Content uploaded by Andrew J Hill
Author content
All content in this area was uploaded by Andrew J Hill on Mar 07, 2022
Content may be subject to copyright.
arXiv:2202.02959v1 [cs.LG] 7 Feb 2022
Mathematical Geoscience manuscript No.
(will be inserted by the editor)
A Machine Learning Approach for Material Type
Logging and Chemical Assaying from Autonomous
Measure-While-Drilling (MWD) Data
Rami N. Khushaba ·Arman Melkumyan ·
Andrew J. Hill
Received: Dec 2020 / Accepted: date
Abstract Understanding the structure and mineralogical composition of a region is
an essential step in mining, both during exploration (before mining) and in the min-
ing process. During exploration, sparse but high-quality data is gathered to assess the
overall orebody. During the mining process, boundary positions and material proper-
ties are refined as the mine progresses. This refinement is facilitated through drilling,
material logging, and chemical assaying. Material type logging suffers from a high
degree of variability due to factors such as the diversity in mineralization and geology,
the subjective nature of human-measurement even by experts, and human errors in
manually recording results. While laboratory-based chemical assaying is much more
precise, it is time-consuming and costly and does not always capture or correlate
boundary positions between all material types. This leads to significant challenges
and financial implications for the industry, as the accuracy of production blast-hole
logging and assaying processes is essential for resource evaluation, planning and ex-
ecution of mine plans.
To overcome these challenges, this work reports on a pilot study to automate the
process of material logging and chemical assaying. A machine learning approach has
been trained on features extracted from measure-while-drilling (MWD) data, logged
from autonomous drilling systems (ADS). MWD data facilitates building profiles of
physical drilling parameters as a function of hole depth. A hypothesis is formed to
link these drilling parameters to the underlying mineral composition. The pilot study
results in this paper demonstrate the feasibility of this process, with correlation coef-
ficients of up to 0.92 for chemical assays and 93% accuracy for materials detection,
depending on the material or assay type and their generalization across the different
All authors are within the Rio Tinto Centre for Mine Automation, Australian Centre for Field Robotics,
The University of Sydney
8 Little Queen street, Chippendale, NSW 2008.
Tel.: +61 2 9351 4209
E-mail: Rami.Khushaba@sydney.edu.au,
Arman.Melkumyan@sydney.edu.au,
Andrew.Hill@sydney.edu.au
Corresponding author ORCID: https://orcid.org/0000-0001-8528-8979
2 Rami N. Khushaba et al.
spatial regions. The achieved results are significant, showing opportunities to guide
further drilling processes, provide chemistry data with a down-hole resolution, and
continuously update mine plans as the mine progresses.
Keywords mining ·measure-while-drilling ·logging and assaying ·machine
learning
1 Introduction
The mining industry has embarked on a journey to use automation to improve the
accuracy and consistency of mining processes within its surface mines. Blast-hole
drilling is a key task in most surface mining, as the accuracy of the information ob-
tained affects the entire downstream mining process. This has impacts on i) schedul-
ing, ii) excavation, iii) slope stability, iv) material handling, v) beneficiation, vi)
ore loss and vii) final product blending (McHugh et al. 2012). Measurement-while-
drilling (MWD) enables the collection of accurate, fast and high-resolution informa-
tion from production blast-hole drills for equipment automation and monitoring the
health of major drilling items (Rai et al. 2015). In open-pit mining, MWD systems
monitor several performance factors including, but not limited to, i) rate of penetra-
tion, ii) torque, iii) rotation pressure, iv) specific energy of drilling, v) weight on bit
and vi) rotary speed. These factors are becoming standard features on the blast-hole
drill rigs supplied by most equipment manufacturers (Hatherly et al. 2015). MWD
data has already been utilized in applications related to i) rock type recognition
(Zhou et al. 2 010), ii) boundary identification and surface upd ates (Silversides and Melku myan
2020), iii) automated coal seam detection (Leung and Scheding 2015), iv) estima-
tion of rock mass rating during tunneling (Galende et al. 2018), v) rock fracture den-
sity characterization (Khorzoughi et al. 2018), vi) improving rock-breakage efficien-
cies (Park and Kim 2020), vii) developing a prediction model of over- and under-
excavation depths from blasting (Navarro et al. 2018), viii) characterizing a coal mine
roof (Khanal et al. 2020), and ix) identifying the top of coal seams allowing drilling to
be halted to prevent unintende d blasting of coal and associated problem s (Hatherly et al.
2015). However, an extensive literature review has not uncovered any use of MWD
data to estimate material types or chemistry assays.
Boreholes and blast-holes are routinely logged and assayed to assist in the under-
standing of th e structures and mineralogical co mpositions of an area (Sommerville et al.
2014), the accuracy of which is essential for resource evaluation and planning in the
minerals industry. Material logging is the process of recording manual geological
observations to identify the material types present in the sample, and chemistry as-
saying is generally a lab-based process to quantify the chemical make-up of samples
(valuable minerals, impurities and water content).
To standardize primary characteristics, such as mineralogy and texture, in a hier-
archical manner, the Material Type Classification Scheme was developed (Box et al.
2002; Wedge et al. 2019). This scheme models physical and chemical attributes for
predicting metallurgical behavior and the quality of the product for optimal ore pro-
cessing (Paine et al. 2016). Logging of material-types is usually performed on chip
Title Suppressed Due to Excessive Length 3
samples produced by reverse circulation (RC) drills during exploration (in 2m inter-
vals) or blast-hole cone samples (one sample per blast hole). This generates a huge
number of samples and significant costs and time in preparation and assaying. For
each sample, a geologist manually handles the material and estimates several param-
eters including (Sommerville et al. 2014; Wedge et al. 2018): (i) the percentages of
various material types present in the sieve (usually in increments of 5%), (ii) the
sample color, (iii) the shape of the chips, and (iv) the percentage of material recov-
ered. Another sample is usually sent for laboratory X-ray fluorescence assay analysis
to measure i) aluminium oxide (Al2O3), ii) iron (Fe), iii) silicon dioxide (SiO2), iv)
phosphorus (P), v) sulfur (S), vi) manganese (Mn), vii) magnesium oxide (MgO),
viii) titanium dioxide (TiO2), ix) calcium oxide (CaO), and x) total loss on ignition
(LOI). However, this approach has limitations as exploration holes are sparse in the
horizontal direction, while horizontally denser blast-holes provide only a single sam-
ple for the entire hole (e.g. 10-15m depth). For these reasons, the literature criticizes
assaying from RC boreholes or blast-hole samples in that they do not always pro-
vide truly representative analysis (Clark and Dominy 2017), are not always effective
at characterizing thin layers, and the sampling is costly to perform. As a result, there
has been an increased interest in downhole, in-situ assaying. Downhole Pulsed Fast
and Thermal Neutron Activation (PFTNA) has been proven through its operations
to date to provide a downhole in-situ assaying technique that has several advantages
over conventional sample-based assays in terms of safety, cost, cycle-time of results,
and accuracy (Jeanneau et al. 2017; Market et al. 2019). By removing the need for
manual samples taken on-site, tools relying on PFTNA technology limit the exposure
of workers to manual handling, operational and environmental safety risks typical on
many mine sites.
It is important to mention here that the use of PFTNA-based technologies alone
does not completely obviate the need for manual handling for logging purposes, as
such technologies are primarily used for assaying, while material logging is not facil-
itated by PFTNA. It has been reported that using PFTNA for in-situ measurements,
as an alternative to traditional chemical analysis, may be overly-simplistic, as it does
not provide some vital information such as hardness (Smith et al. 2015). Addition-
ally, PFTNA does not produce estimation for all assays, especially trace elements
which can be of particular importance in some deposits, such as P in iron ore, or Au
and Ag in copper.
This paper argues that improvements in drilling technology can change the way
mineral deposits are observed. The motivation of this work is driven by the fact
that MWD data has already been deemed useful for several applications (Zhou et al.
2010; Silversides and Melkumyan 2020; Leung and Scheding 2015; Galende et al.
2018; Khorzoughi et al. 2018; Park and Kim 2020; Navarro et al. 2018; Khanal et al.
2020; Hatherly et al. 2015), which lends itself as a potential tool for determining the
material-types and chemical assays. Hence, the primary hypothesis of this work is
that using MWD and machine learning can help predict chemical assays and materi-
als. This is a novel application of MWD data, which has not been previously verified
and is based on a common conclusion that the drilling behavior is related to the me-
chanical properties of the material being drilled. The result would be a downhole
in-situ assaying or logging “equipment-as-a-sensor” tool that lends itself particularly
4 Rami N. Khushaba et al.
well to bulk mining operations such as that found in the iron ore operations in the
Pilbara 1. Additionally, MWD data has good down-hole resolution and can be made
available in real-time, during drilling. A key challenge is the difficulty modelling the
bit–rock interaction because of high-order variability in different rocks or even sim-
ply at different sample points in the same rock material due to the presence of cracks,
fissures and a host of other discontinuities (Rai et al. 2015). This is where the use of
machine learning (ML) models to capture the relation between the MWD data and
materials or assays become increasingly important. The application of ML in this
problem is supported by wide-ranging successes of using ML in several applications
in the era of big data (L’Heureux et al. 2017).
2 Data Sources
This pilot study was conducted on an Iron Ore mine site in the Hammersley Range,
which is located in the Pilbara region of Western Australia. MWD, material logging
and chemistry assay data was collected from two physically separate regions, labelled
here as Regions A and B.
2.1 MWD Data
MWD data represents real-time measurements of several mechanical signals, col-
lected from sensors equipped on relatively large drill-rigs used in mining for blast-
hole drilling. These signals are commonly used to control and monitor the perfor-
mance of drilling. Fig. 1 shows the autonomous blast-hole drill rig that collects the
MWD data used in this paper. The parameters of interest considered in this paper in-
clude: i) the time taken for hole development (start to finish), ii) measurement depth,
iii) average head rotation speed (rotationRPM), iv) average bit air pressure (airPres-
sure), v) average feed pressure (feedPressure), vi) average torque (torque), vii) av-
erage rate of penetration (rop), viii) average force on bit (fob), ix) average rotation
pressure (rotationPressure), x) adjusted penetration rate (apr), and xi) specific energy
of drilling (sed). Also considered is the ratio of rotationPressure to feedPressure as
another parameter. The MWD time-series data were logged by the drilling system
as discretized sequences, typically discretised into depth segments of 0.1 meter res-
olution (Hatherly et al. 2015). A total of almost 7000 holes’ data were logged by the
ADS system, as MWD is available for every hole. This is not the case for materi-
als/assays data as mentioned in the next section.
2.2 Logging and Assaying
The logged material-types and chemistry assay data were provided from multiple
blasts within both Region A and Region B. While the autonomous drilling system
1https://en.wikipedia.org/wiki/Pilbara
Title Suppressed Due to Excessive Length 5
Fig. 1 Autonomous blast-hole drill rig used for collecting experimental MWD data
provided MWD data along the full depth of every blast-hole, the data for material-
types and chemistry is only available for a subset of the holes, and where available is
a single value per hole for each material and chemistry parameter (depicted in Fig.2).
This is primarily due to the cost of data collection, including manual logging, manual
physical sampling and laboratory analysis costs. Additionally, this data is delayed by
the collection and analysis time, while MWD data is collected electronically and is
immediately available after drilling.
The resulting data set comprises MWD for all holes, and subsets of holes that also
include material logging and/or chemistry assays, as noted in Table 1.
Additionally, wherever material logging is available, a theoretical chemistry as-
say is also estimated by geologists. A validation process is usually performed by
geologists after logging, adjusting the percentages of the logged material types or
adding/removing material types where necessary, using geologically informed sub-
stitutions so that the theoretical assay values are within an error tolerance of the lab-
oratory assay values obtained from the interval’s chip samples (Wedge et al. 2018).
6 Rami N. Khushaba et al.
Data Sources Region A Region B Total
MWD + Material 805 993 1,798
MWD + Chemistry 2,176 2,408 4,584
MWD + Mat. + Chem. 537 804 1,341
Table 1 Material logging and chemistry assays are collected more sparsely than
MWD data, resulting in multiple subsets of holes with different combinations of data
for the same hole
MWD data
rotationRPM
airPressure
feedPressure
torque
rop
fob
rotationPressure
apr
sem
sed
measurementDepth
Aautonomous Drill
Machine learning
algorithms
Chemical Assays
Materials Logging
The suitability of different
feature extraction,
reduction, regression and
classification methods is
investigated
Fig. 2 The three sources of data available for this pilot study. MWD is sampled at
0.1m intervals across the entire hole depth, while assay and material types offer a sin-
gle value per hole for each parameter, for the subset of holes where they are collected
3 Details of the Machine Learning Pipeline
3.1 Feature Extraction
The literature review showed that there is no generally agreed-upon quantitative
method for the description of the MWD data. As such, several feature extraction
methods are investigated in detail in this study. As our hypothesis attempts to relate
the changes in the drilling parameters to the changes in materials and chemical assays
underground, we have sought features focusing on changes in the amplitude, energy,
complexity, frequency contents and shape of the power spectrum, signals’ peak val-
ues, root mean square and entropy of the signals to denote the dispersion. The selected
methods have been utilized across different time-series data in other domains, demon-
strating significant success in describing signals (Hjorth 1970; Khushaba et al. 2014;
Phinyomark et al. 2010), and hence are seen as good candidates to employ to describe
the MWD signals. The investigated feature extraction methods are described below.
In the following features’ descriptions, we assume a signal x, with x[j]representing
its j-th sample, where j=0,2,3,...,N−1, with a length of Nsamples.
Title Suppressed Due to Excessive Length 7
Hjorth parameters:
These parameters were originally developed in an approach to describe the shape
of the frequency spectrum of any signal directly from the time-domain by using Par-
seval’s theorem and Fourier transform properties (Hjorth 1970). Several spectral anal-
ysis methods (fast Fourier transform, wavelets transform, Wigner-Ville transform,
plus many others) can be used to study the frequency contents of the underlying
signals. However, Hjorth parameters were chosen because of their low computational
cost which makes them favorable for any online data processing task (Khushaba et al.
2014). The description of these parameters starts by first observing Parseval’s theo-
rem, which states that the sum of the square of the function is equal to the sum of the
square of its transform
N−1
∑
j=0
x[j]2=1
N
N−1
∑
k=0
|X[k]X∗[k]|=
N−1
∑
k=0
P[k],(1)
where X[k]is the discrete Fourier transform (DFT) of the original signal x,P[k]is
the phase-excluded power spectrum, that is the result of a multiplication of X[k]by
its conjugate X∗[k]divided by N, and kis the frequency index. As the complete fre-
quency description derived by the Fourier transform is symmetric with respect to
zero frequency (having identical branches stretching into both positive and negative
frequencies) (Hjorth 1970), all odd moments will become zero. This is according to
the definition of a moment mof order nof the power spectral density of the signal x
which is given by
mn=
N−1
∑
k=0
knP[k].(2)
Hjorth parameters are hence mainly based on the lower order even moments,
denoted as m0,m2, and m4:
m0=
N−1
∑
k=0
k0P[k] =
N−1
∑
j=0
x[j]2,(3)
m2=
N−1
∑
k=0
k2P[k] = 1
N
N−1
∑
k=0
(kX[k])2=1
N
N−1
∑
j=0
(
∆
x[j])2,(4)
m4=
N−1
∑
k=0
k4P[k] = 1
N
N−1
∑
k=0
(k2X[k])2=1
N
N−1
∑
j=0
(
∆
2x[j])2,(5)
where
∆
nis the n-th derivative of a function in the time-domain, this is according to
the time-differentiation property of the Fourier transform. This property states that
the power spectrum of a signal in the frequency domain multiplied by kraised to
the n-th power is equivalent to n-th derivative of the same signal in the time-domain.
Hjorth parameters of interest are then calculated based on the above moments as
shown below
Activity =m0,(6)
8 Rami N. Khushaba et al.
Mobility =rm2
m0
,(7)
Complexity =qm4
m2
Mobil ity ,(8)
Waveform Length (WL):
This is an intuitive measure to describe the cumulative length of the waveform
over the considered segment which can also indicate the waveform complexity. The
resultant value of the W L calculation also indicates a measure of the waveform ampli-
tude, frequency, and duration (Hudgins et al. 1993). For a given hole depth, the WL
feature value grows larger for signals with higher frequencies than those with lower
ones. This is given by
W L =
N−2
∑
j=0
|x[j+1]−x[j]|,(9)
Simple Square Integral (SSI):
SSI captures the energy of the signal under consideration as a feature (Phinyomark et al.
2010). It can be expressed as
SSI =
N−1
∑
j=0
|x[j]|2,(10)
Crest Factor (CF):
CF is defined as the ratio of the absolute value of the peak in the signal under
consideration divided by the Root Mean Square (RMS) value of the same signal
(Silva 2005). The crest factor indicates how extreme the peaks are in a waveform,
with CF of 1 indicating no peaks and larger values indicating more peaks. CF also
expresses the size of the dynamic range for an input signal. This is given by
CF =|xpeak|
xRMS
,(11)
where
xRMS =v
u
u
t
1
N
N−1
∑
j=0
x[j]2,(12)
Pressure Ratio (PR):
A new measure was created by taking the ratio of the rotationPressure to that
of the feedPressure, this was inspired by the earlier work of Leung and Scheding
(2015). This ratio forms a derived drill performance indicator, the characteristics of
which this work attempts to link to the different materials type and chemical assays.
These characteristics included all the following:
I summation of the absolute difference between consecutive samples of the PR
Title Suppressed Due to Excessive Length 9
II summation of the PR squared (SPR2) and its logarithmically scaled version
III summation of the first derivative of the PR squared (SDPR2)
IV summation of the second derivative of the PR squared (SDDPR2)
V a logarithmically scaled version of SDPR2/SPR2
VI a logarithmically scaled version of SDDPR2/SDPR2
VII the maximum of PR multiplied by the max of f ob.
Singular-Value Decomposition Entropy (SvdEn):
The dispersion of the singular values
λ
kalso indicates the complexity of the sig-
nal dynamics and is an indicator of how many vectors are needed for an adequate
explanation of the signals being studied. To calculate SvdEn, the singular values are
first normalized by their total summation
¯
λ
k=
λ
k
∑
λ
k
,(13)
where ∑¯
λ
k=1. To denote the dispersion characteristics of the singular values, SvdEn
is defined with the Shannon formula applied to the elements of singular values of the
matrix and is calculated as shown below (Li et al. 2008).
SvdEn =−∑¯
λ
klog¯
λ
k,(14)
SvdEn can also be considered as a measure of feature-richness in the sense that
the higher the entropy of the set of SVD weights, the more orthogonal vectors are
required to adequately explain it.
Signal Flatness (F):
This is defined as the ratio between geometric and arithmetic means and is con-
sidered as an important measure to distinguish signals that are flat or do not change
much, from those that have the amplitude concentrated across small ranges. A high
signal flatness (approaching 1.0) indicates that the signal has a similar value across
all samples, while a low flatness (approaching 0.0) indicates that the power is con-
centrated in a relatively small number of samples (Johnston 1988; Dubnov 2004).
Flatness =
N
q∏N−1
j=0x[j]
∑N−1
j=0x[j]
N
=
exp1
N∑N−1
j=0ln x[j]
1
N∑N−1
j=0x[j],(15)
Descriptive Statistics:
The following common statistical measures are also included as features: the sig-
nal maximum, standard deviation, skewness, kurtosis, mean, geometric mean, and
median of each signal.
3.2 Regression and Classification Models
Several regression and classification models were utilized in this research. The Sup-
port Vector Machines (SVM), multivariate Gaussian Process (GP), and Random Forests
(RF) models were all utilized for MWD versus assays regression analysis, while
10 Rami N. Khushaba et al.
SVMs classification models were also utilized for the detection of the different material-
types as described in the experiments section. The details of these models are omitted
from this paper as these are traditional algorithms for which the description can be
found in general pattern recognition references (Theodoridis and Koutroumbas 2009;
Rasmussen and Williams 2005). These models were utilized with the extracted MWD
features described in the previous section, with the output of the models being the es-
timation of the chemical assays for regressions models and the classification labels
(material detected or not) for the material detection problem.
The testing scheme utilized with the above models employed a cross-validation
process across two modes, these are: (i) random k-fold cross-validation (Stone 1974;
1977) and (ii) spatial k-fold cross-validation (Roberts et al. 2017; Meyer et al. 2019;
Talebi et al. 2020). Overall, the goal of cross-validation is to test the model’s ability to
predict based on new data that was not used for training the models, to flag problems
like overfitting or selection bias (Cawley and Talbot 2010), and to give an insight on
how the model will generalize to an independent dataset (an unknown dataset). While
the cross-validation procedure is the same for random and spatial cross-validation
(the process of dividing the data into separate segments for training and testing),
the major difference is how the data points are split into folds. The chosen mode of
validation is hence largely dependent on the goal of the experiments.
In the case of random k-fold cross-validation, it is generally known that blast-hole
sampling is a sparse process, where not all the drilled holes would be considered for
logging physical materials or laboratory sampling of chemical assays. In this regard,
the general approach is to sample every n-th hole, with nvarying across the mine sites
depending on design specifications (roughly 1/6 or 1/8 for materials logging and 1/4
for assays sampling). A potential application of this work would be to fill these data
gaps with inferenced materials and chemical assays, using the nearby holes where
data has been recorded as the training set. This mimics the process of random k-fold
cross-validation, and the short distance scales do not require spatial k-fold cross-
validation for this demonstration. The random k-fold cross-validation experiments
are included as a proxy for this use case.
In the case of spatial k-fold cross-validation, recent literature has shown that when
using a model trained on spatial data to make inferences on data collected from a spa-
tially distant environment, the commonly used random cross-validation may provide
considerably over-optimistic error estimates due to the problem of spatial autocorre-
lation (Roberts et al. 2017; Meyer et al. 2018). Cross-validation strategies based on
random data splitting fail to assess models’ performance in terms of spatial mapping.
If the objective is to test the model’s performance upon spatially distant datasets,
then a spatially-aware cross-validation scheme will provide more realistic outcomes
to accurately estimate the generalization errors. In this regard, we have implemented
a leave-one-blast-out scheme of testing in which the data from all holes belonging to
a specific blast is kept away for testing, while the remaining data from all other blasts
are used for training (a process that is repeated across all blasts). This mimics the use-
case of predicting assays in a new blast in real-time from MWD data, before assays
and material logs take place, or for selected blasts where these are not economically
viable.
Title Suppressed Due to Excessive Length 11
For most of the regression models, Bland-Altman graphs and QQ-plots are shown
to verify the outcome of these models in comparison to the laboratory measurements.
To verify the results statistically, the following measures are also shown on the cor-
relation graphs:
Ieq - slope and intercept equation
II r- Pearson r-value
III RMSE - root mean squared error
IV p- Pearson correlation p-value
Vn- number of data points used
VI RPC - reproducibility coefficient (1.96*SD)
VII CV - coefficient of variation (SD of mean values in %)
The Bland-Altman plot is presented as a scatter plot in which the x-axis represents
the average of a pair of measurements (A+B)/2, and the y-axis shows the difference
between the two paired measurements (A−B). The Bland-Altman plot allows visual
inspection for several aspects of the comparability of outcomes of the utilized model
against the actual laboratory measurements. First, a consistent measure of bias can be
described. This is the mean of all the differences between the algorithms’ outcomes
and the laboratory measurements. This mean is represented as a line across the x-
axis of the plot, with the difference between this value and y-values describing the
magnitude and direction of the bias. Bias can be reported in absolute terms or as a
percentage (bias/mean value). The Bland-Altman plots are also utilized to study any
proportional bias between the model outcomes and the laboratory measurements. The
existence of proportional bias indicates that the model and laboratory measurements
do not agree equally through the range of measurements (the limits of the agreement
will depend on the actual measurement).
4 Experimental Results
In this section, the feasibility of predicting several assays from MWD data with the
various regression models is investigated. The proposed approach is demonstrated
on selected assay types, as a general proof of concept. The results in the following
sections are acquired using the random k-fold cross-validation mode (as filling in the
gaps was the primary motivator of this research), unless otherwise specified that the
spatially-aware cross-validation mode was used.
4.1 Iron (Fe) Prediction
The regression results for predicting iron (Fe) using GP, SVM, and RF models are
shown in Fig. 3 and Fig. 4 for Regions A and B respectively. For Region A, both SVM
and GP had a Pearson correlation coefficient of 0.79 with the laboratory readings for
Fe, with an RMSE of 2.7 and 2.8 respectively for GP and SVM. RF performed sim-
ilarly to GP and SVM, with a correlation coefficient of 0.78 and an RMSE value
of 2.8. For Region B, both GP and SVM again performed similarly, though with a
12 Rami N. Khushaba et al.
lower Pearson correlation coefficient of 0.64, and RMSE of 5.4, while RF achieved
a correlation coefficient of 0.63, and RMSE of 5.4. All regression models from both
sites had p-values less than 0.001, which indicates a significant correlation between
the Fe estimates from these models vs. that from the corresponding laboratory mea-
surements. From these results, it is clear that all models achieved better results on
Region A data in comparison to the performance on Region B. Therefore, further
analysis to understand the differences between Regions A and B were carried out.
The distributions of the laboratory measurements of Fe from both sites were then
studied to understand the differences between Regions A and B. Fig. 5 shows the
QQ-plot for both sites individually against the quantiles of normal distribution, and
also against each other’s quantiles. The results in Fig. 5 clearly demonstrate that Fe
samples from each site are not normally distributed and that the samples of Region A
Fe do not come from the same distribution as Region B Fe. Further evidence that
these distributions are different is seen in the histogram of the Fe samples from each
site, as shown in Fig. 6. Both histograms are skewed, but Region B Fe exhibits twice
the kurtosis of Region A Fe.
As the performance of these models on the Fe data was similar across both mining
sites, the estimates generated by the SVM models were selected for further analysis.
The histogram of the residuals, that is the difference between laboratory Fe mea-
surements and SVM Fe estimations, are shown in Fig. 7. These histograms clearly
indicate that the majority of Region A samples had a difference between laboratory
and estimated values of ±2.5, which is the typical tolerance in the assay error for Fe
and SiO2as indicated in the literature (Wedge et al. 2018). However, the Region B
residuals were concentrated around ±5. It is important to note here that the tolerances
are usually set independently for each element and may vary from project to project,
according to requirements or the interval’s iron grade, as less accurate validation is
required for low-grade (waste) intervals. In general, low Fe values of Fe <50 are
classified into waste, while 50 ≤Fe <60 are classified as med-grade and Fe ≥60
are classified as high-grade. When plotting the residuals across these three classes of
Fe grades, it was observed that the tail of the histograms in Fig. 6 with large negative
values were all related to low-grade iron as shown in Fig. 8. These larger errors occur
well under the waste cut-off, so are unimportant as the exact Fe content will not be
considered in further decision-making at such low values.
4.2 Phosphorus (P) Prediction
In these experiments, the two datasets from Region A and Region B were combined
to generate one large set of samples, as each site contains a relatively small number
of samples. The results in Fig. 9 demonstrate the performance of the machine learn-
ing models in predicting P values from the MWD data. In this case, the RF model
demonstrated the highest correlation with the laboratory measurements of 0.81, with
an RMSE of 0.03. The GP model had a correlation coefficient of 0.79, and SVM of
0.78, both with an RMSE of 0.03.
In order to understand more about what impacted these correlation values, we
plotted the difference between the laboratory P measurements and RF predictions of
Title Suppressed Due to Excessive Length 13
40 50 60 70
Algorithm
35
40
45
50
55
60
65
70
Laboratory
y=1.00x+0.02
r=0.79
RMSE=2.7
p<0.001
n=2176
40 50 60 70
Mean Algorithm & Laboratory
-15
-10
-5
0
5
10
15
Laboratory - Algorithm
5.4 (+1.96SD)
0.00 [p=0.94]
-5.4 (-1.96SD)
RPC: 5.4 (9.7%)
CV: 4.7%
(a) GP model results
40 50 60 70
Algorithm
35
40
45
50
55
60
65
70
Laboratory
y=1.09x-5.36
r=0.79
RMSE=2.8
p<0.001
n=2176
40 50 60 70
Mean Algorithm & Laboratory
-15
-10
-5
0
5
10
15
Laboratory - Algorithm
5.3 (+1.96SD)
-0.14 [p=0.02]
-5.6 (-1.96SD)
RPC: 5.5 (9.9%)
CV: 4.7%
(b) SVM model results
40 50 60 70
Algorithm
35
40
45
50
55
60
65
70
Laboratory
y=1.14x-8.00
r=0.78
RMSE=2.8
p<0.001
n=2176
40 50 60 70
Mean Algorithm & Laboratory
-15
-10
-5
0
5
10
15
Laboratory - Algorithm
5.6 (+1.96SD)
0.06 [p=0.36]
-5.5 (-1.96SD)
RPC: 5.6 (10%)
CV: 4.8%
(c) RF model results
Fig. 3 Region A Bland-Altman iron (Fe) regression plots with several models
P versus the laboratory P measurements. In this way, one can clearly see where the
laboratory and model disagree, and observe the actual range of P across the corre-
sponding samples. The results in Fig. 10 clearly show that very few of the approx-
imately 4000 samples had a measured P value >0.35, and these were essentially
‘missed’ by the model, as the model did not observe sufficient samples of this mag-
nitude to learn the relationship between MWD and P across that range. Hence, the
14 Rami N. Khushaba et al.
20 40 60
Algorithm
10
20
30
40
50
60
70
Laboratory
y=0.99x+0.43
r=0.64
RMSE=5.4
p<0.001
n=2408
20 40 60
Mean Algorithm & Laboratory
-30
-20
-10
0
10
20
30
Laboratory - Algorithm
11 (+1.96SD)
-0.01 [p=0.93]
-11 (-1.96SD)
RPC: 11 (22%)
CV: 9.2%
(a) GP model results
20 40 60
Algorithm
10
20
30
40
50
60
70
Laboratory
y=0.94x+3.19
r=0.64
RMSE=5.4
p<0.001
n=2408
20 40 60
Mean Algorithm & Laboratory
-30
-20
-10
0
10
20
30
Laboratory - Algorithm
10 (+1.96SD)
-0.54 [p=0.00]
-11 (-1.96SD)
RPC: 11 (21%)
CV: 9.1%
(b) SVM model results
20 40 60
Algorithm
10
20
30
40
50
60
70
Laboratory
y=1.09x-5.06
r=0.63
RMSE=5.4
p<0.001
n=2408
20 40 60
Mean Algorithm & Laboratory
-30
-20
-10
0
10
20
30
Laboratory - Algorithm
11 (+1.96SD)
0.21 [p=0.05]
-10 (-1.96SD)
RPC: 11 (22%)
CV: 9.3%
(c) RF model results
Fig. 4 Region B Bland-Altman iron (Fe) regression plots with several models
sample size available for this experiment did impact the accuracy of estimation. It is
believed that having much larger datasets with a significantly larger number of sam-
ples would enhance the models’ performance. Considering the available samples, the
performance of the models with an RMSE of 0.03 is quite promising, given that the
literature has no previously reported figures in this regard, to the best of the authors’
knowledge.
Title Suppressed Due to Excessive Length 15
-4 -3 -2 -1 0 1 2 3 4
Standard Normal Quantiles
35
40
45
50
55
60
65
70
75
80
Quantiles of Input Sample
QQ Plot of Sample Data versus Standard Normal
(a) Region A Fe
-4 -3 -2 -1 0 1 2 3 4
Standard Normal Quantiles
10
20
30
40
50
60
70
80
Quantiles of Input Sample
QQ Plot of Sample Data versus Standard Normal
(b) Region B Fe
35 40 45 50 55 60 65 70
Region A Quantiles
10
20
30
40
50
60
70
Region B Quantiles
(c) Region A Fe vs. Region B
Fe
Fig. 5 QQ-plots displaying empirical quantile-quantile plot for Regions A and B.
Samples from both mining sites are not normally distributed, and both samples do
not come from the same distribution
10 20 30 40 50 60 70
0
200
400
600
800
1000
1200
Sample Count
Region B
Region A
Fig. 6 Histogram of Fe samples from Regions A and B showing clear differences in
distributions between the two mining sites
4.3 Sulfur (S) Prediction
The results for predicting S based on the combined MWD data from Regions A and
B are shown in Fig. 11 using the GP, SVM, and RF models. These results clearly
demonstrate that the accuracy of the proposed approach generally depends on the as-
say type (for example, Fe, P, S, etc.) and distribution within the two mining sites, with
Sulfur having correlation coefficients as high as 0.91 using GP model with an RMSE
of 0.0027. On the other hand, the raw measurements of S and their GP predictions
based on MWD data are also shown in Fig. 12, which further visually demonstrates
the effectiveness of the method proposed in this paper.
16 Rami N. Khushaba et al.
(a) Region A (b) Region B
Fig. 7 Histogram of the difference between laboratory Fe and SVM estimation of Fe
on Regions A and B samples
-50 -40 -30 -20 -10 0 10 20
Difference Between Laboratory and SVM Estimation
10
20
30
40
50
60
70
Laboratory Fe Measurements
High-Grade
Waste
Low-Grade
(a) Region A
-50 -40 -30 -20 -10 0 10 20
Difference Between Laboratory and SVM Estimation
10
20
30
40
50
60
70
Laboratory Fe Measurements
High-Grade
Watse
Low-Grade
(b) Region B
Fig. 8 The difference between laboratory Fe and SVM estimation of Fe against the
classifications of Fe into low, med and high-grades, on Regions A and B samples
4.4 Spatially-Aware Cross-Validation
In order to develop a spatially-aware cross-validation scheme, we first plot the spatial
distribution of the available blast-holes across the patterns from the two mining sites
of Region A and Region B as shown in Fig. 13. Our spatial cross-validation process
involved a leave-one-blast-out process, in which all of the holes related to an individ-
ual blast are kept for testing while training is based on the data from all other holes
from all remaining blasts. The process is repeated across each of the blasts to ensure
that the testing is performed across all of the blasts in the data set. Performance esti-
mates are computed based on the final results to validate the overall assay estimation
performance across spatially different blasts.
The spatially-aware regression results for both P and S are shown in Fig.14 using
an RF model. These results demonstrate that the model’s performance when estimat-
Title Suppressed Due to Excessive Length 17
0 0.1 0.2 0.3 0.4
Algorithm
0
0.1
0.2
0.3
0.4
0.5
Laboratory
y=1.00x+0.00
r=0.79
RMSE=0.03
p<0.001
n=4584
0 0.1 0.2 0.3 0.4
Mean Algorithm & Laboratory
-0.2
-0.1
0
0.1
0.2
Laboratory - Algorithm
0.06 (+1.96SD)
0.00 [p=0.93]
-0.06 (-1.96SD)
RPC: 0.06 (41%)
CV: 23%
(a) GP model results
0 0.1 0.2 0.3 0.4
Algorithm
0
0.1
0.2
0.3
0.4
0.5
Laboratory
y=1.05x0.00
r=0.78
RMSE=0.03
p<0.001
n=4584
0 0.1 0.2 0.3 0.4
Mean Algorithm & Laboratory
-0.2
-0.1
0
0.1
0.2
Laboratory - Algorithm
0.06 (+1.96SD)
0.00 [p=0.00]
-0.06 (-1.96SD)
RPC: 0.06 (41%)
CV: 24%
(b) SVM model results
0 0.1 0.2 0.3 0.4
Algorithm
0
0.1
0.2
0.3
0.4
0.5
Laboratory
y=1.04x-0.01
r=0.81
RMSE=0.03
p<0.001
n=4584
0 0.1 0.2 0.3 0.4
Mean Algorithm & Laboratory
-0.2
-0.1
0
0.1
0.2
Laboratory - Algorithm
0.05 (+1.96SD)
0.00 [p=0.18]
-0.06 (-1.96SD)
RPC: 0.05 (39%)
CV: 22%
(c) RF model results
Fig. 9 Combined Phosphorus (P) data, Regions A and B, Bland-Altman regression
plots with several machine learning models
ing these two chemical assays using a spatially-aware mode significantly dropped
in comparison to the previously demonstrated P and S results when using random
k-fold cross-validation. This is in line with the literature findings in that, in the ran-
dom k-fold cross-validation mode, autocorrelation could have impacted the results.
Given the purpose of our first experiment was to fill in the missing gaps of chemical
18 Rami N. Khushaba et al.
-0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2 0.25 0.3
Difference Between Laboratory and RF Estimation of P
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Laboratory P Measurements
Fig. 10 Difference between laboratory measurements and RF estimation of P vs.
laboratory P measurements
assays for blast-holes, the previous performance measures are advantageous within
the context and objective of this first testing scheme. In comparison, the results of the
spatially-aware cross-validation scheme could have been impacted by several factors
including first, the limited size of the available data, with only 4584 samples across
both sites, and second, the different distributions of the assays across the two sites.
Hence, the authors conclude that further experimentation with a significantly larger
dataset would be required to properly judge the modelling accuracy of assays from
MWD across spatially distance locations.
4.5 Cross-Assay Predictions
The goal of this experiment is to verify whether the knowledge about one assay type
can better inform the models about the distributions of other assays, that is to indicate
if, for example, the knowledge of Fe or SiO2can inform Al2O3, or other such assay
combinations. For this specific test, the RF model was selected as an example of the
available regressions models, as the performance is similar across GP, SVM and RF,
and the purpose is simply to demonstrate the concept. The experiment proceeded by
first analyzing whether there is any correlation between the different assays, which
is shown to be the case in Fig. 15. The results for predicting Al2O3using MWD
alone, MWD together with Fe, and MWD together with SiO2are each shown in Fig.
16. These results show significant enhancements to Al2O3accuracy (correlation and
RMSE) when using the combination of MWD with one additional chemistry assay.
This supports the hypothesis that knowing one assay type can significantly enhance
the prediction performance for other assay types. The benefit of this finding is that
if one assays type can be measured or predicted with some sensor/model, then these
measurements can be used to improve estimates for other assays, a subject of future
investigation.
The analysis also included a spatially-aware cross-validation test on cross-assay
prediction. In this regard, the results for predicting Al2O3using MWD together with
Fe, and MWD together with SiO2are shown in Fig. 17. It is evident from these
Title Suppressed Due to Excessive Length 19
0 0.01 0.02 0.03 0.04
Algorithm
0
0.01
0.02
0.03
0.04
Laboratory
y=1.01x0.00
r=0.91
RMSE=0.002738
p<0.001
n=4584
0 0.01 0.02 0.03 0.04
Mean Algorithm & Laboratory
-0.02
-0.01
0
0.01
0.02
Laboratory - Algorithm
0.01 (+1.96SD)
0.00 [p=0.94]
-0.01 (-1.96SD)
RPC: 0.01 (56%)
CV: 27%
(a) GP model results
0 0.01 0.02 0.03 0.04
Algorithm
0
0.01
0.02
0.03
0.04
Laboratory
y=1.03x0.00
r=0.90
RMSE=0.002836
p<0.001
n=4584
0 0.01 0.02 0.03 0.04
Mean Algorithm & Laboratory
-0.02
-0.01
0
0.01
0.02
Laboratory - Algorithm
0.01 (+1.96SD)
0.00 [p=0.06]
-0.01 (-1.96SD)
RPC: 0.01 (59%)
CV: 29%
(b) SVM model results
0 0.01 0.02 0.03 0.04
Algorithm
0
0.01
0.02
0.03
0.04
Laboratory
y=1.03x0.00
r=0.92
RMSE=0.002604
p<0.001
n=4584
0 0.01 0.02 0.03 0.04
Mean Algorithm & Laboratory
-0.02
-0.01
0
0.01
0.02
Laboratory - Algorithm
0.01 (+1.96SD)
0.00 [p=0.06]
-0.01 (-1.96SD)
RPC: 0.01 (53%)
CV: 26%
(c) RF model results
Fig. 11 Combined Sulfur (S) data, Regions A and B, Bland-Altman regression plots
with several machine learning models
results that the correlation coefficients dropped in comparison to those obtained by
the same models with a random cross-validation scheme, which indicates the impact
of autocorrelation in this case. However, these results are also encouraging as the
spatially-aware models’ performance can be largely enhanced by knowing the distri-
bution of other assays estimates. For example, the same approach can be used with
20 Rami N. Khushaba et al.
0 500 1000 1500 2000 2500 3000 3500 4000 4500
Samples
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
Sulfur (S) Readings
GP - Sulfur (S)
Laboratory - Sulfur (S)
Fig. 12 Raw Sulfur (S) laboratory and GP model predictions showing high agreement
between measurements and predictions
(a) Region A blast-hole distribution (b) Region B blast-hole distribution
Fig. 13 Spatial distribution of the blast-holes across the two sites, with colors rep-
resenting the different blasts across the two sites. The x-y coordinates have been
re-scaled due to commercial sensitivity
a technology relying on PFTNA that estimates most assays apart from P, in order
to estimate these missing P values. Combining MWD with a selection of PFTNA-
estimated assays to fill in the missing chemistry values warrants further investigation.
4.6 Multivariate vs. Univariate Response Predictors
While the main aim of this paper was to prove the concept of assay and material pre-
diction via machine learning models, it could be argued that using machine learning
approaches to predict multivariate response values are now available (for example,
multivariate random forests) (Segal and Xiao 2011), and thus the univariate predic-
tion of assay grades is a limited step forward. To validate the effectiveness of the
multivariate RF response predictor against the traditional univariate version, a com-
parison between both models was performed to predict all assays types. The results
in Fig. 18 show no significant differences in terms of the correlation coefficients be-
tween the two multivariate and univariate RF models, except for Mn, which could be
Title Suppressed Due to Excessive Length 21
0 0.1 0.2 0.3 0.4 0.5
Algorithm
0
0.1
0.2
0.3
0.4
0.5
Laboratory
y=1.00x+0.00
r=0.67
RMSE=0.04
p<0.001
n=4584
0 0.1 0.2 0.3 0.4
Mean Algorithm & Laboratory
-0.2
-0.1
0
0.1
0.2
Laboratory - Algorithm
0.07 (+1.96SD)
0.00 [p=0.00]
-0.07 (-1.96SD)
RPC: 0.07 (48%)
CV: 28%
(a) Phosphorus (P) regression results
0 0.01 0.02 0.03 0.04
Algorithm
0
0.01
0.02
0.03
0.04
Laboratory
y=0.94x+0.00
r=0.84
RMSE=0.00
p<0.001
n=4584
0 0.01 0.02 0.03 0.04
Mean Algorithm & Laboratory
-0.02
-0.01
0
0.01
0.02
Laboratory - Algorithm
0.01 (+1.96SD)
0.00 [p=0.00]
-0.01 (-1.96SD)
RPC: 0.01 (66%)
CV: 35%
(b) Sulfur (S) regression results
Fig. 14 Spatially-aware regression results for Phosphorus (P) and Sulfur (S) using
individual RF models
due to the limited data size and the distribution of the data from the two sites. How-
ever, it is important to note that the multivariate version predicted all responses in one
step. In a real-time system deployment, this may translate to a significant reduction in
the computational requirements associated with building several individual chemistry
assay predictors. For offline predictions, it appears both multivariate and univariate
approaches are suitable.
4.7 Material Type Classification
In this section, an experiment employing SVM as a classifier to predict the pres-
ence or absence of different material-types was conducted. The purpose is to estimate
whether or not a given percentage of the material under consideration is present at the
specified location. This is a process for which the threshold utilized to define the ex-
istence/absence of a certain material can vary from one material type to another, as
different materials may become operationally relevant at different proportions. For
the purpose of this paper, as we are primarily interested in a proof of concept, a
22 Rami N. Khushaba et al.
Fig. 15 Pearson correlation coefficients among all assays from both mining sites
threshold of 0% was selected; any value greater than zero is taken to indicate the
presence of the material under investigation. The SVM model is well suited to this
analysis as SVM models were originally developed for binary classification problems
such as this one.
Eight material-types were randomly selected from all existing material types, in
no specific order, to demonstrate the feasibility of this concept. The results for this
section are shown in Fig. 19 for the following material types: i) shale (SHL), ii)
banded iron formation (BIF), iii) powdery banded iron formation (BPO), iv) goethite
ochreous (GOL), v) hematite goethite medium (HGM), vi) hematite goethite friable
(HGF), vii) shale ferruginous (SHF), and viii) goethite martite vitreous goethite ma-
trix (GMO). Given the available number of samples, these results demonstrate that
MWD can be used to predict the presence/absence of different material-types with
accuracies in the order of 80–90% across the selected material types.
4.8 Feature Importance and Selection
The importance of the various features used for assay prediction is compared in this
section, to provide insights into the different features’ relevance for this application.
This is aided by algorithms such as RF providing a measure of predictor importance
for process discovery analysis.
In this case, the analysis showed that the top ten features that were highly ranked
by RF for Fe predictions in Region A included i) the maximum values of torque, rota-
tionPressure, and pressure ratio, ii) the waveform length of rotationPressure and pres-
sure ratio, iii) the standard deviation of pressure ratio, iv) Hjorth activity of pressure
ratio, torque, and sed, and v) median of airPressure. For predicting Fe in Region B,
the following features were ranked as top ten by the utilized RF model, including i)
the flatness of sed and rop, ii) Hjorth activity of sed, iii) Hjorth mobility and com-
plexity of rop, iv) standard deviation of sed and rop, v) median of airPressure and
rotationPressure, and vi) kurtosis of rop. For Sulfur prediction across the combined
data from the two regions, the top ten features included all of i) the median of rop,
Title Suppressed Due to Excessive Length 23
0 5 10 15
Algorithm
0
5
10
15
Laboratory
y=1.15x-0.61
r=0.65
RMSE=1.6
p < 0.001
n=4584
0 5 10 15
Mean Algorithm & Laboratory
-5
0
5
Laboratory - Algorithm
3.1 (+1.96SD)
-0.07 [p=0.00]
-3.2 (-1.96SD)
RPC: 3.1 (77%)
CV: 45%
(a) RF model results predicting Al2O3using MWD only
0 5 10 15
Algorithm
0
5
10
15
Laboratory
y=1.17x-0.62
r=0.90
RMSE=0.92
p<0.001
n=4584
0 5 10 15
Mean Algorithm & Laboratory
-5
0
5
Laboratory - Algorithm
1.9 (+1.96SD)
-0.01 [p=0.47]
-1.9 (-1.96SD)
RPC: 1.9 (47%)
CV: 27%
(b) RF model results predicting Al2O3using MWD with Fe
0 5 10 15
Algorithm
0
5
10
15
Laboratory
y=1.15x-0.58
r=0.85
RMSE=1.1
p<0.001
n=4584
0 5 10 15
Mean Algorithm & Laboratory
-5
0
5
Laboratory - Algorithm
2.2 (+1.96SD)
-0.03 [p=0.05]
-2.2 (-1.96SD)
RPC: 2.2 (53%)
CV: 32%
(c) RF model results predicting Al2O3using MWD with SiO2.
Fig. 16 The impact of combining MWD with other individual chemistry assays to
predict Al2O3
airPressure, torque and sed, ii) Hjorth activity of airPressure, rop, torque and sed, and
iii) geometric mean of rop, airPressure. On the other hand, for predicting phosphorus
based on the combined data from two regions, the utilized RF model ranked the top
features to include i) geometric mean of rop and airPressure, ii) median of rop air-
24 Rami N. Khushaba et al.
0 5 10 15
Algorithm
0
5
10
15
Laboratory
y=1.16x-0.67
r=0.79
RMSE=1.3
p<0.001
n=4584
0 5 10 15
Mean Algorithm & Laboratory
-5
0
5
Laboratory - Algorithm
2.5 (+1.96SD)
-0.1 [p=0.00]
-2.7 (-1.96SD)
RPC: 2.6 (65%)
CV: 36%
(a) RF model results predicting Al2O3using MWD with Fe from a spatially-aware cross-validation
scheme
0 5 10 15
Algorithm
0
5
10
15
Laboratory
y=1.12x-0.60
r=0.74
RMSE=1.4
p<0.001
n=4584
0 5 10 15
Mean Algorithm & Laboratory
-5
0
5
Laboratory - Algorithm
2.6 (+1.96SD)
-0.17 [p=0.00]
-3.0 (-1.96SD)
RPC: 2.8 (69%)
CV: 39%
(b) RF model results predicting Al2O3using MWD with SiO2from a spatially-aware cross-validation
scheme
Fig. 17 Spatially-aware cross-validation test on cross-assay prediction
Pressure and rotationRPM, iii) Hjorth activity of airPressure, rop, and sed, iv) Hjorth
complexity of sed, and v) the integral sum of rotationRPM.
As can be seen from the above list of features, the ranking and the type of top
ten contributing features to each assay estimation accuracy would be different from
one chemical assay to another and could be even different from one region to another
(depending on the distribution of the data and the amount of available data). In the
context of the available data within this study, Hjorth parameters (specifically the
activity), waveform length, flatness and basic statistics were almost always among the
top-ranked features. This does not exclude the importance of the remaining features
but is simply intended to highlight the common key features. As the primary goal of
this paper was to prove the feasibility of predicting assays and materials from MWD
data, a more thorough analysis of the feature importance and ranking will be included
in future work on a larger dataset.
Title Suppressed Due to Excessive Length 25
Al2O3 CaO Fe LOI MgO Mn P S SiO2 TiO2
0.4
0.5
0.6
0.7
0.8
0.9
1
Pearson Correlation Coefficient
multivariate-RF
traditional-RF
Fig. 18 Multivariate vs. univariate RF response predictors
5 Discussion
The results presented in this paper indicate that the approach of mapping MWD to
materials types and chemical assay values is feasible, leading to the conclusion that
machine learning algorithms can be employed to predict materials and assays. While
this was only demonstrated on eight materials types, the findings do generalize across
a broader set of materials (excluded for brevity). It is important to mention here that
our findings also suggest that the accuracy of this mapping process is primarily de-
pendent on (i) assay/material type and its distribution across the mining sites from
which the data was acquired and (ii) the amount of available data. It was also found
that the choice of the regression model, among GP, SVM and RF models, did not
significantly alter the findings, as the quality of the estimation performed by these
machine learning models mainly depended on the quality of the extracted features.
The analysis started with Fe estimation using GP, SVM, and RF models, with
results suggesting that Fe estimation across Region A was more accurate than that
across Region B. The distribution of Fe across the two mining sites was then exam-
ined in Fig. 6 and it was found that the two mining sites had significantly different
distributions of iron as the histogram of Fe across Region B spiked much higher than
that across Region A (larger kurtosis value). This in turn explains part of the variabil-
ity of the results from both sites. Further variability appears to come from limitations
on the available data, with approximately 2000 samples available for each site re-
sulting in fewer samples within some classes (waste, low-grade, high-grade). It is
important to note that the largest Fe prediction errors were actually made within the
waste class, as was shown in Fig. 8, while much more consistent predictions were
made along low-grade and high-grade iron where the predicted values would actually
be relevant.
In terms of Phosphorus (P), the largest prediction errors were made along values
larger than 0.35, as when combining data from Regions A and B it was found that out
of the set of nearly 4000 samples, only around 11 Phosphorus samples were avail-
26 Rami N. Khushaba et al.
1 2
Predicted Class
1
2
True Class
6.2%
10.0%90.0%
93.8%
(a) SHL confusion matrix
1 2
Predicted Class
1
2
True Class
13.0%
9.8%90.2%
87.0%
(b) BIF confusion matrix
1 2
Predicted Class
1
2
True Class
19.7%
15.7%84.3%
80.3%
(c) BPO confusion matrix
1 2
Predicted Class
1
2
True Class
13.9%
22.7%77.3%
86.1%
(d) GOL confusion matrix
1 2
Predicted Class
1
2
True Class
14.0%
18.2%
81.8%
86.0%
(e) HGM confusion matrix
1 2
Predicted Class
1
2
True Class
12.4%
17.3%
82.7%
87.6%
(f) HGF confusion matrix
1 2
Predicted Class
1
2
True Class
5.5%
16.3%83.7%
94.5%
(g) SHF confusion matrix
1 2
Predicted Class
1
2
True Class
18.9%
10.9%
89.1%
81.1%
(h) GMO confusion matrix
Fig. 19 Classification confusion matrices to inform us about the accuracy with which
one can identify the existence of different material-types: 1 - Material does not exist,
2 - Material exist
able with values larger than 0.35. Hence, there was not enough data for the GP, SVM
and RF models to learn the relationship between MWD and high Phosphorus val-
ues. However, the overall prediction accuracy of Phosphorus suggests its estimation
is useful, with a Pearson correlation coefficient of 0.81, RMSE of 0.3, and a p-value
of <0.001. This indicates significant correlations between laboratory measurements
Title Suppressed Due to Excessive Length 27
and estimations by our models. In terms of Sulfur (S), the predictions shown in this
paper were found to be most accurate when combining the data from the two min-
ing sites, with the best results achieved by RF with a correlation coefficient of 0.92,
RMSE of 0.002604, and a p-value of <0.001.
The effectiveness of using knowledge about one assay type to assist the prediction
of other assays was investigated. The hypothesis, in this case, is that if one algorithm
or sensor can predict one assay type accurately, then one can make use of this knowl-
edge to better predict other assays. For this part of the analysis, when using the MWD
features with an RF model, the combined Regions A and B data had a correlation co-
efficient of 0.65 with an RMSE of 1.6 when predicting Al2O3. However, these results
were significantly enhanced by adding either Fe (r=0.9, RMSE =0.92) or SiO2
(r=0.85, RMSE =1.1), as was shown in Fig. 16. The reason for selecting these two
assays to augment MWD data when predicting Al2O3is that high correlations were
observed between these assays, as depicted in Fig. 15. Hence, the knowledge of either
of these assays supported the model to more accurately estimate the Al2O3values.
In the final part of the experiments, we presented a proof of concept that MWD
can also predict the presence or absence of material types. Eight different materials
types including i) SHL, ii) BIF, iii) BPO, iv) GOL, v) HGM, vi) HGF, vii) SHF
and viii) GMO were randomly selected to demonstrate this. The confusion matrices
generated when using the SVM as a classifier to predict material presence or absence
showed, in general, accuracies over 80% achieved with the available data.
6 Conclusion
In this paper, a proof of concept has been presented for using machine learning and
MWD data to predict the presence or absence of material types and estimate chem-
ical assay values. It is important to note that while different applications of MWD
have been previously considered in the literature, the study presented in this paper is
the first to use MWD for material logging and assaying purposes, to the best of the
authors’ knowledge. The findings of this study strongly support the feasibility of the
proposed approach, with results showing correlations between MWD features and
assays types of up to 0.92 for individual assays. The analysis has also shown that the
accuracy depends on the distribution of the data and the assay type being predicted. It
was also demonstrated that the presence or absence of material types can be predicted
while drilling by using MWD data based on a limited dataset.
The findings of this paper are important to the mining industry as the timeliness
and quality of these estimations cascade through the downstream mining processes.
Knowledge of material-types and chemical assays can play a significant role in min-
ing, guiding the drilling process, orebody modeling, and providing chemistry data
with down-hole resolution. The generated predictions and estimates of material types
and assays can also further help guide mine planning. The work in this field continues
as the authors plan to investigate the impact of using deep learning models on much
larger datasets from a wider range of sites. Further studies will be conducted on auto-
matic feature extraction in comparison to the handcrafted feature extraction approach
28 Rami N. Khushaba et al.
utilized in this paper, which demonstrated the feasibility of using this MWD data for
logging and assaying.
Acknowledgements This work has been supported by the Australian Centre for Field Robotics and the
Rio Tinto Centre for Mine Automation. The authors would also like to acknowledge the support of Anna
Chlingaryan and Katherine Silversides in the manuscript review and editing process.
References
Box J, Phillips J, Clout J (2002) Use of geological material types for predicting iron ore product charac-
teristics. FUWA – Ward Iron and Steel Making Workshop
Cawley G, Talbot N (2010) On over-fitting in model selection and subsequent selection bias in performance
evaluation. Journal of Machine Learning Research 11:2079–2107
Clark I, Dominy S (2017) Underground bulk sampling, uniform conditioning and conditional simulation -
unrealistic expectations? Eighth World Conference on Sampling and Blending :61
Dubnov S (2004) Generalization of spectral flatness measure for non-gaussian linear processes. Signal
Processing Letters, IEEE 11:698 – 701
Galende M, Men´endez M, Fuente M, Sainz-Palmero G (2018) Monitor-while-drilling-based estimation of
rock mass rating with computational intelligence: The case of tunnel excavation front. Automation in
Construction 93:325–338
Hatherly P, Leung R, Scheding S, Robinson D (2015) Drill monitoring results reveal geological conditions
in blasthole drilling. International Journal of Rock Mechanics and Mining Sciences 78:144–154
Hjorth B (1970) Eeg analysis based on time domain properties. Electroencephalography and Clinical Neu-
rophysiology 29(3):306 – 310, ISSN 0013-4694
Hudgins B, Parker P, Scott R (1993) A new strategy for multifunction myoelectric control. IEEE transac-
tions on bio-medical engineering 40:82–94
Jeanneau P, Flahaut V, Maddever R (2017) Iron ore benefits from neutron pulsed geochemical tools. Con-
ference proceedings: IRON ORE 2017, Perth, Australia) 2017:387–396
Johnston J D (1988) Transform coding of audio signals using perceptual noise criteria. IEEE Journal on
Selected Areas in Communications 6(2):314–323
Khanal M, Qin J, Shen B, Dlamini B (2020) Preliminary investigation into measurement while drilling as
a means to characterize the coalmine roof. Resources 9:10
Khorzoughi M, Hall R, Apel D (2018) Rock fracture density characterization using measurement while
drilling (mwd) techniques. International Journal of Mining Science and Technology 28:504–516
Khushaba R N, Takruri M, Miro J V, Kodagoda S (2014) Towards limb position invariant myoelectric
pattern recognition using time-dependent spectral features. Neural Networks 55:42 – 58, ISSN 0893-
6080
Leung R, Scheding S (2015) Automated coal seam detection using a modulated specific energy measure
in a monitor-while-drilling context. International Journal of Rock Mechanics and Mining Sciences
75:196–209, ISSN 1365-1609
L’Heureux A, Grolinger K, Elyamany H F, Capretz M A M (2017) Machine learning with big data: Chal-
lenges and approaches. IEEE Access 5:7776–7797
Li S y, Yang M, Li C c, Cai P (2008) Analysis of heart rate variability based on singular value decomposi-
tion entropy. Journal of Shanghai University (English Edition) 12:433–437
Market J, Byrne C, Robinson D, Jeanneau P, Rossiter H (2019) Downhole assays in the pilbara. Conference
proceedings: IRON ORE 2019, Perth, Australia) 2019:366–386
McHugh C, Stokes A, Oppolzer F, Rogers B (2012) Automated multi-drills in rio tinto. Conference pro-
ceedings: Proceedings Eighth Open Pit Operators’ Conference , (The Australasian Institute of Mining
and Metallurgy: Melbourne) 2012:83–88
Meyer H, Reudenbach C, Hengl T, Katurji M, Nauss T (2018) Improving performance of spatio-temporal
machine learning models using forward feature selection and target-oriented validation. Environmental
Modelling and Software 101:1–9, ISSN 1364-8152
Meyer H, Reudenbach C, W¨ollauer S, Nauss T (2019) Importance of spatial predictor variable selection
in machine learning applications – moving from data reproduction to spatial prediction. Ecological
Modelling 411:108815, ISSN 0304-3800
Title Suppressed Due to Excessive Length 29
Navarro J, Sanchidri´an J, Segarra P, Castedo R, Costamagna E, L´opez L (2018) Detection of potential
overbreak zones in tunnel blasting from mwd data. Tunnelling and Underground Space Technology
82:504–516
Paine M, Boyle C, Lewan A, Phuak E, Mackenzie P, Ryan E (2016) Geometallurgy at rtio – a new angle
on an old concept. Third AusIMM International Geometallurgy Conference :55–61
Park J, Kim K (2020) Use of drilling performance to improve rock-breakage efficiencies: A part of m ine-to-
mill optimization studies in a hard-rock mine. International Journal of Mining Science and Technology
30(2):179–188, ISSN 2095-2686
Phinyomark A, Hirunviriya S, Limsakul C, Phukpattaranont P (2010) Evaluation of emg feature extraction
for hand movement recognition based on euclidean distance and standard deviation. International Con-
ference on Electrical Engineering/Electronics Computer Telecommunications and Information Tech-
nology (ECTI-CON), Chiang Mai 2010:856 – 860
Rai P, Schunesson H, Lindqvist P A, Kumar U (2015) An overview on measurement-while-drilling tech-
nique and its scope in excavation industry. Journal of The Institution of Engineers (India): Series D
96:57–66
Rasmussen C E, Williams C K I (2005) Gaussian Processes for Machine Learning (Adaptive Computation
and Machine Learning). The MIT Press, ISBN 026218253X
Roberts D R, Bahn V, Ciuti S, Boyce M S, Elith J, Guillera-Arroita G, Hauenstein S, Lahoz-Monfort
J J, Schr¨oder B, Thuiller W, Warton D I, Wintle B A, Hartig F, Dormann C F (2017) Cross-validation
strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography 40(8):913–
929
Segal M, Xiao Y (2011) Multivariate random forests. WIREs Data Mining and Knowledge Discovery
1(1):80–87
Silva C (2005) Vibration and Shock Handbook
Silversides K L, Melkumyan A (2020) Boundary identification and surface updates using mwd. Mathe-
matical Geosciences :1–25
Smith C, Jeanneau P, Maddever R, Fraser S, Rojc A, Lofgren M, Flahau V (2015) PFTNA logging tools
and their contributions to in-situ elemental analysis of mineral boreholes. TOS forum 5:157 – 165
Sommerville B, Boyle C, Brajkovich N, Savory P, Latscha A A (2014) Mineral resource estimation of the
brockman 4 iron ore deposit in the pilbara region. Applied Earth Science 123(2):135–145
Stone M (1974) Cross-validatory choice and assessment of statistical predictions. Journal of the Royal
Statistical Society: Series B (Methodological) 36(2):111–133
Stone M (1977) An asymptotic equivalence of choice of model by cross-validation and akaike’s criterion.
Journal of the Royal Statistical Society Series B (Methodological) 39(1):44–47, ISSN 00359246
Talebi H, Peeters L, Mueller U, Tolosana-Delgado R, Van den Boogaart K (2020) Towards geostatistical
learning for the geosciences: A case study in improving the spatial awareness of spectral clustering.
Mathematical Geosciences 52
Theodoridis S, Koutroumbas K (2009) Pattern Recognition. ISBN 978-1-59749-272-0
Wedge D, Hartley O, McMickan A, Green T, Holden E J (2019) Machine learning assisted geological inter-
pretation of drillhole data: Examples from the pilbara region, western australia. Ore Geology Reviews
114:103118, ISSN 0169-1368
Wedge D, Lewan A, Paine M, Holden E J, Green T (2018) A data mining approach to validating drill hole
logging data in pilbara iron ore exploration. Economic Geology 113:961–972
Zhou H, Monteiro S, Hatherly P, Ramos F, Nettleton E, Oppolzer F (2010) Automated rock recognition
with wavelet feature space projection and gaussian process classification. Proceedings - IEEE Interna-
tional Conference on Robotics and Automation :4444–4450
