Drill Process Monitoring in Percussive Drilling (PDFDrive)
Drill Process Monitoring in Percussive Drilling (PDFDrive)
Drill Process Monitoring in Percussive Drilling (PDFDrive)
by
HÅKAN SCHUNNESSON
Division of Mining Equipment Engineering
TEKNISKA
HÖGSKOLAN I LULEÅ
LULEÅ UNIVERSITY OF TECHNOLOGY
·,:
DRILL PROCESS MONITORING IN PERCUSSIVE DRILLING - A
MULTIVARIATE APPROACH TO DATA ANALYSIS.
By
Hakan Schun12esson
Division of Mining Equip77Zent Engi,neering
Lule&. University of TechnoZogy, Lulea, Sweden
i
HÖGSKOLAN 1 LULEA
1990 -03- 2 4
PREFACE BIBAJOTEKET
Results presented in this thesis are all based on drilling data obtained from
LKAB Kiirunavaara mine, Kiruna, Sweden. Contribution by drilling operators
are therefore greatly acknowledged. I wish to express my thanks to Lars
Hermansson and Ulf Enback, LKAB, for moral support and for providing all
necessary help during the project completion, and also to LKAB mining
company and all other personnel at LKAB/Kiruna that have contributed to this
work
I sincerely thank Dr. Thomson Sinkala (LuTH) for a detailed scrutiny of the
early manuscript of this thesis, especially concerning the drilling aspects and
revision of the English.
I wish to express my thanks to Prof Svante Wold of Umeå University for valuable
comments on the manuscript, especially concerning the multivariate theory and
the treatment of data.
Thanks are also due to Torgny Almgren, Hans Wirstam and Stefan Wänstedt, all
of LuTH, for comments on the final version of the manuscript.
I thank Dr. Jonathan Peck, CCARM, Canada, for helpful discussions and
comments on chapter 3.1.3.
Håkan Schunnesson
11
ABSTRACT
In a field test i the OSCAR area in LKAB's Kiirunavaara mine, penetration rate,
thrust, torque, rotation speed and air pressure were recorded on a production
drill rig for every 10 mm of hole length.
In the treatment of the drilling data Principal Component Analysis (PCA) is used
to locate outliers in the data, to study drilling in rock with different ore grades
and drilling in rock with different mechanical properties. Partial Least Square
(PLS) regression is used to correlate the drilling response with the content of
iron in the rock. To perform a prediction of iron ore grade over a larger area
several different models are used to achieve a reasonable prediction. Finally,
PLS is used to locate the ore boundary from the recorded parameters.
The result shows that for drill monitoring of percussive drilling where all
parameters carry information about the rock and at the same time are inter-
related, PCA provides a substantial simplification of the data analysis.
Using PLS, prediction of rock properties can be done reasonably well even for
small changes in rock response to drilling.
TABLE OF CONTENT
PAGE
PREFACE ii
1. INTRODUCTION 1
3. DATA ANALYSIS 7
3.1 Past attempts to analysis of drill monitoring data 8
3.1.1 Single parameter response 8
3.1.2 Variable interaction between drilling parameters 8
3.1.3 Pattern recognition 9
3.2 Multivariate analysis of data 10
3.2.1 Principal Component Analysis (PCA) 12
3.2.2 Partial Least Square (PIS) regression 14
iv
7. DISCUSSION AND CONCLUSION 45
7.1 PCA applications 45
7.2 PIS applications 46
8. RECOMMENDATIONS FOR FURTHER STUDIES 48
REtiERENCES
1. INTRODUCTION
Due to the general interest in large scale mining methods the mining industry
today requires more detailed information about the ore and the surrounding
rock mass.
Today the information about the ore is often based on widely spaced core
drilling. In stoping methods the estimation of rock conditions for entire stopes
can therefore be based on few core holes, resulting in high uncertainties for
uncored areas.
Drilling data are often very complex. Several different features may interact, and
random components and measurement errors of different magnitudes may also
be associated with variables in the measured data.
Monitoring of drill rigs has to date mostly been applied to rotary drilling. In
Sweden where percussive drilling is the commonly used technique, drill
monitoring has hardly been used. The reason for this is partly the increased
complexity of monitoring data in percussive drilling compared to other drilling
methods. As an example, the useful energy that reaches the bit/rock contact
depends on factors such as the impact energy, the transmission efficiency of the
wave through the drill string, the size and shape of the bit, the contact conditions
between the bit and the rock. It is therefore difficult to accurately determine the
amount of energy used for rock fragmentation in a drilling situation where all
these conditions vary. Furthermore, the rock is an extremely inhomogeneous
material. Properties such as rock strength and structure will often vary, thereby
introducing variations in the drilling process.
This thesis attempts to show how analysis of drill monitoring data from
percussive drilling can be improved by the use of multivariate techniques. These
techniques are used to evaluate drilling data obtained from field drilling tests in
Kiirunavaara, Sweden.
2
2. DRILLING TEST AND PARAMETERS
OBTAINED
2.1 IN'1RODUCTION
The drill monitoring test was performed in the OSCAR area in the central part
of LKAB Kiirunavaara mine in Kiruna.
The Kiirunavaara ore mainly consists of magnetite, with apatite as the most
significant polluting mineral. The ore body is approximately 4 km long, has an
average width of 80 m and dips about 60 degrees eastward.
In the test area, traditional sub-level caving has been replaced by sub-level
stoping in order to reduce development work. The area is divided into 4 stopes
by intermediate pillars (Figure 1). Each stope is approximately 30 by 30 m in
width and between 60 to 80 m in height.
Sublevel stoping
Heigt: 60-80 m
Raise boring diam: 2 400 mm
Hole length: 45-60 m
Hole diam: 115-165 mm
Max deviation: 1%
Tons of ore/drilled meter: 45-60
Tons/blast: 200 000 (max)
Explosives: Anfo, emulsion
3
All production holes (Figure 2) are drilled downwards from 596.5 m level, using
an Atlas Copco SIMBA 269, a highly automated production drill rig. By the aid
of a TV camera, the rig is monitored and controlled remotely from a cabin
usually located in a separate drift.
The rig was equipped with a 4.5" ITH hammer (COP 42), using 4.5" button bits.
The drill tubes used had an outer diameter of 76 mm and were 1875 inm long.
During the test period, drilling parameters were recorded for every 10 mm of
hole length for 156 holes totalling 5118 m. The recorded parameters are:
4
ORT 298
ORT 297
..._,
e•
.ee
'1' 1
ORT 295 1.
•
ORT 296..• i
e• •
/ • • I •
e •
ORT 290 ORT 294 / •
• I
• • • • •
i •
29024200 •
• •
ORT 289
290211200 • • ••• • • • • • • • • •
• • • •
• • ••• • • • • • • • • •
• • bib. • • • S • • • • • • •
• • • • •
81111 i Il i I I III I 1 I *
• •
• 29308200
•
29805201
• • • • X1
• • d
•
2900-3200
I. • •
29000200
.I .
I
•
•
• •
• •
v rcorehole
8,-129800200
• • x
• • „. 29896200
29016100 (29696200
• . • • XX. x K
• f 29093200'291932(4
• rif L. ix
x xir x • X X X •NIC AI • • • • • • • • ••• • • 5 • X X X • • •
•
• • • 4.4 • • • • X X X • • •
x • x K .4( x. • • II • •
X X X • X X *IN 4. • • • • • • • • ea* • • • • X • X • • •
6
3. DATA ANALYSIS
100.0 -
80.0 -
It
60.0 A I A 11
- il Il 1I \ 11 ;1 Rotation peed ' Il
it I It
J \ t1 11 /1‘ /1
11 41 z
I `, 1, it ,
-• i J
J t1s 11: r
40.0- — ti ti I 1 /4 I 1
•
20.0 -
Torque.
,
Thrust
0.0 1111111111(1 1 I
6.7 6.8 6.9 7.0 7.1 7.2 7.3 7.4 7.5 7.6 7.7 7.8 7.9 8.0
Hole Depth (111)
The air to a drill rig is normally supplied from the central air system in the mine.
The air pressure to the rig is therefore dependent on the compressed air
distribution system and air consumption elsewhere in the mine.
7
During the drilling process the air pressure also varies depending on the flow
resistance in the drill system. The higher the resistance the smaller the flow and
the higher the input air pressure to the system.
Hence, all drilling parameters vary more or less with rock properties while they
also all interact during the drilling process. It is therefore imperative that all the
parameters be considered in the model for predicting rock properties.
The earliest approach to analysis of drill monitoring data was based on single
parameter response as an indication of rock condition. Normally penetration
rate was used to indicate the quality of the penetrated material.
Morfeldt et al (1973) report early use of this technique for the foundation of high
rise buildings in Stockholm, Sweden. The technique was here used with success
to locate the solid rock under the overburden.
Similar application has been reported by Homer and Sherrell (1977). Here one
also aimed at discovering cavities and othe—r discoiifiriaties in the rock mass. In
order to correlate penetration rate with the rock properties, all other drilling
parameters were as much as possible maintained constant.
This type of data analysis has been applied with success when there is a
substantial difference in rock properties between different geological domains
(for example between solid rock and overburden in surface drilling). However,
when the difference in rock properties between different rock types is small, and
the contribution from the drill parameter interaction become more dominating
this type of analysis is not effective.
8
Pfister (1985) reports successful practical use of several relations in rotary
drilling based on the required drilling energy.
Brown and Phillips (1977) have extensively reviewed literature on relations for
rotary drilling, and Barr (1984) has tested a few of them in his experiments.
Results obtained have varied from one test to another.
Let us assume that for a certain rock there is a mismatch between, for example,
the true compressive strength and the predicted compressive strength based on
drilling parameters. This means that the relation used is not very well suited for
the specific type of rock, and has therefore to be re-calibrated. This calibration is
however not always a straight forward matter since not only the contribution
from each parameter involved has to be considered, but normally also the inter-
relation between the parameters. Even if the true properties of the rock are
available for calibration, it is rare to obtain an optimal solution by manual
calibration.
9
Griffiths (1983) used the technique to classify different rock types during oil well
drilling. Due to the interaction between drilling parameters he treated the
parameters as fuzzy variables. In the best cases he succeeded to correctly classify
70 % of the rock mass.
Peck and Pollitt (1990) used pattern recognition to classify rock as coal,
mudstone or siltstone based on drill parameter response, in an open pit coal
mine in Canada. In order to separate the different rock types single parameter,
cross- and 3-dimensional plotting was used. In general, the 3-dimensional plot
gave better separation between the three rock types than the other plots. The
technique was very successful in classifying coal and siltstone, but not mudstone
that in terms of rock properties lies between coal and siltstone.
A general problem when trying to separate data groups in a space with more
than two dimensions, is to decide from where to view the data. Viewed from one
direction the different data groups may appear to belong to the same population
while viewed from another direction the differences may clearly be seen. See, for
example, Figure 4.
10
• 0 • • • CD0 0 0 X
1
• 0 0 0 • • 0 COO 0 X 2 (a)
• 0 C900 00000 x
3
• o
xi
(b) (c)
In order to use all three variables in the analysis, a three dimensional coordinate
system can be used with x, y, z axes corresponding to penetration rate, trust and
air pressure respectively. Each measurement, involving the three variables, is
marked as a point in the coordinate system (Figure 4b),//
Looking at the points through a window defined by the axes x1 and x2, the
difference between the two groups, ore and side rock, can still not be
distinguished. However, viewing this data from some suitable angle(s) a clear
difference between the groups is observed, as shown in Figure 4c.
The difference between the two groups in figure 4 is distinguished through the
difference in variable interaction, and not through magnitudes of the drilling
variables.
11
63(
If in Figure 4c the coordinate system is rotated so that the first axis coincides t,
with the direction of the largest data variation (called the first principal
component), the second axis coincides with the direction of the second largest
data variation (called the second principal component), and so on, we will have
transferred the three dependent drilling parameters into a few independent
Principal Components, thus capturing the dominating data structure in the
original data set, Furthermore the difference between ore and side rock (in
Figure 4) can now be distinguished through the value of the second principal
component.
This type of analysis is not limited only to 3 dimensions, but can be performed
for any number of dimensions (variables), even though a space with more than
three dimensions may not be as easy to comprehend.
The description above is the ground for Principal Component Analysis (PCA).
PCA normally forms the basis for multivariate data analysis, and is normally
used for simplification of a data table, creating models, noise reduction, outlier
detection, variable and object selection, correlation evaluation, classification,
and prediction of different features (Wold et al, 1987).
The idea of PCA is to find directions in the data space that will indicate typical
features. Unless the swarm of data points is spherical (for a 3 dimensional space)
it is usually possible to identify a dominant direction of the data and with
regression fit a line to the points. In most cases this direction. will not coincide
with any single variable.
The drilling data analysed in this thesis is stored in a matrix X with p columns
relating to the recorded parameters, and n rows relating to different data
sampling locations along respective holes.
12
0
•••••••
e
Average point
VARIABLE 1
13
The elements in X are modelled by;
where the matrix ej contains the residuals not explained by the Principle
Component model.
The number of significant dimensions (a), with one t and one c vector for each
dimension, has to be determined. In this thesis cross-validation (Jolliffe, 1986)
has been used.
a
= E tik cki + eii (2)
k. 1
By plotting the score (t) vectors against each other, one obtains informative 2-
dimensional pictures of the objects configuration in the multi-dimensional data
space. The effect of each originally measured variable on each score point can be
identified by studying the loading (c) vectors together with the score vectors.
The technique used here, PLS, is an extension of PCA to include the dependency
structure between sets of variables. The goal is to find a relation between X and
Y such that Y is modelled by X (Ståhle and Wold, 1986).
14
Like in PCA, the elements in the X-block can be modelled by
The inner relation can then be found by looking at a graph of the t-vector against
the u-vector. The simplest model for this relation is a linear one. Thus
U = dt (5)
This model, however, is not the best possible since the principal components are
separately calculated for both blocks and have therefore a weak relation between
them. In this model the best possible correlation is obtained by a simultaneous
calculation oft and u in the PIS algorithm, so that the components become
slightly rotated and lie closer to the regression line.
z
• 1, egg 4(
• z
Z
em
•r I
,
15
When model (3), (4) and (5) is not sufficient to describe the systematical
structure of X and Y, it may be possible to extract new components from the
residual matrix.
The number of significant dimensions (a) has also here been determined with
cross-validation.
Thus, the general properties in PCA and PLS are the same. The main difference
is that in PIS the knowledge about the design of an experiment is allowed to
affect the direction of the components. Differences between a control group and
a treatment group that are of a subtle nature may therefore be detected by PLS
but not by PCA (Stähle, 1987).
When a model has been created for the relation between the X-block and Y-
block, new objects defined by the variables in the X-block can be used to predict
the values in Y. Since both the model and the new object is well defined in the
multivariate space, the distance between them can be calculated. This distance is
called "Distance to Class" (DCL), and indicates how well the data point fits the
model. A small DCL value indicates a high similarity between the point and the
data set used to create the model, and will therefore also indicate a more
reliable prediction. The converse is also true.
The calculations done in this thesis are all performed with a commercial
software package called SIMCA. SIMCA uses the NIPALS method (V/old, 1966)
to calculate components. Different other methods has been studied by Ståhle
(1987) concluding that the NIPALS method is the most convenient, since it
converges rapidly, allows incomplete data and is particular well suited for cross-
validation.
16
4. APPLICATION OF PRINCIPAL COMPONENT
ANALYSIS TO DRILL MONITORING DATA.
In Figures 8a and 8b the first two Principle Component score vectors and the
corresponding two loading vectors are plotted, respectively, for a section of hole
29086100. It is observed that the majority of the data points belong to a well
defined group in the lower part of Figure 8a. However, some outliers can
immediately be identified as not belonging to the main part of the data
population in the figure.
PC2 8
A
[o]
B
6- fl
4-
_ C_
033
2- DO
• 0 ci
a i U.
-7.
0
—2-
—4 1 I
—5 —3 —1 3
PC1
Figure 8a. Score plot of the first and second principal component for
hole 29086100.
17
PC2 1.0
0 Penetration rate
a Torque
0.5
0.0
Thrust o 0 Rotation peed
o Air pressure
-0.5
-1.0
-1.0 -0.5 0.0 0.5 1.0
Figure 8b. Loading plot of the first and second principal component for
hole 29086100.
By a first analysis of the outliers data can be arranged into three different groups
marked (A), (B) and (C). From Figures 8a and 8b the following can be
concluded:
Group (A) consists of only one data point and is characterised by high torque
pressure and penetration rate. By examining each and every measured point in
the massive raw data for every drilling parameter, it is observed that torque
pressure suddenly increases above the pre-set value in the anti-jamming system,
after which the pressure attains its normal operating level. Here only penetration
rate is observed to be slightly higher than normal.
The data points in group (13) are observed to behave similar to group (A), but
with a tendency towards higher penetration rate rather than tor_que_pressure. By
examining each and every measured point in the massive raw data for every
18
drilling parameter, it can be shown that the data points in this group are located
on either side of the point at which a new drill tube is added. This behaviour of
the drilling parameters identified just before or drier a drilling tube is added may
be attributed to the function of the drill rig rather than rock, and will therefore
also be excluded before further analysis.
19
The most significant polluting mineral in the Kiirunavaara ore body is apatite, in
which both phosphorus and calcium oxide are present. Figure 9, based on 416
independent cutting analyses in the high phosphorus area, shows that in this area
there is a high correlation between the content of phosphorus and calcium oxide
in the ore. This indicates that the dominating parts of these substances are
bounded in the apatite.
Calcium Oxide Content (%)
25
232
202
18
15 2
8-
52
o
2 4 6 8
Phosphorus Content (%)
Figure 9. The relation between phosphorus and calcium oxide in the
high phosphorus part of the ore.
The drilling parameters obtained from each of the three holes are averaged for
each rod. Drilling with the first three rods from hole collaring is done manually
at rather low thrust. These data will not be discussed here. Of interest is only the
drilling events after the collaring phase. In Figures 10a and 10b the first two
Principle Component score vectors and the corresponding two loading vectors
are plotted, respectively.
20
PCa 4
1
. A
2 A e
eA
A di IttAziA
AA-4,
A
a
mu al
0
so la
al II
0
.-or
8° ° o
0 9
Q)
—2
o Hole 1
la Hole 2
4 Hole 3
4
—4 —2 0 2 4.
P
Figure 10a. Score plot of the first and second principal components for
holes 1, 2 and 3.
F C1 1.0
o Torque
0.5
o Thrust
0.0
C] Rotation spe d
° Penetration r te
—0.5 -
o Air pressure
-1.0
—1.0 —0.5 0.0 0.5 1.0
PC1
Figure 10b. Loading plot of the first and second principal components
for holes 1, 2 and 3.
21
In Figure 10a, the data comprising only 5 input variables from the three different
rock types form three distinct classes. The response of data from hole 3 with high
content of quartz and apatite is clearly different from holes 1 and 2. This
difference is indicated primarily by high torque and thrust and a low air pressure
(Figure 10b). The boreholes 1 and 2 that geologically differ mainly in apatite
content appear to differ as regards the response of penetration rate and rotation
speed (Figure 10b).
A comparison between all the holes in the area shows that the average
penetration rate in the high phosphorus part of the ore is on average 15 % lower
than in the low phosphorus part (Figure 11), indicating a possible difference in
rock response.
20
Phosphorus ore
i F2 P
20 40 60 80 100
PENETRATION RATE (cm/mm)
Figure 11. Distribution of the average penetration rates in the high and
low phosphorus ore regions, respectively.
22
PC2 1.0
cAir pressure
0.5
aTorque to Quartz
D Thrust
ca Iron
-1.0
-1.0 -0.5 0.0 0.5 1.0
PC'
Figure 12a. Loading plot of the first and second principal components
for hole 1.
PC2 1.0
e_
0.5
a Air pressure
a Quartz
0 Iron
0.0 _
a Calcium o2d.de
Phosphorus
a Thrust
a Torque
-0.5 a Rotation speed
a Penetration rate
-1.0 i ,"
-1.0 -0.5 0.0 0.5 1.0
PC'
Figure 12b. Loading plot of the first and second principal components
for hole 2.
23
PC8 1.0
Cl Air presssure
0.5
a Torque
a Quartz
Cl Iron
0.0 _
Rotation speed
a a Penetration rate
a
Thrust
rt Calcium oxide
C Phosphorus
-0.5
-1.0
-1.0 -0.5 0.0 .5 1.0
PCI
Figure 12c. Loading plot of the first and second principal components
for hole 3.
For hole 1 in the low phosphorus part of the ore, it seems as if the decreasing
iron content is mostly replaced by quartz, and not apatite (Figure 12a). The
geological effect on penetration rate is not very clear in the figure. However,
since the rock is geologically rather homogeneous dominated by a high content
of magnetite and only a small amount of other minerals, the variation in
mechanical properties of the rock is likely to be rather small. Therefore no
strong conclusions can be made from the figure alone.
For holes 2 and 3 in the high phosphorus part of the ore, the dominating
geological variation is the iron/apatite interaction (Figures 12b and 12c). It can
also be noted that the quartz content appears to be un-correlated to this
variation.
Although the correlation between geology and penetration rate is not obvious in
the plots, it seems, however, that the penetration rate is more correlated to
quartz than to the iron and apatite occurring in common.
It has been shown above (Figure 11) that penetration rate in low phosphorus ore
is higher than in the high phosphorus ore. In the high phosphorus ore both the
content of apatite and quartz are high. But since quartz has a greater hardness (7
on Mohs scale) while apatite has a lower hardness (5 on Mohs scale) than
magnetite (5.5 on Mohs scale), the hardness of the two minerals occurring
together is somewhere in between, and the corresponding penetration rate will
24
probably respond in a similar manner. That is, while penetration rates will tend
to be low in quartz, the rates will tend to be high in apatite. From Figures 12 a-c
it appears generally that quartz has a greater effect on penetration rate than
apatite.
A sudden drastic change in rock strength changes the drilling conditions. In order
to study this phenomena a hole was core drilled between two production holes in
the OSCAR area (Figure 2). The distance from the cored hole to each of the
production hole was 1.75 m. At the bottom of the core drilled hole, a zone of
breccia with different properties was traversed. Some properties of the ore and
the breccia are given in Table 4.
The low density of the breccia is an indication that the presence of magnetite is
low.
In Figures 13a and 13b, the first two Principle Component score vectors and the
corresponding loading vectors are plotted for data from one of the production
holes close to the cored hole. This group of data is obtained from the lower half
of the bore hole.
The drilling conditions in the breccia are clearly different from the rest of the
hole. This difference is indicated by high torque pressure and rotation speed
(Figures 13a and 13b).
It is likely that the high torque pressure is due to high penetration rates in the
breccia with low compressive strength, resulting in a larger resistance to bit
rotation.
25
PC2 4
2-,
2-
4 Ii fill i IIII II [
i II I I
—4 —2 0 2 4 6
PC1
Figure 13a. Score plot of the first and second principal component for
hole 29800200.
PC' 1.0
El Air pressure
0.5
0 Torque
0.0 ci Rotation spe ed
Ei Thrust
0 Penetration rate
—0.5
—1.0 if
Figure 13b. Loading plot of the first and second principal component for
hole 29800200.
26
In Figure 14 the response of drilling parameters is shown for drilling through a
fracture zone. In a fracture zone the bit has a tendency to get jammed, resulting
in an increased torque. The torque increases either until the resistance is
overcome, often indicated by a peak in rotation speed (Figure 14), or until the
anti jamming function is activated to retrieve the bit.
200.0
175.0 -
Penetration Rate
150.0 -
Rotation
125.0 -
100.0
75.0 -
iI I
50.0 _I L I;
\t \-ovt , «
25.0 - :, içp
, V v ,,,
,„ reasure ,.... , , s„, v v ,:.%)
4VX.:vv,
Thrust .,- --
4i
ss s, _____„, ....s, .3.,
9.5 9.7 9.9 10.1 10.3 10.5 10.7 10.9 11.1 11.3 11.5
Hole Depth (m)
27
5. DETERMINING ORE GRADES FROM DRILL
MONITORING DATA.
5.1 INTRODUCTION
Varying ore grades normally affect the mechanical properties of the rock, and
will therefore also have an effect on the behaviour of the drilling process.
To make predictions of ore grades and other geological properties of the rock,
the drilling response has to be carefully calibrated against the geology in a
number of test holes. In order to do this detailed geological knowledge of the ore
and side rock is necessary.
The ore mineral in the Kiirunavaara ore body is magnetite of varying phosphorus
content. For a more detailed geological description of the Kiirunavaara ore
body, see Frietsch (1978) and Parak (1975).
The OSCAR area is located at the hanging wall side in the central part of the ore
body. The high/low phosphorus contact crosses the area so that approximately
half of the ore is high phosphorus and the other half is low phosphorus.
To get a better and more detailed knowledge of the ore grades in the area,
cuttings were sampled during drilling. The samples were taken for each rod
(1.875 m) from every hole in the middle row of every drift (Figure 2). The
samples were analysed for 13 different substances of relevance to the mining
company. In Table 5 the most important mineral substances are presented for
the high and low phosphorous ores in the test area. The table is based on 416
samples from the high phosphorus area and 495 samples from the low
phosphorus area.
Looking at the average values, the major differences between high and low
phosphorus ores are in the contents of iron, phosphorus and calcium oxide. The
phosphorus and calcium oxide are both present probably as apatite, which is
reported to be one of the dominating polluting minerals in the Kiirunavaara ore
body. The content of iron in the high phosphorus ore is approximately 10 %
lower than in the low phosphorus ore, while the apatite is correspondingly higher
in the high phosphorus part.
28
Table 5.
High phosphorus ore Low phosphorus ore
°lc) Standard deviation 0/0 Standard deviation
Fe 57.63 (9.87) 68.26 (3.24)
P 2.23 (1.48) 0.03 (0.24)
SiO2 2.18 (8.10) 2.22 (2.24)
CaO 7.48 (4.63) 0.37 (1.02)
MgO 1.07 (0.92) 1.08 (0.60)
A1203 0.18 (1.28) 0.24 (0.23)
The standard deviations indicate that the low phosphorus ore tends to be rather
homogeneous while the compositions of minerals in the high phosphorus part of
the ore have large variations. See, for example, the standard deviation in quartz
content in the high phosphorus part of the ore.
2.0
0.0 -I
—2.0 1-
_
—4.0
—4.0 —2.0 2.0 4.0 6.0
PC,
Figure 15a. Score plot of the first and second principal component for
high phosphorus ore.
29
PC2 1.0
0.5 "Ai2 03
Si0
ID2
cl Mg0
• o Fe
0.0
0 Ca0
-0.5 op
PC,
4.0
a
a a o 'a a
a
3.0 - a
a h m°
o 9:1
2.0 - %
pa a
1.0 - cia a a
a Se coo %
D ° nil a
a a
aa a
0.0
jeep te is. aa
El «Fr, a " .
-1.0 - a
ti•4 pia 41410 ° rb a a
a
-2.0-
-3.0
-3.0 -2.0 -1.0 0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 10.0 11.0
PC,
Figure 16a. Score plot of the first and second principal component for
low phosphorus ore.
30
P C2 1.0 op
0.5 -_
ci A1202
a Fe 0 Ca0
0.0
SilD2
ri Mg°
••••1
—0.5
—1.0
—1.0 —0.5 0.0 0.5 1.0
PC1
Figure 16b. Loading plot of the first and second principal component for
low phosphorus ore.
In the high phosphorus ore calcium oxide and phosphorus are well correlated
(Figure 15b), which may be verifying the assumption that the two minerals are
contained mostly in apatite. Quartz, aluminium trioxide, and magnesium oxide
are also fairly well correlated, indicating that they may all be belonging to the
same rock type, possibly a side rock (gangue). Thus, Figure 15a in principle
illustrates the interaction between three different rock types.
In Figure 15a the first principal component describes the variation between iron
and other minerals, while the second principal component separates the apatite
from gangue minerals. In the figure the variation along line (1) is an interaction
between only iron and apatite while the variation along line (2) is an interaction
between only iron and gangue minerals. Most of the data points are obtained
from an area where the iron-apatite replacement seems to dominate. However, a
number of data points indicate mainly the iron-gangue interaction. Most of these
data points are located close to the hanging wall, but a few are spread out in the
rest of the test area. There is also a large group of da.L. points (between line 1
and line 2) where both the content of apatite and gangue minerals increases
when the iron content decreases. In the further analysis, this group of data
requires special attention.
In the loading plot for the low phosphorus ore (Figure 16b) the interaction of the
different substances looks slightly different. Calcium oxide and phosphorus do
not correlate, indicating that some of the calcium oxide is not contained in
31
apatite. Calcium oxide instead correlates well with quartz, aluminium trioxide
and magnesium oxide. In the low phosphorus ore however, the amount of
accessory minerals is less, which may result in lower accuracy and higher physical
disturbances in the analysis of cuttings.
The effect on drilling parameters can be divided into three basic categories:
To make reliable predictions of the iron ore content, features due to non-
continuous properties and measurement errors must be separated from those
related to the variation in iron content.
The first attempt to predict iron ore content from drilling parameters was made
for a single hole (29896200). The important issue here is to study if the variation
in iron ore content affects the mechanical properties of the rock, and if this effect
is large enough to allow predictions of the ore content on the basis of the
characteristics of only five drilling parameters, namely thrust, torque, penetration
rate, percussive air pressure and rotation speed.
A first PLS run was performed using the five drilling parameters in the X-block
and only Iron in the Y-block, see chapter 3.2.2. The data from the hole was
averaged over every 10 cm, resulting in 447 objects along the hole. The iron
content was measured once for each tube. For the analysis it was assumed that
the samples were taken at the middle of the tube, and a simple interpolation
procedure was used to obtain the iron content every 10 cm between the sample
points. The ti and ui vectors for the first PLS run are shown in Figure 17.
32
the hole. The iron content in this zone may therefore not be well predicted with
the same model used in the rest of the hole.
1.0 a e a
Da 0
aa
ar?cteraa aScloZE Sg3 %.1
a a 13
0.5 - 0 D3 cra3413%eiek, 23rfi afaacie aa
0.0 -
-0.5-
o
aa
a
-1.0 - a
a
a
a
-1.5- a
a
-2.0 -
-2.5 II I II , 1 •1
-3.5 -2.5 -1.5 -0.5 0.5 1.5 2.5 3'.5
ti
In order to model the entire hole, the data set was divided in two parts, one part
consisting of the recorded data in the fracture zone between 25 and 31 m in the
hole (64 objects) and one part consisting of the rest of the hole (383 objects).
Two separate models was created for each data set, model 1 for data from the
solid part of the hole and model 2 for data from the fracture zone.
Comparison between results of the prediction using model 1 and the cutting
analysis is shown in Figure 18a. Figure 18b shows a comparison between results
of the prediction using model 2 and the cutting analysis.
33
Iron Content Os )
75
45 1 1
0 5 10 15 20 25 30 35 40 45 50
Hole Length (in)
Figure 18a. Prediction of iron content for hole 29896200 with model 1.
70 - % Cutting Analysis
_ ,---s. ,..'s.__,....."--,
..
..
je
,.........../...—.............. «.... s,... Prediction
inf. . .L.x...i.
65 -
_
60 - • ..."
s„---
55 -
50 -
0 5 10 i15 20 25 30 35 40 45 50
Hole Length (in)
Figure 18b. Prediction of iron content for hole 29896200 with model 2.
34
In order to select the best model for each object, the DCL value was utilized. In
Figure 19, the DCL values for both predictions in Figures 18a and 18b are
presented. From the two curves the selection of the model with the smallest DCL
value can be made. From the figure, it can be seen that the DCL value for model
2 is smaller than the DCL value for model 1 only in the fracture zone.
DCL
3.5
3.0 -
J
2.5 -_
2.0 -
_
1.5 -
0.5 -_
Figure 19. DCL values for the predictions in Figures 18a and 18b.
The final prediction using both models, selected by the size of the DCL value, is
presented in figure 20.
35
Iron Content (x)
75
70 Cutting An lysis
Prediction ..
65 -_
60 -
55 -_
50
45 /111111(1(111f ill111111IIIE1 . 1
0 5 1C) 15 20 25 30 35 40 45 50
Hole Length (m)
In order to slightly improve the overall prediction, single predictions with large
DCL values may be removed from the final prediction.
The above example verifies that there is a correlation between the iro' n content
and drilling response for the hole, and that the correlation is significant enough
to make predictions of iron content with a reasonable accuracy. However, since
the prediction in this case has been done using the same data that is used for
model development, no conclusions can be made as to how universal the model
is.
36
5.3.2 Prediction of iron ore content in the high phosphorus area
In this section an attempt is made to make predictions for large areas in the
OSCAR site in Kiirunavaara mine, by developing a model for the ore containing
the high phosphorus. High phosphorus ore is in this case defined by a
phosphorus content exceeding 0.25 %.
Here, the drilling data was averaged over each drill tube since we could obtain
only one geological data sample per tube. Input data with both drilling and
geological information was obtained from 295 tubes in the following 14 holes;
In this case, half the data set was used to create models for the iron content in
the area, while the second half was used to check the models validity. Every
other 20 points in the initial data set was selected as modeling data set and the
rest as a check data set.
Initially an attempt was made to model the iron content of the entire area with a
single model, (see figure 21a). All drilling parameters including their products
and their squared value (all together 20 variables) are gathered in the X block
(see Section 3.2.2) while the iron content is in the Y block. This first attempt has
been done without rejecting any of the data points.
Single prediction in this case has poor precision, and the model has a tendency to
average the content of iron over the area.
A closer study of the prediction including both quartz and phosphorus contents
(Figures 21b and 21c) indicates that in zones with low quartz content (e.q. tubes
30 to 80) the iron content is in general under-estimated by the model, except for
samples with very high content of phosphorus where the model gives reasonable
predictions. On the other hand the model generally over-estimates the iron
content in zones with high quartz content.
A conclusion from the first prediction is that quartz content has a large effect on
the quality of the prediction. The content of quartz has therefore to be taken into
account when modelling. However, since the interaction between the drilling
parameters varies according to rock, quartz effect on the prediction cannot be
fully described in a single model.
37
11 1
1 1I1
It 11
if If 1 11
I III
i Ii
Prediction II I
I/
Cutting analysis
O s °lllIIIIiIiI 111111!
15.0 -
10.0 -
_
5.0
0.0
0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150
Tubes (index)
Figure 21b The quartz content from the check data set.
6.0 -
4.0 -
_
2.0 -
0.0
0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150
Tubes (index)
Figure 21c The phosphorus content from the check data set.
38
In this case four different models were developed for the high phosphorus zone
of the mining area. The differences between the models, distinguished by the
content of quartz, is given in table 6.
Table 6
Since the hanging wall has a very high content of quartz and contains hardly any
phosphorous, this zone must be treated separately. However since there were
just a few data points from the hanging wall it was not possible to separate this
zone for modeling and verification purposes. Therefore no hanging wall objects
were present in the check data set.
In Figure 22 the prediction of the iron content in the test data set with four
different models is presented, where the selection of models for individual data
points is based on the DCL value. A substantial improvement in prediction
compered to figure 21a is observed except for a few tubes.
The selection of model for different tubes in different holes for the prediction in
figure 22 is shown in figure 23.
39
Iron content (%)
70.0
Prediction
65.0 -
60.0
55.0
50.0
•
45.0
40.0
35.0
Cutting analysis
30.0 -
25.0 ---
20.0 — 11111111111111111l111111111111111111111)11TrillliffT11-11111[FilliiIIIiIII{fil
0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150
Tubes (index)
Model
4.0
3.0 - 1
iv> 'Om
40
6. LOCATING ORE BOUNDARIES FROM DRILL
MONITORING DATA
6.1 INTRODUCTION
6.2 GEOLOGY
The rock on the hanging wall side of the Kiinmavaara ore body consists of quartz
porphyry and the foot-wall side syenite porphyry. The hanging wall consists of a
contact zone with varying thickness. Figure 24, shows an example of how the
dominating substances and minerals from a bore hole in the area are changing,
when drilling from the ore through the contact zone and into the side rock. The
Figure is based on cutting samples taken during the drilling operation.
70.0
50.0 -
40.0 -
30.0 -
20.0 -
10.0 -
0.0
0 3 5 8 10 13 15 18 20 23 25 28 30 33 35 38 40
Hole length (m)
41
6.3 DATA ANALYSIS
In Figure 25, the recorded drilling parameters from a test hole (hole 29024200) is
shown. The hole is drilled from the ore body, almost parallel to the hanging wall,
into the side rock. To make a definite conclusion about the location of the ore-
side rock contact based on the figure would be difficult, and the risk of drawing
the wrong conclusion is obvious.
150.0
Penetration Rat
125.0 -
100.0
75.0 1-
50.0
25.0
ure
0.0
10 11 12 13 14 15
Hole length (in)
Figure 25. Recorded drilling parameters versus hole length for hole
29024200.
In this case the X-block contained all drilling variables while the Y-block
contained the geological information from the bore hole. For this specific
application it was sufficient to use the iron content as an indicator of the ore
boundaries. The task was then to find the relation between the blocks, so that the
iron content could be predicted based on the drilling parameters.
42
In Figure 26, the prediction of iron content is shown for bore hole 29024200. In
the figure, the cutting analysis on which the model is based is also included.
Fe-content (%)
70.0
ORE CONTACT ZONE SIDE ROCK
1
60.0 -
_
Prediction of Fe
content based on
drill parameters s,
Fe content based
on cutting analysis
0.0 ‘„,„„„iliiii-lu,-1-11111111irrri-rfr i ri
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 10.0 11.0 12.0 13.0 14.0
Hole length (m)
Figure 26. Prediction of iron content for the contact zone between ore
and side rock. The cutting analysis is given as a reference.
It is seen from the figure that the overall agreement between the predicted value
and the cutting analysis is reasonably good, even though the curve of the
predicted iron content appears slightly noisy.
Let us assume that the true iron content in the rock agrees with the prediction
shown in Figure 26. The analysis of drill cuttings may still coincide with the
dashed curve in the Figure 26, due to the fact that samples are collected at the
surface far away from the point where they are generated. Thus, this introduces
substantial mixing and segregation during transportation to the surface.
Therefore analysis of cuttings will always reflect an average of the grades in the
bore hole.
The capability of the PLS method to identify major changes in the rock seems to
be good. The hanging wall contact can clearly be recognised at 5.6 m depth. A
possible streak of side rock between 4 and 4.5 m is also indicated. The thickness
of the contact zone is approximately 3 m (between 6 and 9 m). But since the hole
43
penetrated the hanging wall at an oblique angle, the true thickness of the contact
zone is probably much less.
Using PLS the positions of different features along the hole have been identified
almost accurately. The hanging wall contact can in this case be identified within
1 cm, which is the length between individual measurement points.
The above analysis is based on the assumption that conditions of the hanging
wall zone does not vary in the area. However, hanging wall conditions such as
geology and geomechanics will in some cases vary substantially, while the ore
often is more defined and properties more uniform.
To solve this problem a model based only on data from ore was created. The
location of ore boundaries was then based only on the DCL value (Figure 27).
With this technique it is not necessary to be concerned with properties outside of
the ore. Rather, it is only needed to identify whether it is different from ore or
not.
2.5 -
2.0 -
e(
1.5- DCL
1.0
0.5 - 1
0.0 III{
0 1 2 3 4 5 6 7 8 9
Hole length (m)
Figure 27. The DCL value versus hole length for bore hole 29024200.
The model is based only on iron ore.
44
7. DISCUSSION AND CONCLUSION
In this thesis it has been demonstrated how multivariate techniques can be used
to treat this type of data.
It has been demonstrated that each recorded variable carries information about
the rock, even if individual parameters have different magnitude and importance
for specific properties in the rock mass. It is therefore important to look at the
total interaction between all the recorded drill parameters in the analysis.
To identify outliers is important in any type of analysis. In this case a point with
high penetration rate in combination with high air pressure and low torque can
be an outlier, even though none of the individual drilling parameters show any
abnormal behaviour. Such outliers are normally difficult to detect by using only
one drilling parameter.
45
The score plots provide a substantial simplification of the interpretation of
drilling data compared to ordinary plots of each raw parameter versus hole
length, for instance. The score plots also allow grouping of specific features in
drilling data and the evaluation of similarities and differences between individual
data points or groups from the hole, see Figures 10a and 13a. For example, the
separate locations for each of the three groups in Figure 10a indicate the
difference in relations between the drilling parameters.
In the corresponding loading plots (Figures 10b and 13b) the inter-relations
between all the drilling parameters can be evaluated. In Figure 10b, the
interaction of the drilling parameters for each group of data in Figure 10a, can
be evaluated, and the significance of each parameter in relation to each data
group established.
PLS is an efficient tool to relate specific properties in the rock mass to the
response of drill parameter.
The difficult step in this analysis is to choose (from the entire test data set)
reliable modelling test sets for different drilling conditions, each accurately
reflecting the properties in the rock mass that one likes to predict. To do this, a
detailed record of this property in the test hole is required, but detailed
knowledge of all other properties likely to affect the analysis is also necessary.
Once reliable test sets has been selected this can be used to predict the iron
content of a new object, for example. The DCL value related to each prediction
gives an indication of how reliable the prediction is, and can therefore also be
used to select the best model, see Figure 19. Furthermore, a persistently high
DCL value can indicate that the hole is drilled into a new type of rock.
Multiple models were also used in the attempt to predict iron content over a
lager area. Here, one rock model was not sufficient to describe the relation
between drilling parameters and iron content for all types of rock. Several
models, each covering a more well defined type of rock, was used to improve the
prediction.
46
In order to separate the data population into smaller more uniform groups a first
PCA of the whole data population can be helpful. This separation can also be
based on detailed geological knowledge through the response of drilling
parameters. The latter procedure was followed in this case.
One important task is to decide how many models to use in order to describe a
whole area. The fewer the models the less accurate the rock description will be,
especially for uncommon types of rock properties. Use of many models may
however increase the risk of each model being too specialised. Furthermore, the
risk that the models may overlap leading to wrong prediction in combination
with low DCL value, will generally increase.
The last part of the thesis shows how multivariate technique can be used to
identify the boundaries of the ore body. With this technique this can be done at
low cost and with reasonably high accuracy. The information is also available
almost immediately after completion of the hole. Any other method available
today would be more time consuming.
47
8. RECOMMENDATIONS FOR FURTHER
STUDIES
This thesis mainly presents the multivariate techniques as tools to interpret drill
monitoring logs. The need for further research and practical field applications to
verify the potential of these techniques in different geological environments is
therefore large.
A major problem in this study has been to retrieve detailed information about
the rock that could match the accuracy of the drilling parameters that is
monitored for each centimetre of drilling. The use of geophysical logging in the
monitored hole would provide a more detailed record of the rock properties,
which is likely to improve analysis and predictions based on drilling parameters.
The use of more drilling parameters in the analysis is likely to improve the
accuracy of prediction of different properties in the rock mass. Machine
vibrations is one example that in simple form has been used for this application.
More sophisticated vibration monitoring can probably provide a number of new
parameters reflecting variations in different properties in the rock mass.
48
REFERENCES
Barr, M. V. (1984)
Instrumented horizontal drilling for tunnelling site investigations.
- Ph.D. thesis, Imperial Collage of Science and
Technology, London, UK.
Carlson, R. (1986)
How to optimize a chemical reaction.
- US-Swedish Symp. on Synthetic Selectivity, Tynningö,
Sweden, 25-29 Aug, 1986.
Frietsch, R. (1978)
On the magmatic origin of the iron ores of the Kiruna type.
- Econ. Geol., 73, 478-485.
Griffiths, C. M. (1983)
Strata recognition from a computer analysis of drilling variables.
- Ph.D. thesis, University of Newcastle upon Tyne,
Newcastle, UK.
Jolliffe, I. T. (1986)
Principal Component Analysis.
- Springer Verlag.
Leighton, 3. C. (1986)
Correlating rotary drill performance to powder factors for improved blasting
control.
- Int. Symp. on Geot. Stab. in Surf. Min., Calgary, 1986.
Parak, T. (1975)
The origin of the Kiruna iron ores.
- SGU, series C, Nr 709.
Peck, J. P. (1989)
Performance monitoring of rotary blasthole drills.
- Ph.D. thesis, McGill University, Montreal, Canada.
Pfister, P. (1985)
Recording drilling parameters in ground engineering.
- (Geodrilling, Okt.)(Ground Engineering, April), 1985.
Sinkala, T. (1989)
Hole deviations in percussion drilling and control measures.
- Ph.D. thesis, Luleå University of Technology, Luleå,
Sweden.
Stähle, L. (1987)
Pharmacological studies on behavioural changes induced by dopamine agonists
in the rat: A multivariate approach.
- Ph.D. thesis, Karolinska Institutet, Stockholm, Sweden.
Wold, H. (1966)
Nonlinear estimation by iterative least squares procedures.
- Research paper in statistics, Ed David FN, Wiley, New
York.
50
ISSN 0280-8242
HÖGSKOLANS TRYCKERI
LULEÅ