0.0793 0.819 0.6524 0.773 I 0 0
3.3--33.7 0.8952 0.3138 0.1287 0.224 0 0 1
0.0307 -0.9397 0.7809 1.155 -1 0 0 -.

4. Results analysis

Based on this trained neural network model, boreholes B33, B34, El36 were tested. The testing results are
shown in figures 2-4. The thick line is the results of neural network model. The thin line represents the core
log.
4.00 -
1
I
YB
*-0° 4
..
; 0.00 -
0.00 10.00 20.00 30.00 40.00
Depth (m)

Figure 2. The prediction result vs. core log in borehole B33

From figure 2, for borehole B33, it can be seen that the difference 'between neural network model and the
core log are negligible. The rock classes ore and semi-ore are selected based on subjective decisions b!y the
geologist. The 2 spikes at 27m are probably misclassified due to flaws in the measurement technique.

4.00 -
1

,-IIvI-
0.00

0.00 10.00 20.00 30.00 4Cl.00


Depth (m)

Figure 3. The Prediction result vs. core log in borehole B34

2151
From figure 3, for borehole B34, it can be seen that except depth 20m-25m7 along the whole borehole
length, the difference between neural network model and the core log is negligible. In depth 20m-25m, even
though one could assume that rock is homogeneous from looking at the core log, the gamma log shows that
it isn't. There are distinct zones of ore within the semi-ore between 20m to 25m.

1
73

B
2.00

1
Figure 4. The prediction result vs. core log in borehole B36

From figure 4, for borehole B36, the results from the neural network model appear to be least successhl.
The model disagrees with core between 20m to 26m. In this area, there are parts with disseminated metals
on the core log. Furthermore, the geophysical logs suggest metals as well. Hence, some of the rock
classified as waste could just as well be semi-ore. It is possible that the metals within the rock in this hole
are of no economic value, hence classified as waste. Nevertheless, they affect the geophysics and the neural
network model.

Based on this analysis and re-examining results, the real misclassification area is very small. The results for
ore boundary delineation are satisfactory.

5. Conclusions and discussion

Result presented here are, if not great, will be within whst's acceptable for the mine.
(1) There is the relation between these 4 geophysical logging parameters and the rock type. Even though thls
relation is very complicated, it can be learned by neural network models.
(2) The prediction accuracy of rock type by neural network model based on these 4 geophysical logging
parameters is much higher than statistical analysis. This is very good result for mining engineering.
(3) Training data should be chosen very carefully. It should be very representative and be handled in proper
way to minimize the errors caused by the logging operation.

Acknowledgement

Authors are very thankful for the Swedish Board for Technical Development (NUTEK) for sponsoring and
financing this project and Dr. Uday Kumar for his comments and manuscript improvements.

2152
References

[l] J. Aymler, P. Eisler, P. Methew, and A. Wylie, T h e Use of Natural Gamma Radiation for Estimating
the Iron Content of Sedimentary Iron Formations Containing Shale Bands", Nuclear Techniques in
Geochemistry and Geophysics. Proceedings of a panel, Vienna, 1976.
[2] D. Barker.., "Analysing Financial Health Integrating Neural Networks and Expert System", Int. J. PC
Artificial Intelligent, May, 1990.
[3] S. Carlsten, L. Lindqvist and 0. Olsson, Comparison Between lbdar Data and Geophysical, Geoltogical
and Hydrological Borehole Parameters by Multivariate Analysis of Data, SKB Technical Report 89-15,
Swedish Geological Company, Uppsala.
[4] P. Elkingtoin, "The Application of Wireline Logging Methods to Surface Mining", In Proceedings of 2nd
Surface Mining and Quarrying Symposium of IMM, Bristol, England.
[SI J. Fan, M. Nikolaou and R. White, "An Approach to Fault Diagnosis of Chemical Processes via Neural
Networks", Int J. AIChE, Vo1.39, No. 1, 1993.
[6] N. Karunanithi., D. Whitey and Y. Malaiya, "Using Neural Networks in Reliability Prediction", Int. J.
IEEE Software, July, 1992.
[7] C. Lee and R. Steming., "Identify Probable Failure Modes for Underground Openings Using a Neural
Network", Int. J. Rock Mech. Min. Sci & Geomech. Abstr. Vol. 29., No. 1, 1992.
[8] Y. Ohga and H. Seki, "Abnormal Event Identification in Nuclear Power Plants Using a Neural Network
and Knowledge Processing", Int. J. Nuclear Technology, Vol. 101, 1993.
[9] P. Simpson, Artificial Neural Systems - Foundations, Paradigms, Applications and Implementations,
Pergamon Press, New York, 1990.
[lo] T. Ryan, P. Sementilli, and B. Hunt, "Extraction of Shorelilne Features by Neural Nets and Image
Processing", Int. J. Photogrammetric Engineering & Remote Sensing, July, 1991.
[ 111. R. Schalkoff, Pattem Recognition - Statistical, Structural and Neural Approaches, John Willey &
Sons, Inc., Singapore, 1992.
[ 121 T. Urbancic and R. Bailey, "Statistical Techniques Applied to Borehole Geophysical Data in Gold
Exploration", Geophysical Prospecting 36, 1988.
[ 131 S. Wbst edt, Geophysical Borehole Logging in Malmberget, Technical Report 1991:14T, LuleH
University of Technology, 1991.
[14] S. Wbstedt, Geophysical Borehole Logging Applied to the Location and Characterisation of Base
Metal Ores, Licentiate Thesis, Lulei University of Technology, Sweden, 1992.

2153

You might also like