See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/332143100

Evaluation of sensor technologies for on-line raw material characterization in

"Reiche Zeche" underground mine - outcomes of RTM implementation

Conference Paper · April 2019

CITATION READS

1 908

2 authors:

Feven Desta Mike Buxton

Delft University of Technology Delft University of Technology
10 PUBLICATIONS   51 CITATIONS    43 PUBLICATIONS   723 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Real-Time Mining View project

INCO-Piles 2020: International consortium to recover CRMs from stockpiles/tailings targeting RIS View project

All content following this page was uploaded by Feven Desta on 02 April 2019.

The user has requested enhancement of the downloaded file.

REAL TIME MINING - Conference on Innovation on Raw Material Extraction Freiberg 2019

Evaluation of sensor technologies for on-line

raw material characterization in “Reiche Zeche” underground mine -
outcomes of RTM implementation

F. S. Desta, M. W. N. Buxton

Resource Engineering, Faculty of Civil Engineering and Geosciences, Delft University of Techno-
logy, P.O. Box 5048, 2600 GA Delft, The Netherlands

ABSTRACT:

The increasing advances in sensor technology have resulted in greater availability of sen-
sor data for a wide range of applications. One such application is raw material characteri-
zation in mining operations. Sensor technologies operate over certain range of the elec-
tromagnetic spectrum and provide information on several aspects of material properties.
The sensitivity and the material properties the instrument detects and measures varies
from sensor to sensor. The purpose of this study was to synthesize and evaluate the use
of sensor technologies for characterization of a polymetallic sulphide deposit in “Reiche
Zeche” underground mine. This paper discusses the material characterization methodolo-
gy using sensor technologies, demonstrates how it fits within the Real-Time Mining (RTM)
framework, identifies the interface for both software and hardware requirements and de-
fines the gaps and limitations of application of sensors. It provides a brief overview of the
use of sensor and data fusion for material characterization to convey a high-level context
in raw material characterization. The sensor technologies considered in this study include
RGB imaging, visible–near infrared (VNIR), short wave infrared (SWIR), mid-wave infrared
(MWIR), long-wave infrared (LWIR) and Raman spectroscopy.

The required information from sensor data in mining operations is not limited to grade con-
trol applications. Information on co-occurring minerals or elements are also important for
definition of requirements in mineral processing, to identify indirect proxies of ele-
ments/minerals of interest, to understand the formation of minerals, to define requirements
for blasting parameters, to improve safety and to define requirements for environmental
monitoring of toxic material. In view of these points, there is a need for combinations of
sensors to achieve a near complete description of material composition and properties.
The methodological approaches developed for information extraction from each sensor
data and fused data are presented. This includes both direct mineral fingerprinting and
indirect proxies using spectral data. The efficient sensor data processing methods and the
acquired results from the use of individual sensor and the fused data are summarized.
Overall, the acquired results from the use of each sensor technology and the data fusion
approach significantly contributed to an improvement of data quality and illustrate the effi-

32
REAL TIME MINING - Conference on Innovation on Raw Material Extraction Freiberg 2019

ciency of use of sensors in the mining industry. However, some of the observed limitations
include lack of system robustness, a need for test case specific mineral libraries, the need
for development of an integrated principled tool for efficient data collection, processing and
knowledge generation. Going forward, automated material characterization is possible with
robust system design (exemplified by portable and ruggedized system) and efficient soft-
ware (test case specific mineral libraries) that can be developed using a combined sensor
signal.
Keywords—sensor technologies, data fusion, material characterization, on-line data, min-
ing

1 Introduction

Sensor technologies that produce high-throughput multi and megavariate data are advancing. Sen-
sors derived data are in current use in a wide range of applications. Raw material characterization in
mining operations is one application area. Sensor technologies measure different aspects of material
properties. Material property is a broad term that includes physical, chemical, optical, mechanical
and atomic properties. Fundamental understanding of material characteristics is crucial in selecting
suitable sensor solutions for operational decision making using raw material characterization. In
addition, the selection of sensors for a specific application requires knowledge of sensor parameters.
These parameters include operating wavelength range, spatial resolution, spectral resolution, ac-
curacy, precision, sensors field of view (spot size), robustness for environmental conditions (such as
vibration, humidity and dust), detection limit and depth of penetration (e.g surface or volumetric
measurements).
The operating wavelength range of a sensor is the window of the electromagnetic spectrum on
which the given sensor operates. Spatial resolution specifies the pixel size of an image that provide
details or the smallest addressable element the image holds (the distinct detail in the image). Spect-
ral resolution is a measure of sensor ability to resolve spectral features and bands into separate
components (width of spectral band). Finer spectral resolutions enable the higher resolution spectral
characteristics of the targets to be captured by the sensor. Accuracy is a measure of the closeness of
a result to the true or known standard value. Precision refers to the reproducibility of multiple mea-
surements. Sensor field of view or spot size refers to the size of the measured area of a single mea-
surement. Robustness of sensor systems for harsh environmental conditions (e.g vibration, humidity
and dust) is important for in-situ applications (e.g for underground applications). Detection limit of
a sensor is the lowest quantity of a substance that can be detected by the system with a general con-
fidence level of 99%. It is a key parameter for application of sensors in low-grade mines. Depth of
penetration is the depth light or electromagnetic radiation can penetrate into a material.

The field of view or the extent of observable scene of sensors differs from technology to technolo-
gy. Thus, some technologies produce point data, while the others produce image data. For example,
the imaging technologies cover a larger area of a target and provide a 2D image that shows the spa-
tial and spectral distribution of entities under investigation. Whereas, the point techniques cover a
very small area (spot size) and generate point data, discrete unit of information that is acquired from
a single spot (though the spot size varies from instrument to instrument). Depending on the techno-

33
REAL TIME MINING - Conference on Innovation on Raw Material Extraction Freiberg 2019

logy, the point data are spectra which consist of wavelength and reflectance or intensity informati-
on. Based on the output data type, sensor technologies are divided into point spectroscopies and
imaging spectroscopies.

Crucial information from sensor derived data are applications dependent. For example, the key ma-
terial properties (geological attributes) in mining operations include mineralogy, geochemistry,
fragmentation and ore geometry. Knowledge of these properties plays a key role in supporting
effective decision making in mining operations. Thus, improves the economic and environmental
benefits. This paper presents the use of RGB imaging, visible–near infrared (VNIR), short wave
infrared (SWIR), mid-wave infrared (MWIR), long-wave infrared (LWIR) and Raman technologies
for polymetallic sulphide ore characterization, highlights the developed methodological approaches
for knowledge generation, addresses the opportunities with sensor combinations and defines the
gaps and limitations for future research works. The sensor technologies are described in Table 1.

Table 1: Sensor technologies operating wavelength range, the material properties the systems
measure and the geological attributes that can be derived from the sensors signal
Operating
Material Properties/ Geological attri-
No. Sensor Wavelength
Type of energy transfer butes
range (µm)
Mineralogical
1 RGB Imaging 0,4 - 0,7 Reflection
Texture
(Fragmentation)
Hyperspectral Imaging Reflection\Absorption
2 0,4 - 1,0 Mineralogical
(VNIR)
Hyperspectral Imaging Reflection\Absorption
3 1,0 - 2,5 Mineralogical
(SWIR)
Mid Wave Infrared Reflection\Absorption
4 2,5 - 7,0 Mineralogical
(MWIR)
Long Wave Infrared Reflection\ Absorption\
5 7,0 -15 Mineralogical
(LWIR) Emission
Scattering of radiation
6 RAMAN 0,2 -1,4 Mineralogical

2 Study sites

The Reiche Zeche underground mine located in the Freiberg district, eastern Erzgebirge, Germany,
served as the case study area. The deposit is characterized by polymetallic vein type mineralization
formed by two hydrothermal mineralization events of Late-Variscan and Post-Variscan age (Seifert
2008). The Late-Variscan mineralization event dominates in the central part of the mine and mine-
ralization is rich in sulphur, iron, lead, zinc and copper. Typical ore minerals include pyrite, galena,
arsenopyrite, chalcopyrite and sphalerite, along with quartz and minor carbonate gangue. Ore mine-
rals with a smaller Cu, Zn and Fe content characterize the Post-Variscan mineralization event. This
mineralization event consists of a fluorite-barite-lead ore assemblage, mainly containing sphalerite,
pyrite, galena, chalcopyrite and marcasite, as well as quartz, fluorite, carbonates and barite, as
gangue (Seifert 2008; Benkert et al. 2015).

34
REAL TIME MINING - Conference on Innovation on Raw Material Extraction Freiberg 2019

2.1 The nature of materials

Measurements of rock attributes were performed both in-situ and using rock samples collected from
mine face, drill core, muck pile and LHD potential sensor installation sites. Figure 1 shows sensor
technologies used at each site. The samples were collected in the form of channel samples (from the
mine face), rock chips samples (from muck pile and LHD) and drill core samples. The in-situ mea-
surements were performed in the underground mine using an RGB imager.

Drill core Muck pile LHD

Mine face
• Hyperspectral • RGB Imaging • RGB
• RGB • MWIR
• MWIR • MWIR
• Hyperspectral • LWIR
• LWIR • LWIR
• MWIR • Hyperspectral
• Raman • Hyperspectral
• LWIR • Raman
• Raman
• Raman

Figure 1: Potential sensor solutions for characterization of materials at mine face, drill core, muck
pile and LHD potential sensor installation sites in the mining value chain

3 Sensor technologies and test measurements

Out of the selected sensor technologies, the point technologies are MWIR and LWIR. The imaging
techniques are RGB imaging and VNIR/SWIR hyperspectral imaging. In the sections that follow,
the measurement results pertaining the point and imaging techniques are presented. The conventio-
nal techniques namely X-ray diffraction (XRD), X-ray fluorescence (XRF), Electron Microprobe
Analyser (EMPA) and Inductively Coupled Plasma Optical Emission Spectrometry /Mass Spectro-
metry (ICP-ES/MS) data were used to validate the material characterization results.

3.1 Points spectrometers

3.1.1 RAMAN

Raman spectroscopy is a well-established technique that provides mineralogical information. It can

be used for the identification of a wide range of minerals such as iron ore oxides, carbonates, silica-
te, sulphides and sulphate (Gaft et al., 2005; Griffith., 1975; White., 1975; Mernagh and Trudu,
1992). To assess the usability of the technology, two Raman spectrometers with excitation laser
sources of 532nm and 785nm were considered. Measurements were performed using powder and
rock samples. The acquired raw spectra were pre-processed and interpreted using RRUFF mineral
library and other published works..
The 785nm laser Raman measurements resulted in defined Raman peaks for both rock and powder
sample forms. However, the results obtained from the analysis of the rock samples were superior

35
REAL TIME MINING - Conference on Innovation on Raw Material Extraction Freiberg 2019

than the powder samples measurements. Similarly, the 532nm laser Raman measurements resulted
in a better signal for the measurements of rock samples than powder samples. Comparing the two
excitation laser sources for the characterization of the test case materials, the 785nm laser source
outperforms the 532nm laser source. This is likely due to the fact that longer excitation wavelengths
are known to give less fluorescence than shorter excitation wavelength (Bumbrah and Sharma,
2016). The other possible reason could be, for non-transparent samples (e.g sulphide minerals) lon-
ger wavelength excitation laser sources penetrate deeper into the samples, thus provide better signal
than the shorter wavelengths (Tuschel., 2016).

The Raman method provided good results for the identification of most of the test case minerals.
Minerals that were identified using the 785nm laser source Raman system include; calcite, sphaleri-
te, kaolinite, marcasite, pyrite and siderite. For example, Figure 2 shows sample sphalerite Raman
spectra. The analysis of Raman applicability for the characterization of the test case materials was
extended into usability assessment of ore-waste discrimination and elemental concentration predic-
tion using the chemical fingerprints of the minerals. However, based on the analytical measure-
ments of 40 sample, the correlation of the Raman signal to the elemental content of the material
under investigation is very low therefore the prediction accuracy of the model is low. Similarly, ore-
waste discrimination using Raman signal was not possible. The correlation of the Raman signal
with the elemental and ore-waste discrimination was tested using both linear and non-linear techni-
ques. However, neither linear nor non-linear relations could be achieved from the Raman signals.
Therefore, this technique was not further considered for data fusion.

Figure 2: Raman spectrum of Sphalerite

The main challenges of Raman spectra analysis include peak overlap and fluorescence effect. The
former can be minimized by considering spectra decomposition techniques such as Multivariate
Curve Resolution-Alternating Least Squares (MCR-ALS). The latter can be minimized by conside-
ring longer excitation wavelengths. Raman has a good potential for quantitative analysis of the
identified minerals, however real-time application requires deposit specific mineral library that
takes into account materials heterogeneity. The current advancement of the technology resulted in a
hand-held instrumentation permitting in-situ measurements.

3.1.2 MWIR and LWIR reflectance

MWIR and LWIR reflectance data were analysed for ore-waste discrimination using chemometric
analytical techniques. Design of Experiment (DoE) was implemented to identify the optimal inde-

36
REAL TIME MINING - Conference on Innovation on Raw Material Extraction Freiberg 2019

pendent and combined data filtering techniques for discriminating the two classes using the MWIR
and LWIR datasets. The processed data were used to make predictions about the composition of
unknown samples. A series of prediction models were developed using the processed data combi-
ned with Partial least squares-Discriminant Analysis (PLS-DA). Model performance was evaluated
using the calibration, validation and prediction statistics in the form of an estimated prediction error.
When models were applied to the MWIR dataset, the prediction improved to 86.3% after baseline
correction. After normalization of the LWIR data, an enhanced correct classification rate of 84.7%
was obtained. The MWIR data alone provide sufficient information to successfully classify the
samples into ore and waste.

This finding is of interest since this region of the electromagnetic spectrum is the least explored due
to limited instrument development. The two techniques were successfully used to discriminate ore
and waste materials, the reflectance signals of the two techniques combined with PLS-DA has a
great potential for rapid automated online discrimination of ore and waste materials. The details of
the methodological approach is described in (Desta and Buxton, 2018). In addition to ore-waste
discrimination, the use of MWIR for Fe and a combined Pb Zn predictions was assessed. The ac-
quired prediction accuracies are 85% and 86.7% respectively. LWIR is very well known for analy-
sis of rock forming minerals, however using chemometric techniques the use of the technology for
elemental prediction was assessed. The Fe prediction accuracy of LWIR spectral reflectance data
combined with chemometric techniques was 88%. Likewise, a prediction accuracy of 73% was
achieved for prediction of a combined Pb Zn concentration.

In the test case, sulphide minerals are the primary sources of the target elements (e.g Pb, Zn and
Fe). However, identification of the sulphide minerals with direct mineral fingerprinting of MWIR
and LWIR reflectance data is challenging, due to the weak features of the minerals in the spectra.
However, data-divine approach can be used to extract knowledge from the multivariate reflectance
spectra data of MWIR and LWIR techniques. For example, MWIR coupled to chemometric tools
such as a PLS-DA can be used to distinguish polymetallic sulphide ore and waste materials using
the spectra as chemical fingerprints of the mineralogy.

3.2 Imaging technologies

3.2.1 Hyperspectral Imaging (VNIR/SWIR)

Hyperspectral imagers collect image data in hundreds of narrow adjacent spectral bands resulting in
3D multivariate data structures. Hyperspectral imaging is mainly used in airborne or spaceborne
remote sensing application. The recent advancement of the technology resulted in laboratory based
and field-based platforms (FLSMIDTH., 2017; HYSpex., 2017; Specim., 2017; Nasrullah, 2014;
Schneider, 2011; Corescan., 2017). Depending on the sensor type and set-ups, hyperspectral images
with very high spectral and spatial resolution can be acquired. Hyperspectral cameras operate over a
wide range of the electromagnetic spectrum, the choice is application dependent. In this paper, the
use of VNIR (0.4-1.0µm) and SWIR (1.0-2.5µm) hyperspectral images for the characterization of
the materials from the test case using rock chips and drill core samples were assessed.

37
REAL TIME MINING - Conference on Innovation on Raw Material Extraction Freiberg 2019

Prior to data analysis, the raw VNIR and SWIR hyperspectral images were pre-processed using
normalization, spectral subsetting, spatial subsetting, spike correction and masking missing values
techniques. To isolate noise from the signal, the minimum (or maximum) noise fraction (MNF)
transformation was implemented in two cascaded Principal Component transformations. The inver-
se transformed MNF images were used to generate 2D scatter plots. The spectral prototypes (end-
members) were identified from the scatter plots. To cross check the uniqueness of the selected end-
members, Pixel Purity Index (PPI) was computed by projecting n-dimensional scatterplots onto a
random vector. The collected spectral prototypes were interpreted to identify the possible minerals
using spectral libraries such as USGS (Clark et al., 2003), TSG (AusSpec International Ltd., 2008)
and JPL (Groves et al., 1992). Most of the unique spectra were interpreted however there are un-
identified unique spectra as well. Training sets (Region of Interest - ROI) were generated using the
spectrally unique pixels (endmembers). The ROI’s were used to produce mineral maps that show
minerals distribution and pixel abundances using Spectral Angle Mapper (SAM). One of the advan-
tages of SAM classifier is its insensitiveness to illumination and albedo effects (Yoon and Park,
2015).

The minerals identified using the VNIR data include: the sulphides (e.g pyrite, galena, sphalerite
and chalcopyrite), the ferric iron minerals (hematite and goethite) and carbonates (siderites) (Figure
3). Furthermore, mixed spectra were also observed. Whereas, the minerals identified using the
SWIR data include: mica (muscovite), sulphate (gypsum), clay minerals (montmorillonite and illi-
te), carbonates (siderite), tectosilicate (quartz), phyllosilicate (Mg + Fe chlorite ), sulphide ores (just
with no features and results with featureless line) and mineral mixture (e.g Muscovite + siderite)
(Figure 4). The identified minerals were further validated using XRF (oxide analysis results), XRD
and EMPA data. In addition, visual inspection was performed to validate the sulphide minerals
identification (since most of the test case sulphide minerals are visually distinct).

a)
b)
Siderite
Galena
Goethite
Mixed

Pyrite
Unclassified

Figure 3: (a) a drill core color composite VNIR image and a classified image (b) rock sample color
composite VNIR image and a classified image

Sulphide minerals are SWIR inactive thus do not exhibit features in SWIR data. However, the fea-
tureless nature of the minerals in the SWIR spectra was used as characteristic value to map ore ver-
sus waste materials (Figure 5 (b)). Thus, the technique is promising for ore-waste discrimination.
The VNIR data show a great potential to detect and identify among the sulphide minerals. However,
it needs careful analysis and validation since the sulphides do not show any particular absorption
features. Automation of the mineral identification process might be challenging due to lack of parti-
cular absorption features of the sulphide minerals and the matrix effect owing to the mineral mix-
tures. However, the variation in the spectra can be accommodated by considering a training library

38
REAL TIME MINING - Conference on Innovation on Raw Material Extraction Freiberg 2019

with wider range of mineral mixtures simulated based on the mineral compositions of the test case
material. Owing to the acquired promising results of the two techniques and recent advancement of
the technologies that resulted in portable hyperspectral camera (e.g Specim IQ developed by Spe-
cim., 2019) the application can be extended for in-situ application of mine face mapping in under-
ground mine with suitable illumination source and robust system.

a) b)

Siderite
Muscovite + Sulphide
Muscovite
Illite
Illite+ sulphides
Unclassified

Figure 4: (a) a classified SWIR image of a rock sample (b) a drill core color composite SWIR image
with identified minerals

a)

Waste
b) Ore
Unclassified

Figure 5: a) A false color SWIR image (b) a classified image showing ore-waste discrimination

3.2.2 RGB imaging

RGB imagers characterize the reflectance property of a material and deliver 3 (red-green-blue)
spectral bands often using three independent CCD sensors or using complementary metal oxide
semiconductor (CMOS) technologies. RGB imagers are well-established techniques with rapid data
processing capabilities. The recent advancement of the technology resulted in high speed 66000 fps
and 7.5 µm pixel size RGB cameras (JAI., 2019). Portable and ruggedized systems are available
that the systems are ideal for embedding or surface mounting in harsh environment (e.g under-
ground mines) applications. Tomra., 2017 revealed the application of the technology for mineral
sorting (e.g sorting of talc and calcite). In this section, the use of RGB images for mapping of mine-
rals, fragmentation analysis and ore zone delineation in the underground mine was assessed.

39
REAL TIME MINING - Conference on Innovation on Raw Material Extraction Freiberg 2019

RGB images were acquired in-situ at the defined mine face. To cover the mine face both laterally
and vertically, the images were taken from the two vertical and multiple horizontal reference loca-
tions. Ground control points (GCP’s) were marked at the mine face and the geographic coordinates
of the GCP’s were collected using LIDAR scan by Mine Surveying and Geodesy team of TUB
Freiberg. The collected GCP’s were used to georeference and mosaic the RGB images. The data
acquisition process was controlled for illumination and distortion effects. Supervised classification
was performed using training sets (groups of pixels) that represent up to five mineral types. The
performances of three supervised classification algorithms namely Maximum likelihood (ML), Mi-
nimum distance (MD) and Spectral angle mapper (SAM) were compared. The classification ac-
curacy was computed using a confusion matrix (error matrix) that compares the ground truth classes
with the predicted or classified pixels at each ground truth location. For the classification of the
minerals of in the test case using RGB images, ML outperforms the MD and SAM classification
techniques. The details of the data acquisition process and the methodological approach developed
for knowledge extraction from RGB images are presented in (Desta and Buxton, 2017).

Pyrite/ Chalcopyrite

Galena/Sphalerite

Calcite/Quartz

Weathered material
Figure 6: a) RGB image taken at the mine face b) mineral map that shows
four mineral classes

Pyrite/ Chalcopyrite

Galena/Sphalerite

Calcite/Quartz

Host rock

Figure 7: Ore zone mineral mapping showing the three mineral types produced from georefe-
renced and mosaicked images

Another potential application of RGB imaging is for fragmentation analysis. Rock fragmentation
through blasting influences the subsequent crushing and grinding operations. Thus, it is essential
information in the mining value chain. Rock fragmentation analysis can be performed using RGB

40
REAL TIME MINING - Conference on Innovation on Raw Material Extraction Freiberg 2019

images. However, analysis of fragmentation using images has its own limitations. For example;
under-estimation or over-estimation of particle size, shadow effect and piling effect (Kemeny et al.,
1993). In this study, the fragmentation analysis results were maximized by selecting appropriate
sampling locations. To avoid the shadow effect the target areas were illuminated from different si-
des. High quality images were used. To capture the observed grain size variability multiple images
were acquired and suitable image scales were used. Multiple RGB images were taken at the muck
piles (piles generated from fragmented material just next to the blasted face) and LHD locations in
the underground mine. Figure 8 shows fragmentation analysis results of a muck pile image. The
acquired results of the fragmentation analyses are reproducible. The algorithm detects clast sizes up
to 2mm and everything below 2mm is categorized in the same grain size class. Taking in to account
the blasting parameters, the result can further be used for development of models that better predict
fragmentation in the test case.

Figure 8: a) RGB image taken at the muck pile b) the size distribution curve of the image c) grain
size analysis result

The results from the use of RGB imaging for mineral mapping, ore zone delineation and fragmenta-
tion analysis are promising. Thus, RGB imaging can further be considered as a complementary
technique. It is a simple technology with a good potential for material characterization in mining
operations. Looking forward, better results are possible with better quality RGB images.

3.3 Data fusion

Sensor combinations are required to convey a near complete description of materials. Sensor com-
binations can be implemented using a data fusion approach. Data fusion is a wide ranging subject
that can be applied using various techniques such as chemometrics. In chemometrics, data fusion
can be realized at three levels: low-level, midlevel and high-level. Low-level fusion is data level
fusion, mid-level fusion is a feature level fusion and high level fusion is a decision level fusion
(Borràs et al., 2015, Doeswijk et al., 2011, Federico., 2013). The data fusion methodological ap-
proach developed in the H2020 RTM project was presented in Desta and Buxton,( 2018).
Low-level data fusion was implemented using MWIR and LWIR reflectance spectra data. Using the
fused data block elemental concentrations of Fe and a combined Pb-Zn were predicted. Compared
to the individual techniques, the fused data block model resulted in better prediction performance.
For example, the acquired prediction accuracies of the MWIR data model is 85%. LWIR is very
well known for analysis of rock forming minerals, however using chemometric techniques the use
of the technology for elemental prediction was assessed. The Fe prediction accuracy of LWIR
spectral reflectance data combined with chemometric techniques was 88%. However, the prediction

41
REAL TIME MINING - Conference on Innovation on Raw Material Extraction Freiberg 2019

performance of the fused data block is 94% (unpublished results). Therefore, data fusion enhanced
the prediction performances of the models.

4 Discussion

4.1 Utility of different sensor types for characterization of material from the test
case

The usability of sensor technologies for characterization of the test case materials is synthesized and
evaluated in Table 2. The SWOT analysis shows the potential and threats to the application of the
technologies in mining operations. The table also addresses the observed strength of the sensors.
Table 2: SWOT analysis of the investigated sensor technologies for the use of material
characterization and their applications in underground mining operations

Technology Strength Weakness Opportunities Threats

RGB Imag- x Good for qualitative x Surface techni- x Light sources can be x Variable opera-

x Semi-quantification x The information x Improved signal pro- x Can be affected

analysis que optimized ting conditions
ing
is possible with pixel is limited to 3 cessing/Image proces- by surface impu-

x Rapid data proces- x Lower reflection x Surface rough-

count bands sing techniques rities

x Applicable for visually

available

x Can be used in color

sing - ness affects the

x Most advanced tech- x Dust affects the

distinct minerals measurements
detection and shape

x Ruggedized systems
recognition (ore- nology measurements

x Can be used for

geometry)

x Small size: ideal for

available
fragmentation analy-

x No actual contact is
sis embedding and

x Potential for mine-

surface mounting

x Non-destructive
required
ral/lithological map-

x Indirect proxy for

ping

x Good for qualitative x The smaller the x Developments are x Environmental

mineralogy/grade
VNIR
and semi- wavelength ran- dynamic and advan- influence (such

x Can be used to dis- x Potential for sensor

quantitative analysis ge the limited the cing rapidly as water and

x Surface techni-
info dust) can affect

x Well established tech-

tinguish between based sorting in-situ measure-

x Mineral mixtures
some of the sulphide que ments

x Imaging techniques x Rapid data acquisition

minerals nology

x Least commonly
affect the results
do not require actual

x Non-destructive
contact with samples used for quanti-
tative analysis
SWIR x Can be used for x Processing and x Developments are x Environmental
sulphide ore and handling of the dynamic and advan- influence (such

x Can be used for x Most advanced tech-

waste discrimination large volumes of cing rapidly as water and

x Surface techni-
data dust) can affect

x Portable instruments x Least commonly

identification of nology in situ
associated minerals que

42
REAL TIME MINING - Conference on Innovation on Raw Material Extraction Freiberg 2019

Technology Strength Weakness Opportunities Threats

x Imaging techniques -
x Image and point data
are emerging used for quanti-
do not require actual tative analysis

x Can be used for textu-

contact with the can be acquired
samples
ral and mineralogical

x The least explored x No commercial

analysis
MWIR

x Can be used for x Least explored

region of the IR but system for mine-

x Robust system is
ral identification
x Portable instrument
with a good potential

x Lack of well
sulphide ore discri- region
required for un-
x Spectra showed a
mination
x Hyperspectral imager
already available
documented mi- derground (harsh
environment) ap-
x Surface techni-
very good correlati- neral library is developed
on with Fe, the com- plication
bined Pb_Zn, SiO2, que

x Detection limit ~0.01

Al2O3, Fe2O3

LWIR x Can be used for x Surface techni- x Good potential for x Robust system is

x Advanced technology
discrimination of que mining applications required for un-

x Point and imaging

sulphide ore and - derground

x Can be used for

waste (harsh environ-
spectrometers are ment) applica-

x Portable instruments x Camera need

identification of rock emerging tions

x Spectral signal has a

forming minerals
are available robust housing
good correlation
with some of the test
case elements thus
can be used for ele-

x Detect a wide range

mental prediction

of minerals (mainly
rock forming mine-

x Detection limit~0.01
rals)

x Detect a wide range x Detection limit x Mobile units available x Sensitive to

%

x Both imaging and

Raman

x Detect some of the

of minerals (ppm level detec- vibration and

x Conflict with
tion is not at- point techniques are dust

x Enriched spectral x Weak in intensity x Provides complemen-

sulphide minerals tainable ) available
fluorescent mi-
libraries compared to Inf- tary information to nerals

x Raman signal has x High spatial resolution

rared infrared

;фϭʅŵͿ
x Ruggedized system
a very low corre-
lation with the
elemental con- available
centration of the
test case materi-

x Surface techni-
als

que

43
REAL TIME MINING - Conference on Innovation on Raw Material Extraction Freiberg 2019

4.2 Opportunities for sensing systems in mining operations

In mining, sensors can be used for different applications. For example, sensors for machine perfor-
mance monitoring, collision avoidance and material characterization. Sensors for material charac-
terization can be utilized along the mining value chain, to provide usable data on several aspects of
material under investigation. The choice of sensor for characterization of certain deposit type de-
pends on different factors, such as sensor parameters, material type (deposit type) and operational
environment. Sensor parameters are broad and discussed in Section 1. Deposit types define material
properties that are relevant to sensors measurement. Operational environment is the other crucial
factor to consider. For example, some environments require ruggedized systems due to the harsh
environmental condition, the others require sensors with high data acquisition speed such as
conveyor belt applications.
In-situ application of sensor technologies requires portable, ruggedized (that can be used under
harsh mine conditions) and high speed systems. With the current innovative advancements of sensor
technologies, technical solutions both in terms of instrumentation and application are emerging. For
example, high speed NIR sorters that are able to detect 640000 spectra per second per meter
conveyor belt with a belt speed of 3m/s are available (Robben and Wotruba, 2010), portable sys-
tems such as FTIR has ~ 2kg weight (Agilent., 2017), ruggedized systems (e.g Raman from Stel-
larNet Inc. 2019) are evolving. Sensors with enhanced sensitivity detect minerals in lower concent-
ration, thus improved sensitivity is essential for application of sensors in low-grade deposits.
Despite researchers indicated the benefit of the use of sensors in mining industries (Buxton and
Benndorf, 2013, Goetz et al., 2009, Fox et al., 2017), the use of sensors is limited due to various
factors. One of the possible reasons for the limited use of sensor technologies in mining operations
is the initial investment to purchase the instrument might be higher than the benefit to be realized.
Advancement in sensor technologies has resulted in simplified design and low cost systems, in near
future it is likely that even lower cost systems will emerge. This is one of the factors to improve the
use of sensors in mining operations.

The current demands for applications of sensor in mining include requirements in hardware design
and software tools. The hardware requirements include portability and ruggedized system. Robust
systems are required for applications in harsh environment (e.g underground mine).The software
requirements are related to advances in analytics from machine learning to improved statistical
techniques thus to transform the multivariate raw sensor signals into knowledge about the materials
under investigation. Attributed to various factors, direct fingerprinting of minerals or target ele-
ments using sensor signals might be challenging. However, the value of the property of interest can
be inferred from spectral signals through indirect observations using chemometric techniques.
The other key requirement of sensors use in mining is sensor capability for remote applications. In
this context, remote applications refer to few centimeters to meters distance between the material to
be characterized and sensor (without actual contact). The recent advancement of hyperspectral
cameras resulted in sensors that can be operated with field-based platforms (Schneider, S., 2011).
This is useful for open pit mapping and the already existing field-based platform can be modified
for underground applications (e.g application specific design for mine face mapping).

44
REAL TIME MINING - Conference on Innovation on Raw Material Extraction Freiberg 2019

Imaging technologies provide information over wider area and give spatial context compared to
point technologies (that measure spots). For example, georeferencing and mosaicking of the RGB
images provided a comprehensive view of mineral distribution over the imaged part of the mine
face. This is advantageous in understanding the spatial distribution and the relative abundance of
minerals thus to infer grade indirectly. Coordinates of the sampled areas (channel centroids) were
computed using the surveyed points and the point cloud generated using LIDAR. Therefore, spatial-
ly constrained chemical and mineralogical data were generated. This is useful to link the informati-
on from the different data sources based on location. However, the challenges related to sensors
field of view, spatial resolution, positional accuracies and material variability should be taken into
account.

4.3 Prospect for real-time analysis of material

Real-time material characterization requires rapid data acquisition, automated data processing and
rapid return of results. Thus, it involves advanced platform that integrate hardware and a high per-
formance computing software systems. Once integrated systems are developed, the predictive tech-
nologies can be deployed to deliver online data in different application areas. Such as face mapping,
drill core logging and ore sorting applications. Material flow at the potential sensor installation sites
along the mining value chain can be categorized into static and dynamic sites. Static sites are sites
with relatively slow movement of materials such as mine face, drill core logging and muck pile ap-
plications. Whereas dynamic sites are those sites with a quick flow of materials. For example, the
required real-time response of material characterization at the mine face might be in order of few
hours to few days (after blasting was performed) and conveyor application could be in the order of
milliseconds, and sometimes microseconds depending on the conveyor belt speed. Therefore, real-
time material characterization at the potential sensor mounting sites along the mining value chain
has different temporal aspects.
Predictive models were developed using the MWIR and LWIR reflectance spectra data, the models
were used to discriminate the unknown spectra into ore and waste material types. The predictive
models were trained for the prediction of elemental concentration, independent data sets were used
to assess the predictive performance of the models. The acquired ore-waste classification accuracies
and elemental prediction accuracies of the models indicate, the automation potential of the material
characterization process. Going forward, a better prediction accuracy is expected with extended
dataset in the calibration data. Likewise, for visually distinct minerals, mineral mapping using RGB
imaging is a complementary approach to the conventional mapping. However, the former gives
automated, reproducible and objective results. With well-calibrated prediction and classification
models, automation of material characterization is achievable. Fast and better prediction results are
possible with test case specific mineral libraries that take into account the spectral variation resulted
from the heterogeneous nature of the deposit type.

4.4 Opportunities with fusing of data

In diverse areas of application, there has been an ever-increasing interest in a near complete descrip-
tion of materials using multi sensor data. One of such application areas is mining operation. A com-
prehensive view of materials in mining applications is advantageous; in understanding the process

45
REAL TIME MINING - Conference on Innovation on Raw Material Extraction Freiberg 2019

of mineral formation, to understand the requirements in mineral processing, to find a relation with
indirect proxies of minerals of economic interest, to provide mineralogical information for resource
models and to convey safety information. Fusing of different data sources improves models classifi-
cation and prediction accuracies, improves precision, improves availability and reduces uncertainty.
Therefore, it significantly supports effective decision making in mining operations.

5 Conclusions

This contribution has demonstrated the usability of sensor technologies (RGB imaging,
VNIR/SWIR hyperspectral imaging, MWIR and LWIR) for the characterization of a polymetallic
sulphide ore deposit. The methodological approaches developed for each sensor technology resulted
in usable results for identification, predictability and classification of the test case materials. The
use of senor combinations should aim to maximize the accuracy of (classification and prediction)
models by minimizing the uncertainty related to models performance. Accordingly, the use of data
fusion allowed for increased predictability and classification of materials.

The use of sensor technologies for raw material characterization is rapidly growing, and innovative
advancements are observed. However, due to economically marginal deposits, deeper mine and
complex geology, there is still a need to define and develop improved technologies and innovate
approaches that can address the current and future mining challenges. One of the possible approa-
ches is, to define and develop sensor combinations using scalable data fusion algorithms. Depen-
ding on the deposit type, the ultimate sensor combinations can be optimized and deposit specific
mineral library can be developed using a combined sensor signal. Going forward, improved and
automated material characterization is possible with integrated tools that combine sensor signals
with material properties and ruggedized systems. Therefore, future research that address both the
hardware and software requirements should be conducted to fulfil the gap between in-situ and onli-
ne material characterization.

ACKNOWLEDGMENT

This research (Real-Time Mining Project) has received funding from the European Union’s Horizon
2020 Research and Innovation Programme under the Grant Agreement No. 641989. We would like
to thank Thijs Hanemaaijer at TU Delft for his input on Raman data interpretation.

REFERENCES

Agilent. 2017. FTIR Compact & Portable Systems. Retrieved February 2, 2017; Available from:
http://www.agilent.com/en/products/ftir/ftir-compact-portable-systems/4300-handheld-ftir.

Benkert T., Dietze A., Gabriel P., Gietzel J., Gorz I., Grund K., Lehmann H., Lowe G., Mischo H.,
Schaeben H., Schreiter F., Stanek K., 2015. First step towards a virtual mine—generation of a 3D
model of Reiche Zeche in Freiberg. In: Proceedings of the 17th annual conference of the interna-
tional association for mathematical geosciences (IAMG), vol I, II. Freiberg, Germany, Sept 2015,
1350–1356

46
 REAL TIME MINING - Conference on Innovation on Raw Material Extraction Freiberg 2019

Borràs, E., Ferré, J., Boqué, R., Mestres, M., Aceña, L., Busto, O., 2015. Data fusion methodologies
for food and beverage authentication and quality assessment – A review. Analytica Chimica Acta,
891, 1-14. doi: https://doi.org/10.1016/j.aca.2015.04.042
Bumbrah, G. S., Sharma R. M., (2016). Raman spectroscopy – Basic principle, instrumentation and
selected applications for the characterization of drugs of abuse. 6(3): p.209-215. Egyptian Journal of
Forensic Sciences
Buxton, M.W.N., Benndorf, J., 2013. The Use of Sensor Derived Data in Optimization along the
Mine-Value-Chain An Overview and Assessment of Techno-Economic Significance, in Interna-
tional Congress of the ISM 2013. p. 324-336.
Corescan., 2017. Hyperspectral core scanning. Retrieved February 2, 2017, from
http://www.corescan.com.au/
Desta, F. S., Buxton, M. W. N., 2018. Automation in sensing and raw material characterization - a
conceptual framework. Paper presented at the 2018 IEEE/RSJ International Conference on Intelli-
gent Robots and Systems (IROS). Madrid , Spain
Desta, F. S., Buxton, M. W. N., 2018. Chemometric Analysis of Mid-Wave Infrared Spectral Re-
flectance Data for Sulphide Ore Discrimination. Mathematical Geosciences. doi: 10.1007/s11004-
018-9776-4
Desta F.S., Buxton M.W.N., 2017. The use of RGB imaging and FTIR sensors for mineral mapping
in the Reiche Zeche underground test mine, Freiberg. Paper presented at the Real Time Mining -
Conference on Innovation on Raw Material Extraction, Amsterdam, The Netherlands

Doeswijk, T. G., Smilde, A. K., Hageman, J. A., Westerhuis, J. A., van Eeuwijk, F. A. 2011. On the
increase of predictive performance with high-level data fusion. Analytica Chimica Acta, 705(1), 41-
47. doi: https://doi.org/10.1016/j.aca.2011.03.025

Federico, C., 2013. “A Review of Data Fusion Techniques,” The Scientific World Journal, vol.
2013, Article ID 704504, pp 19
FLSMIDTH. 2017. HyLogging. Retrieved June 20, 2017, Available from
www.flsmidth.com/hylogger
Fox, N., Parbhakar-Fox, A., Moltzen, J., Feig, S., Goemann,K., Huntington, J., 2017, Applications
of hyperspectral mineralogy for geoenvironmental characterisation. 107, 63-77. Minerals Engineer-
ing.
Gaft, M., Reisfeld, R., Panczer, G., 2005. Modern Luminescence Spectroscopy of Minerals and
Materials. Germany: Springer Link.
Griffith, W., 1975. Raman spectroscopy of terrestrial minerals. C. Karr Jr.(editor), Infrared and
Raman
Spectroscopy of Lunar and Terrestrial Minerals. Academic press, New York, 35: pp 229-324

Goetz, A.F.H., Curtiss, B., Shiley, D.A., 2009. Rapid gangue mineral concentration measurement
over conveyors by NIR reflectance spectroscopy. 22(5), 490-499. Minerals Engineering.

47

View publication stats