Analytical Techniques For Online Mineral Identification

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/330831768
ANALYTICAL TECHNIQUES FOR ONLINE MINERAL IDENTIFICATION
Thesis · February 2019

DOI: 10.13140/RG.2.2.27721.19043
CITATIONS READS
0 1,046
1 author:
Navid Khajehzadeh
Aalto University
11 PUBLICATIONS 108 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Doctoral Thesis View project
All content following this page was uploaded by Navid Khajehzadeh on 02 February 2019.
The user has requested enhancement of the downloaded file.

Depar tment of Electrical Engineering and Automation
ANALYTICAL
Navid Khaje hzade h
TECHNIQUES FOR
ONLINE MINERAL
IDENTIFICATION
Navid Khajehzadeh
Aalto Universit y
D O C TO R A L
DISSERTATIONS
2018
Aalto University publication series
DOCTORAL DISSERTATIONS 177/2018
ANALYTICAL TECHNIQUES FOR

ONLINE MINERAL IDENTIFICATION
Navid Khajehzadeh
A doctoral dissertation completed for the degree of Doctor of

Science (Technology) to be defended, with the permission of the
Aalto University, School of Electrical Engineering, at a public
examination held at the hall AS1 of the school on 5 October 2018
at 12 noon.
Aalto University
School of Electrical Engineering
Department of Electrical Engineering and Automation
Control Engineering Group
Supervising professor
Professor Arto Visala, Aalto University, Finland
Thesis advisor
Docent Kai Zenger, Aalto University, Finland
Preliminary examiners
Professor Emmanuel Manlapig, University of Queensland, Australia
Associate Professor Juha Toivonen, Tampere University of Technology
Opponent
D.Sc (Tech.) Esko Juuso, Oulu University, Finland
Aalto University publication series

DOCTORAL DISSERTATIONS 177/2018
© 2018 Navid Khajehzadeh
ISBN 978-952-60-8182-3 (printed)

ISBN 978-952-60-8183-0 (pdf)
ISSN 1799-4934 (printed)
ISSN 1799-4942 (pdf)
http://urn.fi/URN:ISBN:978-952-60-8183-0
Unigrafia Oy
Helsinki 2018
Finland
Ab s t ra c t
Aalto University, P.O. Box 11000, FI-00076 Aalto www.aalto.fi
Author
Navid Khajehzadeh
Name of the doctoral dissertation
P u b l i s h e r School of Electrical Engineering
U n i t Department of Electrical Engineering and Automation
S e r i e s Aalto University publication series DOCTORAL DISSERTATIONS 177/2018
F i e l d o f r e s e a r c h Control Engineering
M a n u s c r i p t s u b m i t t e d 18 May 2018 D a t e o f t h e d e f e n c e 5 October 2018
P e r m i s s i o n t o p u b l i s h g r a n t e d ( d a t e ) 21 May 2018 L a n g u a g e English
Monograph Article dissertation Essay dissertation
Abstract
The exploration and mineral processing phases of mining need advanced measurement and
analytical techniques to speed up the process of mineral identification. Rapid mineral identification
is necessary for a better process control and effi cient use of energy and raw materials.
Al t ho u g h t he s p e ct ro s c o p ic t e c hniqu e s p e rf orme d in t he l ab o rat o rie s p ro vid e ac c u rat e re s u l t s , t he

process of measuring, data preparation and analysis is slow and far from the goal of rapid analysis.
Geologists are enthusiastic about the abundance of the minerals as well as the mineral map of the
surface of the drill core samples. Therefore, there is a need to develop new techniques enabling
rapid mineral identifi cation of the rock and ore drill core samples.
Ore beneficiation is yet another process where online mineral identification is required. The control
of the fl otation processes is strongly relying on the online analysis of elemental contents in the
process feed, fi nal product and tailings, as well as in the intermediate material fl ows inside the
process. However, it is the minerals that affect how the ore behaves in the fl otation process.
Therefore, online analysis of minerals would enable more accurate process control in several
fl otation applications when compared to the online elemental analysis.
Th i s t h e s i s i n ve s t i g a t e s w h e t h e r t h e d a t a o f t h e c u r re n t l y a va i l a b l e m e a s u r e m e n t t e c h n i qu e s c o u l d
be utilize d t o e xt ract mine ralog ical informat ion. The t arg e t is qualit ative and quant itat ive mine ral
identifi cation with the smallest amount of investment or modifi cation of the instruments.
The main results of the thesis show that the spectral integration of the commonly used
spectroscopic techniques such as X-ray fl uorescence (XRF), Laser-induced fl uorescence (LIF),
Laser-induced breakdown spectroscopy (LIBS), reflectance spectroscopy and Raman spectroscopy,
enables the rapid identification of mineral contents. Advanced statistical techniques such as partial
least squares (PLS) provide a means of determining the mineral contents from the available
measured spectra.
Keywords Mineral identification, X-ray fluorescence, Laser-induced fluorescence, Laser-induced

breakdown spectroscopy, Refl ectance spectroscopy, Data fusion, Data modeling,
Partial least squares regression
I S B N ( p r i n t e d ) 978-952-60-8182-3 I S B N ( p d f ) 978-952-60-8183-0
I S S N ( p r i n t e d ) 1799-4934 I S S N ( p d f ) 1799-4942
L o c a t i o n o f p u b l i s h e r Helsinki L o c a t i o n o f p r i n t i n g Helsinki Y e a r 2018
P a g e s 150 u r n http://urn.fi /URN:ISBN: 978-952-60-8183-0
Preface
The research described herein was conducted under the supervision of Docent
Kai Zenger to whom I would like to express my sincere gratitude because
this research work would not have been possible without his support. Kai
Zenger always encouraged me when I was in disappointing situations and he
was supportive spiritually, scientifically and financially. What I learned from
Kai was beyond just research and I gained some life experiences from him that
will last a lifetime.
I would like to thank Prof. Arto Visala for taking the role of my supervisor
and for being helpful whenever we had to take care of organizational matters.
Many thanks to Dr. Olli Haavisto who gave me the chance of being a doc-
toral student and always supported me by his helpful technical, mathematical
and experimental comments. I am grateful to Olli for his endless support,
enthusiasm, knowledge and friendship. His professional ideas and positive
attitude have always been a great support and relief for me.
My sincere thanks to Dr. Heikki Häkkänen who gave me the access to his
laboratory and research facilities. Heikki taught me some skills of measure-
ment that have been of immeasurable value.
My gratitude also goes to Dr. Tommi Kauppinen, my colleague, co-author
and friend. Tommi was always supportive in our common research work and
he made the first glance of my research very smooth.
I would like to express my appreciation to M.Sc. Lauri Koresaar because of
his contribution to my research. The breakthroughs came in when Lauri, Olli
and I started working as a team. Lauri’s expertise in measurements and data
collection was of great help to my research.
Furthermore, I would like to thank Dr. Jani Kaartinen for trusting my
abilities and arranging the collaborations with the Outotec Company.
i
Preface
My heartfelt thanks go to all my labmates, colleagues and academic friends:

Kourosh Latifi, Sergey Samokhin, Christian Giovanelli , Pedram Daee, Saeed
Mardani, Saeed Mirzaeifar, Gerardo Santillan, Philipp Nguyen, Ville Liimatainen,
Zoran Cenev and many others for making the atmosphere of our research work
entertaining and happy.
Last but not least, I would like to thank my family: my mother and my
father, my brothers Saman and Hooman and my sister Sara for their unlimited
support throughout my life. You have always been with me by believing in
me and motivating me. My heart just keeps thanking you and thanking you.
Espoo, August 30, 2018,
Navid Khajehzadeh
ii
Contents
Preface i
Contents iii
List of Publications vii
Author’s Contribution ix
Nomenclature xi
1. Introduction 1
1.1 Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Problem statement in exploration phase of mining . . . 1
1.1.2 Problem statement in mineral processing phase of mining 2
1.2 Objectives and scope of the thesis . . . . . . . . . . . . . . . . . 2
1.3 Research Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3.1 Hypothesis 1: Does the combined Raman spectroscopy
and laser-induced fluorescence provide faster detection
of fluorescent minerals on the surface of the drill core
samples? . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.2 Hypothesis 2: Can the elemental data of laser-induced
breakdown spectroscopy be converted to mineralogical
information for surface scanning of the drill core samples? 4
1.3.3 Hypothesis 3: Is it possible to perform quantitative el-
emental to mineral conversion of LIBS data to obtain
online mineralogical information for better control of
flotation processes? . . . . . . . . . . . . . . . . . . . . 5
iii
Contents
1.3.4 Hypothesis 4: Does data fusion of different measurement

techniques improve on-steam and quantitative miner-
alogical identification of slurry samples? . . . . . . . . . 5
1.4 Contributions of the thesis . . . . . . . . . . . . . . . . . . . . . 6
1.5 Publication summaries . . . . . . . . . . . . . . . . . . . . . . . 7
2. Experimental Background 11
2.1 Drill core analysis in the mineral exploration phase of mining . 11
2.1.1 Importance of rapid mineral exploration . . . . . . . . . 12
2.1.2 Rapid analysis methods of drill core samples . . . . . . 12
2.2 Laser-based measurement techniques for dill core surface scanning 14
2.2.1 Laser-induced fluorescence . . . . . . . . . . . . . . . . . 14
2.2.2 Laser-induced fluorescence imaging . . . . . . . . . . . . 16
2.2.3 Raman spectroscopy . . . . . . . . . . . . . . . . . . . . 18
2.2.4 Laser-induced breakdown spectroscopy . . . . . . . . . . 18
2.3 Online mineral analysis of slurry samples . . . . . . . . . . . . 20
2.3.1 Control of flotation process . . . . . . . . . . . . . . . . 21
2.3.2 Online LIBS slurry analyzer . . . . . . . . . . . . . . . . 23
2.3.3 Online XRF slurry analyzer . . . . . . . . . . . . . . . . 24
2.3.4 Online reflectance spectroscopy slurry analyzer . . . . . 26
2.3.5 Data fusion of slurry analyzers . . . . . . . . . . . . . . 27
3. Theoretical background 31
3.1 Digital images and measuring color . . . . . . . . . . . . . . . . 32
3.2 Non-negative least squares . . . . . . . . . . . . . . . . . . . . . 35
3.3 Logistic regression . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.4 Principal component analysis and singular value decomposition 40
3.5 Partial least squares regression . . . . . . . . . . . . . . . . . . 43
3.6 Model selection and validation . . . . . . . . . . . . . . . . . . 45
4. Results and discussion 47

4.1 Drill core surface scanning and mineral map generation . . . . 47
4.1.1 Positives, negatives and significance of the work com-
pared to the state-of-the-art . . . . . . . . . . . . . . . . 48
4.2 Qualitative mineral identification using LIBS technique . . . . 50
iv
Contents
4.3 Quantitative mineral identification of tailing slurries using LIBS

technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.4 Online mineral identification by data fusion of LIBS, reflectance
spectroscopy and XRF measurement techniques . . . . . . . . . 55
5. Conclusions 61
6. Future work 63
References 65
Publications 73
v
Contents
vi
List of Publications
This thesis consists of an overview of the following publications which are

referred to in the text by their Roman numerals.
I Kauppinen, T., Khajehzadeh, N. and Haavisto, O. Laser-induced fluores-

cence images and Raman spectroscopy studies on rapid scanning of rock
drill core samples. International Journal of Mineral Processing, 132, 26-33,
https://doi.org/10.1016/j.minpro.2014.09.003, November 2014.
II Khajehzadeh, N., Kauppinen, T., Häkkänen, H. Laser-induced fluorescence

imaging: detection of fluorescent minerals and estimation of abundance. In
International Mineral Processing Congress, www.gecaminpublications.com/impc2014,
October 2014.
III Khajehzadeh, N., Kauppinen, T. Fast mineral identification using elemen-

tal LIBS technique. In 4th IFAC Workshop on Mining, Mineral and Metal
Processing MMM, 48, 17, 119-124, https://doi.org/10.1016/j.ifacol.2015.10.089,
October 2015.
IV Khajehzadeh, N., Haavisto, O. and Koresaar, L. On-stream and quantita-

tive mineral identification of tailing slurries using LIBS technique. Miner-
als Engineering , 98, 101-109, https://doi.org/10.1016/j.mineng.2016.08.002,
August 2016.
vii
List of Publications
V Khajehzadeh, N., Haavisto, O. and Koresaar, L. On-stream mineral iden-

tification of tailing slurries of an iron ore concentrator using data fusion of
LIBS, reflectance spectroscopy and XRF measurement techniques. Miner-
als Engineering, 113, 83-94, https://doi.org/10.1016/j.mineng.2017.08.007,
August 2017.
viii
Author’s Contribution
Publication I: “Laser-induced fluorescence images and Raman

spectroscopy studies on rapid scanning of rock drill core samples”
This article is an equal contribution between Tommi Kauppinen and Navid

Khajehzadeh. Tommi Kauppinen was responsible for the Raman spectroscopy
measurements of the samples and Raman spectral analysis. The author of dis-
sertation was in charge of LIF measurements of the samples, image processing
and abundance calculation of minerals. The writing of manuscript was divided
between authors accordingly. D.Sc. Olli Haavisto supervised the research and
contributions of the authors.
Publication II: “Laser-induced fluorescence imaging: detection of

fluorescent minerals and estimation of abundance”
The author was responsible for the contents of the publication and he wrote
the whole manuscript. The Raman results reported in Publication I was used
in this publication, and Tommi Kauppinen has been listed as an author ac-
cordingly. D.Sc. Heikki Häkkänen was the supervisor of the laboratory mea-
surements and commented on the manuscript.
Publication III: “Fast mineral identification using elemental LIBS

technique”
The author was solely responsible for the whole publication. Doctoral candi-
date Tommi Kauppinen commented on the manuscript.
ix
Author’s Contribution
Publication IV: “On-stream and quantitative mineral identification

of tailing slurries using LIBS technique”
The author was responsible for the analysis and modeling of the data. The
author wrote the whole manuscript. D.Sc. Olli Haavisto was supervising the
results and progress of the research. He commented on the manuscript. M.Sc.
Lauri Koresaar was responsible for the LIBS measurements and generating
raw data.
Publication V: “On-stream mineral identification of tailing slurries

of an iron ore concentrator using data fusion of LIBS, reflectance
spectroscopy and XRF measurement techniques”
The author was responsible for the modeling and data analysis. He also pro-
duced the corresponding results. The majority of the article was written by the
author. Sections 2.3 and 2.3.1 were written by D.Sc. Olli Haavisto who was
the expert of reflectance spectroscopy and sections 2.4 and 2.4.1 was written
by M.Sc. Lauri Koresaar as the expert of laboratory measurements.
x
Nomenclature
Acronyms
CIE international commission on illumination
e Euro
EDXRF energy-dispersive X-ray Fluorescence
HIS hue, intensity and saturation
HLS hue, lightness and saturation
HSV hue, saturation and value
HVS human visual system
iid independent and identically distributed
IR infrared
keV kilo electron volt
L*a*b* average values of Lab coordinates
Lab Lightness, redness-greeness, yellowness-blueness
LHD load haul dumper
LIBS laser-induced breakdown spectroscopy
LIF laser-induced fluorescence
LIPS laser-induced plasma spectroscopy
LWIR long-wave infrared
xi
Nomenclature
M million
mJ millijoules
MLA mineral liberation analyzer
MSE mean square error
NIPALS non-linear iterative partial least squares
nm nanometer
NNLS non-negative least squares
PCA principal component analysis
PCR principal component regression
PLS partial least squares
RGB red, green, blue
SVD singular value decomposition
SWIR short-wave infrared
TIR thermal
UV Ultraviolet
VNIR visible and near-infrared
WDXRF wavelength-dispersive X-ray Fluorescence
XRD X-ray diffraction
XRF X-ray fluorescence
Symbols
(c1 , ..., cn )T coefficients of reference spectra
(xi , y i ) a training point
:= sign of updating a parameter value
α learning rate
xii
Nomenclature
Ȳ mean value of output variables
ΔE color difference
∂
∂ (.) partial differential
Λ matrix of eigenvalues
λ parameter of regularization
λi variance of each score vector
∇ gradient
. L2 norm
Φ matrix of loading vectors of output variables in PLS
φ loading vectors of output variables in NIPALS algorithm
Ψ right matrix of singular value decomposition
Σ diagonal matrix of singular value decomposition
Θ matrix of eigenvectors or loading vectors of predictor variables
θi eigenvector
Ξ left matrix of singular value decomposition
A absorbance
a molar absorptivity of the absorber
b thickness
Bpls mapping matrix of partial least squares
c concentration
d coefficients of non-negative least squares
E matrix of errors
E(.) cost function
F matrix of parameters
xiii
Nomenclature
G matrix of errors in eigenvalue decomposition of output vari-

ables
g(.) logistic function
H hue
h predictor function
hθ (.) hypothesis function h with parameters θ
i index variable
I0 incident light
IR reflected light
IT transmitted light
j index variable
k index variable
L(.) likelihood function
m number of data points
n number of parameters
P probability
p probability density function
R reflectance
R , G , B intermediate numbers calculated from red, green and blue

color components
R2 coefficient of determination
Rref intensities of the reference spectra
S color saturation
T transmittance
U score matrix of outputs in PLS
V color value
xiv
Nomenclature
W∗ matrix of weights in PLS algorithm
wi vector of weights in NIPALS algorithm
X matrix of independent variables
Y matrix of dependent variables
Z matrix of score vectors
Z a subset of score vectors
zi score vector
xv
Nomenclature
xvi
1. Introduction
1.1 Context
Global warming and environmental changes have increased the concerns about
natural and specifically water resources. Mining industry is a process where
large amounts of energy and water are consumed, and therefore every ac-
tion resulting in energy savings is an advantage. The exploration and min-
eral processing phases of mining need advanced measurement and analytical
techniques to speed up the process of mineral identification. Rapid mineral
identification is necessary for better process control and efficient use of en-
ergy and raw materials. However, the majority of the online and on-stream
analytical techniques provide knowledge about elemental contents or they en-
able mineralogical interpretation only when they are applied in offline mode.
Therefore, there is still a need for online mineralogical interpretation of data
with a least amount of investment in the instrumentation. Online mineral
analysis alleviates the workload of geologists, enable more cost- and energy-
efficient mining activities, save more water, lead to better process control and
provide reasonable chances to develop the sustainability of mining. There-
fore, rapid mineral identification is important in the planning, execution and
operation of a mining business.
1.1.1 Problem statement in exploration phase of mining
Analyzing the contents and properties of rock samples and mineral ores is
an essential part of the whole mining and minerals processing work chain.
During mineral exploration and orebody delineation, diamond drill cores are
1
Introduction
an important source of information for elemental contents, mineralogy and

geophysical properties of the rock (De La Vergne, 2003). In the geological
front, drill core analysis is commonly done by manual logging. Typically the
initial logging of the sampled cores is performed by geologists, who manually
go through the cores and analyze them by visual inspection. For further
analysis, the cores are split and part of them are sent to a laboratory for
analytical measurements. This measurement process is rather expensive, may
take several months and requires a lot of manual work. Although the results
of the spectroscopic techniques performed in laboratory conditions are usually
precise, the process of measuring, data preparation and analysis is slow and far
from the goal of rapid analysis. On the other hand, geologists are enthusiastic
about the abundance of the minerals as well as the mineral map of the surface
of the drill core samples. Therefore, there is a need to develop new techniques
which speed up the process of rock and ore drill core scanning.
1.1.2 Problem statement in mineral processing phase of mining
Ore beneficiation is yet another process where online mineral identification is

required. Ore beneficiation is the process where commercially valuable miner-
als are extracted from the mined ores. Flotation is one of the main methods
for ore beneficiation, and typically the control of the flotation processes is
strongly relying on the online analysis of elemental contents in the process
feed, final product and tailings, as well as in the intermediate material flows
inside the process. However, it is the minerals that affect how the ore behaves
in the flotation process. In mineral processing, on-stream analysis is still done
by elemental identification and quantification. Thus, online analysis of min-
erals is needed and it would enable more accurate process control in several
flotation applications compared to the online elemental analysis.
1.2 Objectives and scope of the thesis
Having discussed about the importance of online mineral identification and the
need for speeding up the geological surveys, there is the question whether min-
eral identification can be done effectively, fast and with minimum investment.
Despite the fact that modern mineral exploration has been driven largely by
technology, no mature mineral analysis technologies for rapid drill core logging
2
Introduction
and online flotation control are available. Therefore, advanced measurement

and analytical techniques are needed to aid and speed up mineral quantifi-
cation. As lasers enable rapid and modern measurement applications, the
laser-based methods have recently become of increasing importance in the
mining industry. However, most of the contemporary online and on-stream
laser-based techniques used for identification are based on elemental analysis
or performed in laboratory conditions.
To this extent, this thesis at first concentrates on the on-site and qualitative
mineral analysis of the drill core samples in the exploration phase of mining.
The central aim is to provide mineralogical information from the surface of
the drill core samples in order to make the process of drill core logging faster.
The thesis aims to utilize the available laser-based measurement techniques to
extract further knowledge from their generated data. Moreover, some laser-
based measurement techniques have either complex measurement setup or
they are naturally slow for online or on-stream mineral identification. Such
techniques are usually applied in laboratory conditions, while they enable
accurate identifications of minerals. Therefore, another objective of this thesis
is to employ all possible information provided by those slow but accurate
techniques in order to enable online mineral identification of drill core samples.
After that, the research is extended to on-stream and quantitative mineral
identification in the concentration phase. Dominant Laser-based measurement
techniques are introduced, their pros and cons are discussed, and the ultimate
target is to apply advanced data pre-processing and statistical methods to
innovatively derive mineralogical information from their integrated spectra.
Developing new analytical methods and measurement techniques can benefit
industry and customers in all phases of mining life cycle.
1.3 Research Hypotheses
This research work is built on the hypothesis that the potential and capabilities
of the available online and laboratory measurement techniques are not fully
utilized. There are still vast amount of knowledge in the produced data that
can be exploited and it makes the geological surveys automated, speeds up
the process of drill core sorting and improves the flotation process controls.
Moreover, the application of the developed measurement devices and ana-
3
Introduction
lytical techniques in this research can be generalized to other phases of the

mineral processing.
1.3.1 Hypothesis 1: Does the combined Raman spectroscopy and

laser-induced fluorescence provide faster detection of
fluorescent minerals on the surface of the drill core samples?
Raman spectroscopy is an accurate but slow method used for direct mineral
identification mainly in laboratories (Das and Agrawal, 2011). Laser-induced
fluorescence (LIF)(Kinsey, 1977) is a sensitive, moderately easy to implement
and a well-studied measurement technique used for detecting the fluorescent
minerals based on their emission colors (Gaft et al., 2015), the technique is
explained in detail in section 2.2.1. Prior to drill core scanning, employing an
accurate technique such as Raman spectroscopy will provide some reference
data about the minerals positions. This geospatial data on the surface of the
samples should enable calibration of LIF technique for automated and fast
surface scanning. Therefore, this research aims to demonstrate the merger
of Raman spectroscopy and LIF image analysis as a powerful combination to
detect and locate the constituent fluorescent minerals of rocks on the round-
side surface of the drill core samples.
1.3.2 Hypothesis 2: Can the elemental data of laser-induced

breakdown spectroscopy be converted to mineralogical
information for surface scanning of the drill core samples?
Laser-induced breakdown spectroscopy (LIBS) (Cremers and Knight, 2006)

is a well-known and widely applied elemental measurement technique in the
mining industry (Section 2.2.4). The wide range of the elements which can be
detected by LIBS and the simplicity of the measurement setup promotes the
idea of LIBS application for fast scanning purposes. Considering the broad
range of elemental information provided by LIBS, advanced data processing
techniques should enable further knowledge extraction about the mineralogy
of the samples. Moreover, since LIBS and LIF techniques can operate with the
same source of excitation light at similar wavelengths (UV range), mineralog-
ical data produced by LIBS is expected to provide supplementary information
for LIF surface scanning. Such combinations have been investigated in some
research works (Haavisto et al., 2013a). Therefore, this research work seeks
to identify minerals qualitatively from the elemental lines of LIBS spectra us-
4
Introduction
ing advanced statistical techniques. Additionally, the acquired mineralogical

knowledge is expected to add complementary information for LIF drill core
scanning.
1.3.3 Hypothesis 3: Is it possible to perform quantitative

elemental to mineral conversion of LIBS data to obtain
online mineralogical information for better control of
flotation processes?
In addition to the elemental knowledge, better control of flotation processes re-

quires close-to-real-time knowledge about the mineralogy of the slurry streams.
On-stream measurement of the aqueous samples is challenging due to the nat-
ural limitations caused by moving particles and liquids. LIBS and X-ray flu-
orescence (Beckhoff et al., 2007) are the prevailing measurement techniques
used in the current on-stream slurry analyzers. The features of LIBS technique
(Sections 2.3.2 & 1.3.2) boosts the idea of quantitative mineral identification
from the elemental LIBS data. A proper sampling system, advanced data
pre-processing techniques and a proficient data modeling method should en-
able the idea of on-stream and quantitative LIBS mineral identification. This
research work attempts to verify the idea of data-based elemental to mineral
conversion and to provide means of proof for executability of the approach.
1.3.4 Hypothesis 4: Does data fusion of different measurement

techniques improve on-steam and quantitative mineralogical
identification of slurry samples?
Although the elemental to mineral conversion of LIBS spectra provides valu-

able mineralogical information, it will be discussed that LIBS technique alone
is incapable of distinguishing between minerals with similar elemental contents
such as magnetite (Fe3 O4 ) and hematite (Fe2 O3 ). LIBS and XRF techniques
both have high capabilities for measuring elements. Nowadays, both tech-
niques are widely applied as cross-belt analyzers of the crushed ore and slur-
ries of the flotation processes. Furthermore, there are several reports about
the capability of reflectance spectroscopy for detection of minerals (Scheinost
et al., 1998), and the technique is described in detail in section 2.3.4. The three
mentioned techniques (LIBS, XRF and reflectance spectroscopy) provide valu-
able information and bring the idea that data fusion of their measured spectra
should enable further utilization of their produced data for better mineralog-
5
Introduction
ical interpretation. Therefore, this research work seeks to validate the im-
provement of the on-stream and quantitative mineral identification of slurry
samples by fusion of LIBS, reflectance spectroscopy and X-ray fluorescence
(XRF) spectra.
1.4 Contributions of the thesis
The main contribution of this thesis is extraction of mineral information from

the elemental data called elemental to mineral conversion. The novelty is
data-based mineral identification with the slightest need to change the instru-
mentation. The measured spectra of multiple analyzers are combined using
statistical techniques and it results in novel and improved methods for assay-
ing drill core samples and slurries. The contributions of this work are listed
below:
• Drill core analysis
– producing a mineral map of the fluorescent minerals present on the surface

of the drill core samples using Laser-Induced Fluorescence (LIF) and image
processing, entitled LIF-imaging (Publication I & Publication II)
– calibration of LIF-imaging technique using precise but slow Raman spec-

troscopy and linking fluorescent colours to the related minerals (Publica-
tion I & Publication II)
– qualitative mineral identification using elemental data produced by LIBS

technique and drill core sorting using the identified minerals (Publication
III)
– presenting the feasibility of identifying multiple minerals in one spectrum

using a proper data analysis technique (Publication III & Publication IV)
• Slurry analysis
– online and quantitative mineral identification of the slurries in flotation

processes using elemental data of LIBS spectra (Publication IV)
6
Introduction
– quantitative analysis of slurry mineral contents using data fusion of 1-

LIBS and reflectance spectra 2- reflectance and iron lines of XRF spectra
(Publication V)
– presenting the feasibility of quantitative identification of multiple minerals

in the accumulated spectra of multiple measurement techniques by linear
combination of the intensities (Publication IV & Publication V)
1.5 Publication summaries
The thesis consists of five publications (appended to the end of the thesis).
This section summarizes their contents:
Publication I
This article is the first reported combination of Raman spectroscopy and LIF-
imaging applied for drill core mineral characterization. Raman spectroscopy is
an accurate but slow method for characterization of the rock drill core samples.
Laser- Induced Fluorescence (LIF) can be used as a rapid method to produce
color maps of minerals based on their fluorescent color. The sample set are
segments of rock drill core collected from a Ni-Cu multi-metal mine situated
in Kevitsa, Finland. Dill core samples are scanned by LIF technique so that
images representing colors based on the emissions of fluorescent minerals are
generated (LIF images). Some parts of the sample set, based on the exposed
colors, are measured by Raman spectroscopy and therefore minerals in the
scanned positions are identified. In this way the fluorescent colors are linked
to the related minerals (called calibration). The RGB values of LIF images
are then converted to CIELab coordinate and for each color a tolerance range
is defined so that colors in the range of tolerance are considered as the same
mineral identified by Raman spectroscopy.
Publication II
This publication is a continuation to Publication I. The main prerequisite
of LIF-imaging technique is a reference data, linking the exposed fluorescent
colors to the minerals. Therefore, LIF-imaging technique is calibrated using
precise but slow Raman spectroscopy. After linking the pixel colors to the
7
Introduction
minerals, a regularized logistic regression technique is applied to classify the

minerals based on their emitted fluorescent colors in the digital images. The
abundance of each mineral group is estimated by pixel counting. Comparison
of results with X-Ray Diffraction (XRD) laboratory assays showed that the
interpretation of the fluorescent minerals (qualitatively) through LIF images
was valid and reliable.
Publication III
This publication deals with the qualitative mineral identification from the sur-
face of rock and drill core samples. Mineral identification from the elemental
results of LIBS measured spectra is achieved using singular value decompo-
sition (SVD). The proposed approach enables detection of multiple minerals
present in each measured LIBS spectrum. It is shown that LIBS can be used
for giving complementary information to other fast scanning techniques like
LIF-imaging. Finally, it is demonstrated that point-wise measurement tech-
niques such as LIBS can be used for rapid sorting of the drill core samples
containing composite minerals of interest where small fraction of the sample
is representative of the large volume.
Publication IV
This article addresses the issue of online mineral identification of aqueous or
slurry samples in the flotation processes. A multivariate statistical technique
called partial least squares (PLS) along with non-linear iterative partial least
squares (NIPALS) algorithm is employed. The method deals with the esti-
mation of mineral contents from the elemental LIBS spectra produced by an
on-stream LIBS slurry analyzer. While the conventional methods used for
LIBS data analysis consider only a few of the significant LIBS lines, the pro-
posed method in this work takes the full spectrum of LIBS measurements into
account and it will enable detection of multiple minerals in one LIBS spec-
trum. This technique is rapid and gives the possibility of online mineralogical
analysis without any modification of the conventional LIBS setup. Quartz and
Fe-oxides were the most essential minerals for the iron-ore concentrator and
they were identified quantitatively with good accuracy.
Publication V
This article is the extended work of Publication IV. It is extremely gainful to
differentiate between minerals having identical elemental contents such as mag-
8
Introduction
netite (Fe3 O4 ) and hematite (Fe2 O3 ) or silicon-bearing minerals such as quartz

and other mixed silica minerals since they have different flotation properties.
LIBS technique alone was incapable of distinguishing between such minerals
and therefore the collection of magnetite and hematite were identified as total
iron-oxides. Fast and early detection of these minerals would improve the pro-
cess performance even further. This article extends the online mineralogical
assay of slurry samples by spectral fusion of LIBS and XRF techniques with
reflectance spectroscopy. The results show that such data integrations enable
on-stream and quantitative identification of slurry mineral contents specially
for hematite, magnetite and quartz which are important minerals in iron ore
beneficiation.
9
Introduction
10
2. Experimental Background
2.1 Drill core analysis in the mineral exploration phase of mining
Prior to mining, extensive mineral exploration activities are required to locate

and map the mineral resources. Exploration activities are also performed dur-
ing the mining when the orebody is characterized or at the time of searching
for new deposits. One of the most common and standard methods of explo-
ration in mining is diamond drilling of the ground and extracting cylindrical
samples called drill core. The drill core samples are placed into boxes and sent
to laboratories for further analysis (see Figure 2.1).
Figure 2.1. drill core samples are stored into boxes for logging and laboratory
measurements
This section familiarizes the reader with the importance of rapid exploration,
contact-less and non-destructive techniques of drill core scanning and advan-
11
Experimental Background
tages of rapid drill core analysis for the sustainability of mining.
2.1.1 Importance of rapid mineral exploration
The exploration and mining activities are strongly increasing both in Finland
and internationally. New mines are opened and the operation of the existing
ones is enhanced in order to utilize the ore resources as effectively as possible.
This development requires new technological solutions, so that new deposits
can be explored and the ore mined and processed with small energy consump-
tion and with a minimized environmental impact. The improved technology
further creates potential for novel types of service business, as the developed
solutions are generalized to service concepts that can be offered to national
and international markets.
According to the report of Haavisto et al. (2013b), in 2010, exploration
activities in Finland were about 69.9 million EUR and in 2011, the exploration
costs rose to 81 million EUR. Therefore, the importance of mineral exploration
in Finland is acknowledged. The exploration phase of mining presents its own
challenges when aiming for sustainability, and it is a time-consuming and more
often a non-rewarding effort. At the same time, one shouldn’t make damage
beyond repair to the environment and keep the local population happy and
receptive to mining activities. In the end, it is very unlikely that excavated
drill core samples give a positive signal for starting a mine.
Consequently, it is important to enhance the results of the drill core min-
eral assays and to shorten the time of analysis in order to reach to the effi-
cient exploration with less consumption of time and natural resources. There
are laboratory techniques which have potential for the online scanning appli-
cations. Furthermore, combining suitable measurement techniques develops
new automated scanning methods. This will enable fast, cost-effective and
objective mineralogical analysis of the drill core samples and provides comple-
mentary and partly substitutive information for the laboratory analysis. All
of these objectives are aimed at developing mining activities to become more
sustainable.
2.1.2 Rapid analysis methods of drill core samples
When visually inspected by geologists, drill core samples are halved and one
half of them is sent to laboratories for detailed analyses. Exact results are ob-
12
tained in laboratories through slow but reliable measurement techniques. For

laboratory analyses, samples need to be prepared, finely ground and sometimes
tested in vacuum conditions. The whole process of sample preparation, ship-
ment and laboratory examines is rather expensive, laborious and tedious. The
final results of the laboratory may take months to be ready. Online and auto-
mated scanning of drill core body using fast, contactless and non-destructive
techniques provides complementary information for laboratory analyses. Scan-
ning the drill core samples directly from the drill core boxes typically needs no
sample preparation, except for surface cleaning. It gives fast results that can
be used to ease the decision-making process during exploration and to guide
the future drilling activities without delays. This improves the exploration
process and provides economic benefits, as well as reduces the environmental
pressure of the operation.
Laser-based measuring techniques were first developed for military appli-
cations, but they have now become widely available for common uses. This
development has given rise to the multitude of application areas for lasers. One
of these application areas is mining. In order to strive for sustainability, laser-
based or contact-less methods are needed. Such techniques typically require
very little or no sample preparation, are non-destructive, require no physical
contact with the sample, and can give analysis results immediately after the
measurement. Consequently, they can potentially be applied to fast automatic
or semi-automatic analysis of large quantities of the drill core samples.
X-ray fluorescence (XRF) analysis which detects elemental contents of the
sample (Arkadiev et al., 2006), visible and near-infrared (VNIR) and in the
short-wave infrared (SWIR) reflectance spectroscopy detecting indications of
the minerals in the drill core (see e.g. Huntington et al., 2006; Keeling et al.,
2004; Tappert et al., 2011) are the predominant methods for drill core analysis
in industry. There are other spectroscopic and laser-based methods mostly ap-
plicable in the laboratory but with a high potential for industrial usage, e.g.
laser-induced breakdown spectroscopy (LIBS) which detects elemental con-
tents of the samples (Fortes and Laserna, 2010), Raman spectroscopy for direct
mineral detection (Das and Agrawal, 2011) and X-ray diffraction(XRD) ap-
plied for mineral analysis of the pulverized samples in the laboratory (XRD)(Escárate
et al., 2009). Laser-induced fluorescence (LIF) is also a suitable and prominent
measurement technique for detecting the fluorescent minerals.
The goal of this research is to demonstrate faster and better characteri-
13
zation of drill core samples by combining different contact-less measurement

techniques. There have been research studies reporting improved results by
integration of measurement techniques together (Sharma et al. (2003) & Haav-
isto et al. (2013a)). Section 2.2 will provide detail descriptions of the measure-
ment techniques applied in this research for rapid scanning of the drill core
samples.
2.2 Laser-based measurement techniques for dill core surface

scanning
Despite the capabilities in mineral and elemental identification, the discussed

laser-based or laboratory measurement techniques have limitations that usu-
ally hinders their online usage. LIF measurement technique is a fast method
with high throughput but the requirement for calibrating using a reference
data of fluorescent colors and minerals restricts its application as an auto-
mated online scanner. Raman spectroscopy is a powerful laboratory method
for direct mineral identification and there is no need for calibration. However,
Raman spectroscopy has relatively long measurement time and it is susceptible
to auto-fluorescence emission of the minerals. LIBS technique is a point-wise
measurement technique and less reliable for inhomogeneous materials while
XRF has higher resolution covering larger area on the surface of drill core. On
the other hand, LIBS has a very simple and easy-to-implement spectroscopic
setup, wider range of elements can be detected with online LIBS, and mea-
suring elements with atomic number less than 20 is rather simpler by LIBS
compared with XRF (Gaft et al., 2007). Khater (2013) points out that measur-
ing elements lighter than titanium using XRF method and without helium gas
purge or without performing within a vacuum chamber is often impractical.
2.2.1 Laser-induced fluorescence
The process of absorption and re-emission of light by a matter is called pho-

toluminescence. More generally, photoluminescence is divided into fluores-
cence and phosphorescence based on the delay between the excitation and
light emission. One way to separate these two phenomena is that in fluores-
cence the photons are emitted quickly after the excitation (delay less than
1e-8 s), whereas in phosphorescence the sample will emit the light for longer
14
duration of time.
Fluorescence enables rapid identification of fluorescent minerals where the
high-energy photons of the incident light are absorbed by the electrons of the
molecule and the molecule is excited. Then the non-radiative transition of the
absorbed energy occurs within other energy levels such as vibrational levels,
photons, thermal energy and etc. The remaining energy is then re-emitted
as a photon at a longer wavelength (lower energy) than the absorbed photon.
This phenomenon is represented by Jablonski Diagram shown in Figure 2.2.
Laser-induced fluorescence (LIF) analysis is based on measuring the fluores-
cence spectrum of minerals that are excited by an UV laser. The fluorescence
spectrum is affected by the combination of activators in the mineral structure,
and thus it is mineral specific. Activators are typically transition metal or
rare earth element ions (Waychunas, 1988). However, since the fluorescence
can be caused by very small concentrations of the activators, the quantitative
analysis based on LIF requires very good reference samples (Broicher, 1999).
Even though the term laser-induced fluorescence is used here, laser-induced
luminescence also exists in the literature.
Figure 2.2. Jablonski diagram of absorbance, non-radiative decay, and fluorescence.

Image used from the public domain, "Wikipedia"
The fluorescence or luminescence spectrum of minerals have been studied

extensively in laboratory conditions (see e.g. (Gaft et al., 1998) & (Gaft et al.,
2008) & (Reisfeld et al., 1996)). Industrial applications of LIF in mining and
minerals processing have concentrated on dilution control, assessment of ore
types and quality control. Applications for classification of phosphorous iron
15
ore and waste rock from belt conveyors and load haul dumper (LHD) buckets
have been presented by (Broicher, 1999). Nienhaus and Bayer (2003) described
an online LIF analyzer and its application for classification of hard coal, potash
ores and lignites.
2.2.2 Laser-induced fluorescence imaging
Qualitative results like mineral maps or abundance of the minerals are useful
for rapid sorting of the rock samples. The diameter of the radiation in the
point-wise measurement techniques (LIBS, LIF and Raman) is small. Calcu-
lating the abundance of the minerals adequate for industrial scales necessitates
radiating substantial number of spots to cover broader area of the drill core
surface. The feasibility of applying point-wise techniques for surface scanning
is dependent on the homogeneity of the samples. Although XRF analyzers
scan large surfaces and they are considered as rapid measurement techniques,
they explore the elemental contents of the rocks and not the minerals. Re-
flectance spectroscopy using line scanning hyperspectral imaging systems is
considerably rapid and the whole visible surface of the drill core can be ana-
lyzed (Haavisto et al., 2013b). However, there are still a noticeable number of
minerals not detectable by infrared reflectance spectroscopy techniques.
Laser-induced fluorescence imaging (LIF-imaging) is the modified version
of the conventional LIF technique. Conventional LIF analysis is based on
detecting and analyzing the measured fluorescent spectrum of minerals. How-
ever, advances in photography and digital image processing have resulted in
a wide usage of digital cameras in the real world applications. LIF-imaging is
the utilization of both LIF color detection of minerals and digital imaging of
the sample surface. Such combination has been reported in various research
works. For example, Chatterjee (2013) applies image processing techniques for
classification of rock types. Adeyemi et al. (2006) presents a high correlation
between light-induced fluorescent images and digital images applied for tooth
stain removal.
In LIF-imaging the laser beam passes through a right-angle prism redirecting
the beam to the surface of the specimen perpendicularly and the beam is then
passed through a Plano-convex lens that causes slight dispersion of the laser
beam on the surface of the sample. It enables analysis of larger surface area of
the samples and considerably improves the pace of the measurement. Figure
16
2.3 represents the schematic diagram of conventional LIF (a) and LIF-imaging
measurement setup (b). However, in order to have better quality of the image
and visible fluorescent colors, an exposure time should be set on the camera
which is the main cause of delay in the measurements. Exposure time is the
length of time that the camera’s shutter is open when taking a photo. The
required exposure time also varies based on the concentration of the fluorescent
minerals in different mines.
YAGLaser YAGLaser
266nm 266nm
PlanoͲconvexlens
Spectrometer
Collimator
Filter
Sample Sample
a) ConventionalLIF b) LIFͲimaging
Figure 2.3. a) Conventional LIF measurement setup. Point-wise scanning of the samples
using a spectrometer b) LIF-imaging measurement setup. Laser beam is
dispersed and the fluorescent colors are recorded using a digital camera
Figure 2.4 represents a LIF image of a rock sample illuminated on the white
rectangle area. The luminous color is scheelite fluorescing during the laser
illumination.
a) b)
Figure 2.4. LIF image of a scheelite rock sample
17
2.2.3 Raman spectroscopy
Raman spectroscopy is a spectroscopic technique operating by measuring the

vibrational or low-frequency modes of a system or material (Das and Agrawal,
2011). Raman is a well-known method of mineral composition analysis. Ma-
terials have a characteristic Raman scattering radiation when excited by light
and this characteristic is used as fingerprints to deduce the chemical contents
of materials. Properties of the material can be determined by analysis of the
spectrum. The most common procedure is to compare the recorded spectrum
with a library of known spectra to identify a substance. In order for species
to be Raman active, they must have vibrations (or rotations) that result in a
change in the polarisability of the molecule or material. This is why Raman
scattering in metallic substances is very weak, and cannot be detected without
special conditions (Goncharov and Struzhkin, 2003).
Raman Spectroscopy has been used for detection of diamonds (e.g. Lam-
precht et al. (2007)) or other non-metallic components of interest. Kauppinen
et al. (2013) applied Raman on the drill core samples collected from a gold
and a copper-nickle mine both situated in northern Finland. The main ap-
proach was to employ an optimization algorithm called non-negative least
squares (NNLS) (Chen and Plemmons, 2009) for detection of minerals from
the collected Raman spectra. However, they selected only signals from non-
fluorescence areas of samples and only those with clear peaks.
Despite advances and novel methods of spectral analysis developed for Ra-
man, it is a point-wise measurement technique, slow and not suitable for rapid
scanning of the samples. In the context of online scanning, this work intro-
duces Raman as a powerful method for calibrating other scanning methods
such as LIF-imaging and for providing mineralogical assays in laboratory.
2.2.4 Laser-induced breakdown spectroscopy
Laser-induced breakdown spectroscopy (LIBS) is an elemental analysis tech-

nique operating based on detection and analysis of atomic emissions. It was
first introduced by (Brech and Cross, 1962), and after the 1990s by improve-
ment in the laser technologies, LIBS became appealing for engineering appli-
cations. It is a simple spectroscopic technique mainly built on a pulsed laser
source, a focusing lens, a spectrometer and a data analysis unit. LIBS oper-
ates by employing a low-energy pulsed laser (usually tens to hundreds of mJ
18
per pulse) and using a focusing lens it generates a plasma where a small vol-
ume of the sample is vaporized. The radiation emitted by the plasma is then
measured using a spectrometer. The plasma contains a collection of colors
(wavelengths) corresponding to the present elements. Each element results
in emission peaks at characteristic wavelengths, and the height of the peaks
corresponds to the intensity of the elemental contents in the sample. Some
literature use the term laser-induced plasma spectroscopy (LIPS). The area
of application for LIBS is broad nowadays. Ayyalasomayajula et al. (2012)
reports application of LIBS for environmental monitoring. The usage of LIBS
extends to other fields like geology (Rakovskỳ et al., 2012) & (Díaz Pace et al.,
2011) or space exploration (McCanta et al., 2013) & (Wiens et al., 2013).
LIBS has been widely applied to the analysis of the elemental contents of
rocks and mineral samples (Harmon et al., 2006). One typical class of studies
include measuring the LIBS spectra of mineral samples in laboratory, and clas-
sifying the spectra using statistical multivariate methods. This approach has
been utilized successfully at least for beryls (McManus et al., 2008); carbon-
ates, pyroxenes, and feldspars (Harmon et al., 2009). Quantitative laboratory
analysis of rock powders, pellets and chips by LIBS has also been keenly stud-
ied for several ore types: iron ores ((Grant et al., 1991) & (Sun et al., 2000)
& (Death et al., 2008)); Zn and Cu ores (Gondal et al., 2007); and phosphate
ores (Rosenwasser et al., 2001).
LIBS has been used as cross-belt analyzer for industrial applications. Gaft
et al. (2007) introduces a LIBS-based analyzer for on-belt evaluation of MgO in
phosphate rock and ash content in coal. Although the accuracy of the results
was not as good as the laboratory measurements for phosphates, they pointed
out the privilege of continuous scanning using online LIBS. Fricke-Begemann
et al. (2010) presented an ore-sorting application for separating limestone and
dolomite rocks based on the Mg measurement. The setup was used as cross-
belt analyzer. The main problem for online application is that LIBS is a point-
wise measurement technique and not suitable for high throughput capacity.
For geological purposes the interest is in bulk materials, therefore a large
quantity of measurements is required to generate statistically acceptable and
reliable conclusions. One common solution is scanning with high sampling
rates typically from 1 to 20 Hz (Gaft et al. (2007) & Fricke-Begemann et al.
(2010)).
19
2.3 Online mineral analysis of slurry samples
Mineral concentration is a process where valuable minerals are separated from

the gangue. Within the variety of concentration methods, flotation, grav-
ity separation and magnetic and electrical separation are the most common
techniques among all.
The operation of gravity separation is based on the principle of different
density properties of different minerals. The relative motions of the minerals
in response to the force of gravity or other external forces like centrifugal or
magnetic forces are different. Some minerals resist against the force and some
not and this phenomenon is used for separation of different minerals. Gravity
separation has been invented prior to other separation techniques. However,
after introducing flotation or magnetic separation, there has been a decrease in
the application of gravity separation for the process of mineral concentration
(Gorain et al., 2007).
Magnetic separation is the process of separating magnetic substances from
the non-magnetic substances in a mixture using an external magnet. The
magnetic susceptibility of minerals is the main property used for mineral ex-
traction, and therefore this technique is useful for separation or concentrating
iron ore (Li et al. (2010) & Svoboda and Fujita (2003)).
Flotation is one of the main methods for ore beneficiation, and it can sep-
arate a variety of ores and particle sizes. The principle of separation in the
flotation is according to the different chemical properties on the surface of
the valuable minerals versus gangue minerals. The difference on the surface
creates hydrophobic and hydrophilic minerals. First, the slurry is formed by
grinding the ore and mixing with water. Then, the slurry is treated in a
flotation cell by adding flotation chemicals based on the properties of the ore.
After that bubbles are introduced to the slurry from the bottom of flotation
cell. Hydrophobic minerals attach to the surface of the bubbles and rise to
the surface of the cell. Usually multiple flotation cells are interconnected in a
flotation circuit to increase the output of the final concentrate.
The detailed structure of mineral concentration and separation is out of
the scope of this thesis. For more information, Gorain et al. (2007) presents a
detailed description of the flotation cell operation principle and the mechanical
structure.
Flotation process consists of other important steps mainly related to chem-
20
ical additions such as addition of flotation reagents or collectors, frothers,

depressants and activators. On the other hand, treating the tailings and re-
covery from them is also extremely important for mineral processing. In direct
flotation, valuable mineral particles are lifted to the froth and tailings are re-
moved from the bottom of the cell, whereas in reverse flotation tailings are
collected by the froth (Filippov et al., 2014).
Typically, the control of the flotation processes is strongly relying on the
online analysis of elemental contents in the process feed, final product and
tailings, as well as in the intermediate material flows inside the process. As
discussed before, the majority of the online measurement techniques are able to
identify elements. However, in mineral flotation it is the mineral contents, not
the elements of the slurries, that determine the process behavior. Direct and
real-time measurement of minerals would greatly provide additional benefit
for the process control.
2.3.1 Control of flotation process
The output of a flotation process is called recovery consisting of better pro-

portion of valuable minerals compared to the flotation feed. The proportion
of a valuable mineral in the final product is called grade. There is a trade-off
between the recovery of a flotation process and the grade of concentrate so
that better recovery leads to lower grade and vice versa. Figure 2.5 represents
the grade-recovery curve (image has been used from doctoral dissertation pub-
lished by Haavisto (2009)). The first step in the flotation control is to make
the operating point stable. This includes stabilization of slurry level and ad-
justment of the air and reagent flows. Next, the operating point should be
moved to such a level that the grade of the concentrate is not less than a
minimum level and meanwhile the recovery has been increased to a possible
maximum level. After that, using optimized control systems the operation
point of the circuit is further moved to make a better proportion of grade and
recovery possible.
Online mineral assaying provides useful information for process control e.g.
in chemical consumption optimization. The performance of flotation can be
greatly optimized by online assaying of the minerals in the slurry flows. How-
ever, online and quantitative mineral identification of the slurries is challeng-
ing. Therefore, the importance of rapid and quantitative mineral analysis is
21
Recovery
optimizing
control
set point
control
stabilizing
control
Concentrate grade
Figure 2.5. Grade-recovery curve of a flotation process. This figure is courtesy of

doctoral dissertation published by Haavisto (2009)
obvious.
Flotation is an expensive process where large amounts of water and energy
are consumed. There are a considerable number of countries having problems
in water resources (due to climate conditions and geographical location) while
having rich mines. Such countries struggle with the amount of water con-
sumed during flotation. Therefore, every action resulting in energy savings is
an advantage for them. For example, in concentrator plants where the quality
of ore is poor, the magnetic and gravity separations are usually followed by a
flotation process. It is beneficial to bypass the flotation when the grade of ore
is high enough after magnetic separation. Another example is the different
flotation properties of silicon-bearing minerals. For instance, quartz has dif-
ferent properties from other mixed silica minerals, and accurate measurement
of quartz and other silicon-bearing minerals improves the process control and
significantly optimizes the amount of chemical additions.
One of the indicators of decreased recovery is the increase in the amount of
valuable minerals in the tailing flows. This means that valuable minerals have
not been extracted properly during the separation process and they appear in
the tailings which leads to significant financial losses. Therefore, online mon-
itoring of tailing mineral contents is important. It is obvious that concurrent
monitoring of the feed and final concentrate mineral contents is also critical
22
in order to have an efficient process with high-quality final products.

Although online mineral analysis of process flows is a novel achievement
and brings substantial benefits, it is a very challenging process specially for
aqueous materials like slurries. Development of new laser-based measurement
setups only for the purpose of online mineral identification is expensive, time
consuming and is difficult to trade. It is not an easy task to convince clients
to purchase multiple analyzers for different purposes. Therefore, extracting
mineralogical information from the available elemental slurry analyzers is eco-
nomic and advantageous. This will bring many benefits for both customers
and developers so that the customer will achieve more efficiency from the
same products and the developer will be able to offer more functionality for
the products. The concept of mineral estimation from the chemical assays is
usually expressed by the term elemental-to-mineral conversion. In this work,
we introduce data-driven elemental to mineral conversion methods applicable
to slurry streams.
2.3.2 Online LIBS slurry analyzer
The structure of LIBS apparatus and its working principle was explained in
subsection 2.2.4 where LIBS was discussed as a competent technique for rock
and drill core characterization. In this section, application of LIBS for online
slurry analysis will be discussed. Since a LIBS measurement system focuses the
laser light using a lens to generate high-temperature plasma, the illuminated
liquid is vaporized and ionized locally, and therefore the plasma contains all
presented elements in the samples. As a result of vaporization, the LIBS
technique is suitable for slurry analysis or generally aqueous samples. However,
for slurry samples part of the energy of the laser light is absorbed by water
and this attenuates the signal to noise ratio. Furthermore, molecular motions,
shaking and splashing of the liquid adds noise to the signal.
Figure 2.6 represents a schematic diagram of LIBS slurry analyzer for on-
line estimation of the elemental contents. However, in order to compensate
instabilities, to have clear elemental peaks and to mitigate the noise, spectral
accumulation is performed. In real world applications the accumulation pro-
cedure is performed using a sampling system (multiplexer or cutter sampler)
which continuously brings side flows for spectral analysis. The sampled slurry
flows into the measurement cell where the LIBS apparatus is scanning a fixed
23
Sample in 25 l/min
Flow cell
Laser
Spectrometer and
camera
Sample out
Figure 2.6. Schematic of on-stream LIBS slurry analyzer. Courier 8 SL, Courtesy of
Outotec Oy (Outotec, 2018)
point on a stabilizer surface. Usually the measurement noise is considered

to be white noise, and therefore the accumulation of spectra will reveal the
elemental peaks and diminish the noisy magnitudes. As the slurry is circulat-
ing and the laser beam hits different particles each time, it is plausible that
the accumulated spectrum consists of the combination of all the elements and
minerals.
2.3.3 Online XRF slurry analyzer
X-ray fluorescence (XRF) has been a dominant elemental analysis technique

for slurry analysis, since 1960’s (Carr-Brion, 1989). X-ray fluorescence (XRF)
is based on the measurement of fluorescent radiation emitted by atoms that
are excited with X-ray radiation. The incoming X-rays eject an electron from
the inner shell of an atom. This empty position is filled by an electron from
an outer shell. When the electron moves to the inner shell, it emits electro-
magnetic radiation in the X-ray area. The frequency content of the emitted
radiation is characteristic to each element, allowing the identification of el-
emental contents of the sample (Arkadiev et al., 2006). The X-rays can be
generated using a radioactive source or an X-ray tube, which generates the x-
rays electrically. There are two approaches for measuring the X-ray radiation
from samples. Wavelength dispersive XRF (WDXRF) in which the charac-
teristic radiation is separated using a crystal. It diffracts and measures only
the intensity of the desired wavelength. The other approach is energy disper-
24
sive XRF (EDXRF) method that the fluorescent X-rays are measured by the
amount of ionization produced on a semi-conductor material. The WDXRF
technique has higher resolution and it can measure lower elemental grades,
while the EDXRF is able to measure multiple elements simultaneously.
Sample in 25 l/min ANALYZER PROBE (PRS)

WDXRF channel
Detector
50 µ thick
window
Jet flow Spectrometer
cell Signal
processing
atic
om
Aut rence X-ray tube
fe
re tter
shu
Signal
Filter processing
1 mm sensitive
layer Detector
Controlled turbulent
flow in the sensitive layer
provides a representative
sample and prevents scaling. EDXRF channel
Sample out COOLING AND TEMPERATURE

CONTROL (CCA)
Figure 2.7. Schematic diagram of on-stream XRF slurry analyzer, Courier 6 product of
Outotec Oy. Figure is Courtesy of Outotec Oy (Outotec, 2018)
Figure 2.7 represents a schematic diagram of XRF slurry analyzer, product

of Outotec Oy, consisting of wavelength dispersive XRF as the main measure-
ment and energy-dispersive technology as complementary. There are two types
of online XRF analyzers: 1- on-stream analyzer where one centralized analyzer
is installed for separate sample lines and the slurry samples are transferred to
the analyzer successively. 2- in-stream analyzer where analyzers are installed
directly on the process flows. It is obvious that the on-stream method is more
economic since one centralized analyzer is sufficient. However, it causes longer
delays between consecutive measurements from the same slurry line because
one slurry analyzer is used for measuring several slurry lines. On the other
hand, the disadvantages of in-stream method are two-folded. In addition to the
large amount of investment for measuring several slurry lines, flow conditions
of the main slurry lines will directly affect the measurements. Consequently,
centralized slurry analyzers are more in use.
25
2.3.4 Online reflectance spectroscopy slurry analyzer
Reflectance spectroscopy of minerals in the infrared (IR) wavelength range is

based on the characteristic absorption properties of minerals. The measured
spectra can be utilized to determine the mineralogy of the given sample. The
fundamental concept in reflectance spectroscopy is the interaction of incident
light with the sample surface. The original application field of IR reflectance
spectroscopy in geology is related to airborne and spaceborne remote sensing of
minerals and mapping of surface mineralogy (see Van der Meer et al. (2012)).
Vibrational and electronic processes are the two main causes of absorption.
The principle of electronic absorption is based on absorbing the energy of the
incident photon by a molecule and shifting electrons to the higher energy or-
bits. This situation is unstable and the electrons fall in to the lower energy
levels by releasing energy and emitting photons. The wavelengths of the in-
cident photon and the emitted one are not similar (wavelength of incident is
absorbed) and this is used as characteristic of the mineral. Vibrational process
happens as a result of energy absorption and vibration of the whole molecule.
Each molecule has characteristic vibrational modes relevant to specific ab-
sorption frequencies. The vibrational absorbance in minerals is characterized
by the vibrational modes of the mineral structure. Based on the molecular
structure, various vibrations are feasible. Carbonate minerals with water or
hydroxyls usually react to the incident light as a result of vibrational absorp-
tion.
Analysis of reflectance spectra is based on measuring the received light from
a sample after excitation with electromagnetic radiation. Depending on the
type of the sample, the incident light might be reflected with similar or dif-
ferent angle (called diffuse reflectance), might be absorbed by the sample or
might pass through the sample. Reflectance R is defined as the ratio between
intensities of the reflected light (IR ) and the incident light (I0 )
R = I R /I 0 , (2.1)
transmittance is defined as the ratio between intensities of the light transmit-

ted through the sample (IT ) and the incident light (I0 )
T = I T /I 0 , (2.2)
and absorbance is defined as
26
A = log(I 0 /I T ) = log(1/T ). (2.3)
According to Beer-Lambert law, there is linear relation between the concen-

tration c and absorbance A of a sample
A = log(I 0 /I T ) = acb, (2.4)
where b is the thickness and a is the molar absorptivity of the absorber (Os-
borne et al., 1993) and they are usually constant values in the calculations.
Concentration is defined as the number of moles of a solute in a volume of the
solution ( mol
m3
). Therefore, absorbance is proportional to concentration
A∝c . (2.5)
Typically the wavelength ranges of IR spectroscopy are divided to visible

near infrared (VNIR, about 350–1000 nm), short-wave (SWIR, 1000–2500 nm)
and thermal or long-wave (TIR/ LWIR, 8000–14000 nm). This study utilized
VNIR range of reflectance spectroscopy. In the VNIR region, the identification
of minerals is based on the presence or absence of transition metal ions, for
example Fe, Cr, Co or Ni, (Van der Meer et al., 2012) & (Kruse, 2012). The
absorption is caused by electronic processes (for more details, see Clark et al.
(1999)), and the resulting features in the spectra tend to be broad and shallow.
Typical minerals that can be detected from VNIR spectra include hematite
(Fe2 O3 ), geothite (FeO(OH)) and jarosite (KFe3 (OH)6 (SO4 )2 ).
Haavisto et al. (2006) reported an on-stream setup based on visible and
near infrared reflectance spectroscopy for elemental analysis of the froths and
slurries. They extended their work by fusion of XRF and reflectance spec-
troscopy measurement techniques (Haavisto and Hyötyniemi, 2009). Several
researchers have reported feasibility of quantitative mineralogical analysis us-
ing reflectance data obtained from visible and near infrared to shortwave in-
frared (VNIR and SWIR) wavelengths, see (Haest et al., 2012a) and (Haest
et al., 2012b).
2.3.5 Data fusion of slurry analyzers
Despite all advances and improvements of the current available online slurry
analyzers, there are still lacks of reliable methods enabling online mineral iden-
tification of the slurries. Considering the fact that the presence of minerals is
27
conditional to the presence of the constituent elements, extraction of miner-

alogical information from the current elemental data is a valid hypothesis and
this thesis seeks to prove the hypothesis.
LIBS and XRF are two elemental measurement techniques widely applied in
the flotation processes (Outotec, 2018). The interest in LIBS application for
slurry analysis has been recently increased ((Ayyalasomayajula et al., 2011) &
(Eseller et al., 2010)) due to multiple advantages over other analytical meth-
ods, e.g. it is simple and economic, it does not require sample preparation,
it can analyze multiple elements simultaneously, and broad range of elements
can be detected. Although LIBS technique provides a broad range of ele-
mental lines in each measured spectrum, conventional methods used for LIBS
data analyzing mainly consider only a few of the significant LIBS lines re-
lated to a few of the specific elements. The proposed method in this work
makes use of most of the important LIBS lines and is not restricted to a few of
the known elemental wavelengths. Results enable quantitative and continuous
mineralogical analysis of the slurry samples.
For online mineral estimation, a reference data of mineral contents is re-
quired. Therefore, the mineralogical assays of the selected samples are usually
provided by precise measurement techniques in laboratories. In Chapter 4,
it will be discussed that each measurement technique enables identification
of specific minerals which varies according to the instrumentation, the min-
eralogy of the samples, the grade and concentration of the present minerals.
It will be presented that linear combination of reflectance spectra with LIBS
spectra can lead to identification of some minerals which were not measured
by LIBS technique alone.
As discussed in Section 2.1.2, XRF technique has high resolution covering
a large area on the surface of drill core samples, and this feature applies to
slurries too. However, the main drawback of XRF measurement is that the
emission intensity of the elements decreases by decreasing atomic number.
Moreover, the weak emission of some elements are absorbed by water in the
slurry, protective windows and air in the spectrometer. Therefore, the spec-
tra generated by online XRF measurement of slurry samples were missing
many important elemental peaks and was therefore not suitable for mineral
identification. Another issue is that due to the slow sampling frequency of X-
ray fluorescence (long measurement time per sample), rapid oscillations and
sudden grade changes are not detectable. To solve this issue, Haavisto and
28
Hyötyniemi (2009) developed a soft-sensor approach predicting the elemental

contents of slurry using visual and near-infrared reflectance spectrum measure-
ments that can be obtained with high sampling frequency. They demonstrated
the applicability of VNIR spectroscopy for accurate assays especially for the
concentrate lines with high sphalerite and chalcopyrite contents and regular
grade variations. They used the XRF slurry elemental assays as the reference
measurements, and the rapid VNIR spectral assays were continuously cali-
brated with XRF results. Figure 2.8 represents a schematic diagram of the
slurry analyzer prototype developed by Haavisto (2009). A side flow from the
primary sample flow is continuously directed to the reflectance analyzer and
an additional slurry flow is sampled from the primary sample flow for XRF
analyzer.
When performing reflectance spectroscopy, solid content and particle size of
slurries influence the measured reflectance spectra. In laboratory conditions
when there is a small number of samples, problems like solid contents or par-
ticle size can be omitted since for small quantities of the samples they are
considered as constant when the calibration model is calculated. However, in
real on-stream applications both the solids content and particle size vary de-
pending on the feed ore properties and operating point of the process meaning
that a constant calibration model does not necessarily hold for longer periods
of time. Haavisto et al. (2008) reported a recursive PLS algorithm which it-
eratively updates the model generated from reflectance spectra with the XRF
assays to maintain a good prediction accuracy.
Some of the previous works could overcome some of the discussed problems
by combination of measurement techniques (Haavisto and Hyötyniemi, 2009).
This doctoral study focuses on the data integration of the slurry analyzers
rather than physical combination of the measurement setups. However, the
scheme presented in Figure 2.8 and introduced by Haavisto (2009) is applicable
to the measurement setups used in this doctoral study. This study extends
the research work conducted by previous research groups such as Haavisto and
Hyötyniemi (2009) or Haavisto et al. (2006) so that the focus is on the online
mineral identification of the slurry samples rather than elemental contents. An
efficient process control of iron ore concentrators is feasible when the grades
of important minerals such as hematite, magnetite or gangue minerals such as
quartz are available. This research proposes the data fusion of online slurry
analyzers for rapid and online identification of important minerals.
29
Primary
Process flows Process flows
sampling
Illumination Window Sample

cutter
Secondary
Spectrograph sampling
Jet
flow
cell
XRF analyzer
Air feed Webcam
Spectrum analyzer
Figure 2.8. Schematic diagram of slurry measurement prototype consisting of a

centralized XRF and an on-stream reflectance spectroscopy. This figure is
courtesy of doctoral dissertation published by Haavisto (2009)
30
3. Theoretical background
Data-driven approaches have recently drawn attention. Model-based system

analysis requires a prior knowledge about the processes. This knowledge is
provided by applying a mathematical model of the system used as a reference
for analyzing the new sampled data. Increase in the complexity of the sys-
tem causes the challenge of ensuring the accuracy of the extracted model. In
contrast with the model-based methods, data-driven approaches require only
the historical observations of the process performance recorded in a database.
There is no need for any accurate mathematical or physical model of the
system. Data-based methods are applied when the physical model of the sys-
tem is complicated or when the basic system operation principles are difficult
to model but there is enough monitored data available concerning the sys-
tem. Although data-based methods resolve most of the difficulties of applying
model-based methods, they bring us to the main challenge of need for large
quantities of good quality training data.
This thesis deals with the analysis of spectra generated by various mea-
surement techniques. For spectral analysis, linear multivariate techniques are
commonly applied. The spectra measured by detectors such as spectrometers
or spectrographs are usually high-dimensional, and linear multivariate meth-
ods are the most typical way of dealing with those spectra.
This chapter reviews and introduces the mathematical techniques and algo-
rithms used for data analysis. In Publication I and Publication II some meth-
ods of image processing have been applied. Raman spectra have been analyzed
using a mathematical optimization technique called non-negative least squares
(NNLS) and the technique will be introduced briefly in Section 3.2. Publica-
tion II contains an application of a supervised classification technique called
logistic regression. In Publication III a method for dimension reduction has
31
Theoretical background
been applied and qualitative mineral interpretation has been achieved using
a similarity search technique which employs the euclidean distance between
spectra. Publications IV and V use partial least squares (PLS) regression tech-
nique for online estimation of mineral contents along with non-linear iterative
partial least squares (NIPALS) algorithm for deriving the latent variables.
This chapter provides brief but comprehensive information about the applied
methods used in the research. Validation of regression models is the funda-
mental and extremely important step of spectral analysis and therefore will
be explained in details.
3.1 Digital images and measuring color
Color digital images are made of pixels and pixels are composed of channels
of primary colors. The most common primary colors used in many electronic
devices are red, green and blue. Each pixel of a digital image is the result
of adding the values of red, green and blue primary colors in different ways.
A digital image that is taken by a digital camera consists of red, green and
blue channels and is stored as an m-by-n-by-3 data array. Therefore, each
m-by-n matrix is called a matrix of primary channel and contains the numeric
values of red, green or blue primary colors. There is also gray image called
grayscale image made of only one channel of gray color. Therefore, RGB color
model is a space holding three channels of red, green and blue colors and each
pixel of a digital image results from concurrent representation of three pixels
in red, green and blue channels. One usual way to express the resolution of
an image is by pixel counting. For example, a resolution of m-by-n means
m pixel columns and n pixel rows of the digital image. Another resolution
term is radiometric resolution which defines how precisely an image displays
the differences of color intensities. Radiometric resolution is referred to as the
number of bits or levels. For instance, a 24-bit image is made of three channels
of 8-bits (unit8), and each pixel of a channel indicates the intensity value by a
value between 0 to 255. If the image is 48-bit, it means that the pixel of each
channel shows the intensity of the primary color using 16-bit (unit16) and is
called high-color depth image. Figure 3.1 represents RGB color cube of unit8
color images.
For mineral detection, RGB color space is not a proper space since three
32
Figure 3.1. RGB color cube for a uint8 image
independent matrices (red, green and blue) should be considered. There are
other color spaces containing components which ease the process of color de-
tection.
Hue is a significant property of color. It refers to pure color and is achieved
by linear transformation of RGB channels. The hue component of color is
widely applied in image segmentation, object recognition and color detection
(Finlayson and Süsstrunk, 2002). Hue-based representations of RGB images
can be achieved using color transformations like HSV (hue, saturation and
value), HLS (hue, lightness and saturation) and HIS (hue, intensity and satu-
ration) (Sangwine and Horne, 1998). These color spaces are linear transforms
of RGB space (Ford and Roberts, 1998). Figure 3.2 represents the HSV color
model mapped to a cylinder. To know the process of color space transfor-
mation, one way to calculate HSV values from RGB is represented below.
Saturation and Value components of HSV are easily derived from equations
3.1 and 3.2
max(R, G, B) − min(R, G, B)
S= , (3.1)
max(R, G, B)
V = max(R, G, B) . (3.2)
First, the following R , G and B values should be calculated
max(R, G, B) − R
R = (3.3)
max(R, G, B) − G
G = (3.4)
33
max(R, G, B) − B
B = , (3.5)
from where H is calculated according to one of the following conditions
H = 5 + B if R = max(R, G, B), G = min(R, G, B)
H = 1 − G if R = max(R, G, B), G = min(R, G, B)
H = R + 1 if G = max(R, G, B), B = min(R, G, B)
H = 3 − B if G = max(R, G, B), B = min(R, G, B)
H = 3 + G if B = max(R, G, B),
H = 5 − R otherwise
and H is then multiplied with 60 and hue is obtained. There are other ap-
proaches of color transformations. The detailed description and formula for
transforming the color spaces are presented by Ford and Roberts (1998).
Figure 3.2. HSV color cylinder
A color-based segmentation can be easily performed by defining threshold

ranges for hue values of the pixels. Therefore, hue component of color is
widely applied in analysis. However, sometimes irrelevant materials cannot be
distinguished using hue component of color. For instance, dusts on the surface
of the drill core samples occasionally fluoresce with colors in the same hue
ranges of the minerals but usually shinier or paler (see Figure 3.3). Therefore,
other transformations of color should be considered.
As the purpose is to perform identification by measuring the difference be-
tween colors, the CIELab color difference is suitable. The International Com-
34
mission on Illumination (abbreviated as CIE from its French title) defines an-
other color system that classifies color according to the HVS (the human visual
system) (Hunt and Pointer, 2011). CIE provides two color spaces, CIELab and
CIELuv. Although those two color spaces have been defined by CIE system
and are based on human vision, they are unintuitive. Lab is a non-linear but
reversible transformation of color. L represents the lightness of color, a and
b define redness-greenness and yellowness-blueness of the color respectively.
Equation 3.6 shows the formula of calculating the color difference in CIE sys-
tem. L, a and b are the CIELab coordinates of each pixel from the sample
(first RGB to Lab conversion is performed) and L* ,a* and b* are the standard
coordinates computed by averaging the L, a and b values. Identification of
materials using ΔE color difference is performed by setting a threshold so that
ΔE value for Lab coordinates of each pixel is calculated. If the value of ΔE is
greater than the threshold, it means that the measured color is significantly
different from the standard values.
ΔE = [(L − L∗ )2 + (a − a∗ )2 + (b − b∗ )2 ]0.5 (3.6)
3.2 Non-negative least squares
In the context of LIF-imaging, different minerals may fluoresce with similar

colors (Miller et al., 2005). Therefore, LIF analysis requires a reference which
links fluorescent colors to the minerals. Publications I and II introduce Raman
spectroscopy for calibrating and linking the fluorescent colors to the relevant
minerals. Non-negative least squares (NNLS) optimization algorithm was ap-
plied in Publication I as a novel technique for spectral analysis of Raman
spectroscopy. The NNLS method is basically a linear least squares approach
(a) (b) (c)
Figure 3.3. Refined image using ΔE value: (a) Separation of quartz pixels using hue
value (b) Pixels of dust on the surface of drill core (c) Pixels of pure quartz
35
but with constraints of having only positive values of coefficients. The applica-
tion of NNLS for Raman spectral analysis was introduced by Kauppinen et al.
(2013). The mineral identification using traditional point-counting techniques
assume that each spectrum is representative of only one mineral. However,
in this work NNLS was used to enable identification of multiple minerals in a
single spectrum.
The NNLS optimization algorithm is defined as follows. Vector d = (d1 , ..., dn )T
defines the coefficients of reference spectra ( 1,...,n) included into the solution
so that the linear combination of reference spectra builds the sample spectrum.
Vector x is the sample spectrum containing the measured Raman counts as a
function of wavenumber. The intensities of the reference spectra as a function
of wavenumber are included into the columns of matrix R. The NNLS problem
is therefore given as
mind FNNLS (d)
= mind ||xsample − Rref d||2 (3.7)
d≥0
where ||.|| denotes the L2 norm. The purpose is to solve d which defines the
coefficients at which different reference spectra are summed up to form the
sample spectrum. Ultimately, the values of d denote the contribution of each
reference spectrum for building the sample spectrum, and this can be used for
identification of multiple minerals per measured spectrum.
3.3 Logistic regression
Classification or categorization of data is divided into supervised and unsu-

pervised learning. Supervised learning is based on making a model using some
historical observations and estimating the outcomes of new events. Unsuper-
vised learning is based on inferring a hidden model from some unlabeled data.
The main difference between the two approaches is the labeling of data which
means that for supervised learning there exist some training examples with
known outcomes and those outcomes are used as label for past events.
Consider x as an input variable or feature and y denotes the output or a
target variable. The pair (xi , y i ) represents a training point and the whole
training set of m data points is represented by: {(xi , y i ); i = 1, ..., m} . X and
36
Y are used to denote the whole space of inputs and outputs respectively.
Figure 3.4 represents a block diagram of the supervised learning algorithms.
The purpose is to find a predictor function, h, that estimates the output y i for
every given input xi . The function h is called hypothesis function in literature.
The parameters of function h are learned using the training data set.
Training
GDWDset
Learning
SDUDPHWHUVRI
PRGHOK
x h predicted y
Figure 3.4. Block diagram of supervised learning algorithm process
When the target is to predict a continuous y, the problem is regression and

when the output y consists of small number of discrete values the problem is
classification. Assume that we decide to predict y by a linear function of x
hθ (x) = θ0 + θ1 x1 + θ2 x2 + ... + θn xn (3.8)
where θ0 is intercept and considering x0 = 1, the notation for hypothesis

function can be written as

n
hθ (x) = θi xi = θT x (3.9)
i=0
where the parameters θ of hypothesis function are learned by minimizing the

following cost function E(θ):
1 m
E(θ) = (hθ (x(i) ) − y (i) )2 . (3.10)
2 i=1
The purpose is to use an algorithm which iteratively updates θ (parameters)
so that the cost function E is minimized. A common optimization algorithm
for logistic regression is gradient descent also known as steepest descent. Gra-
dient descent is a first-order iterative optimization algorithm which is based
37
on recursive updating of parameters in the direction of negative of gradient.

It starts by an initial guess for parameters and takes steps proportional to the
negative of first derivative of cost function
∂
θj := θj − α E(θ). (3.11)
∂θj
For simplification we calculate the derivative of cost function for only one
training point (xi , y i ) and to simplify further one training point is represented
by (x, y). In this way the sum can be neglected for the moment and the
derivative of cost function is
∂ ∂ 1
E(θ) = (hθ (x) − y)2
∂θj ∂θj 2
∂ n
= (hθ (x) − y). ( θk xk − y)
∂θj k=0
= (hθ (x) − y)xj .
Therefore, the update rule of parameters for each training point (xi , y i ) will
be
(i)
θj := θj + α(y (i) − hθ (x(i) ))xj (3.12)
where i is the index of input variables, j is the parameters index and the
notation ’:=’ indicates updating of a parameter (not equality). So far the
discussion was about learning parameters of a hypothesis function (h) which
predicts the outputs or solves a regression problem. When the problem is
about data classification, a simple case is a binary classification where only
classes of 0 and 1 are the outputs and the hypothesis function should be
modified so that the output of h leads to making a decision about class 0 or
1. In logistic regression the hypothesis function changes to a logistic function
e.g. a sigmoid function
1
hθ (x) = g(θT x) = . (3.13)
1 + e−θx
We can represent θT x by the term z. It should be noted that for all values
of z the output of logistic function is between 0 and 1 (see Figure 3.5). When
z is greater than zero, the output of logistic function is between 0.5 and 1 and
when z is less than 0, output of logistic function is between 0 and 0.5. This is
a significant property resulting in its broad usage for data classification.
38
0.9
0.8
0.7
0.6
J]
0.5
0.4
0.3
0.2
0.1
0
−5 −4 −3 −2 −1 0 1 2 3 4 5
Figure 3.5. Plot of sigmoid function
Given the properties of sigmoid function, the probability of being in class 1 or

0 can be written as
P (y = 1|x; θ) = hθ (x)
P (y = 0|x; θ) = 1 − hθ (x),
or in a compact way
p(y|x; θ) = (hθ (x))y (1 − hθ (x))1−y .
Assuming that the variables (X) are independent and identically distributed
(iid), the likelihood function of parameters is

m
m
i (i) )
L(θ) = p(y |X; θ) = p(y i |xi ; θ) = (hθ (xi ))y (1 − hθ (xi ))(1−y . (3.14)
i=1 i=1
In order to simplify the calculations, the logarithm of likelihood function is

commonly used. The updates of parameters will then be performed by maxi-
mizing the log likelihood

m
log L(θ) = y (i) log h(x(i) ) + (1 − y (i) )(log(1 − h(x(i) ))), (3.15)
i=1
Therefore, equation 3.12 will be changed to
θj := θj + α∇θj log L(θ), (3.16)
where ∇ is the gradient sign. It should be noted that the negative sign in
equation 3.11 has now been converted to positive sign because we are maxi-
mizing the the cost function (log likelihood) rather than minimizing. In fact
39
the update rule of parameters is now based on gradient ascent. Given the first
derivative of logistic function
g (z) = g(z)(1 − g(z)) , (3.17)
and applying it to equation 3.16, ultimately the update rule of parameters will
be
(i)
θj := θj + α(y (i) − hθ (x(i) ))xj . (3.18)
It is important to note that equation 3.18 is different from equation 3.12. In

equation 3.18 the hypothesis function (h(x)) is a non-linear function defined
by equation 3.13. The term α is learning rate and large values of that might
cause overshoot and not divergence of the optimization algorithm and small
values of that will significantly slow down the optimization process. For more
information refer to the lecture notes of Ng (2000). When there are large
number of features, there is the risk of over-fitting of the trained model. Over-
fitting happens when a model is trained by too many parameters to fit to the
data. To prevent the over-fitting, regularized logistic regression is applied.
Regularization is performed by adding a penalty term in the cost function
when updating the parameters. The cost function is defined by
−1
m
λ n
E(θ) = [ y (i) log h(x(i) ) + (1 − y (i) ) log(1 − h(x(i) ))] + θ2 , (3.19)
m i=1 2m j=1 j
where λ is the parameter of regularization. Large value of λ pushes majority of

the parameters (θ) to zero or close to zero and vice versa. There is a trade-off
between the size of λ and the quality of model. The regularization term is used
when there are large number of features and only few of them are important.
Ng (2004) provides a detailed description of the regularized regression.
3.4 Principal component analysis and singular value

decomposition
The process of transforming the data from a high-dimensional space to a space

of fewer dimensions is called dimension reduction. Principal component anal-
ysis known as PCA is one of the most common linear methods for dimension
reduction. PCA is a statistical technique that transforms a set of correlated
40
variables into a set of linearly uncorrelated variables called principal compo-

nents (Basilevsky, 2009).
The data are transformed so that the variance of the data in the direction of
the first principal component is maximum and the second principal component
introduces the second direction with second largest variation of data and so
on. However, the principal components should follow the constraint of being
orthogonal. Prior to implementing PCA, scaling of variables is important
because the features with larger scales dominate others and this will result in
erroneous conclusions. PCA method calculates the principal components by
eigenvalue decomposition of data covariance matrix
1 T
R= X X , (3.20)
n
where R is the covariance matrix and X is assumed to be an n by m matrix
of input variables. The covariance matrix can be decomposed to matrices of
eigenvectors and eigenvalues
R = ΘΛΘT , (3.21)
where the eigenvectors Θ = (θ1 |...|θm ) of the data covariance matrix are called
principal components or loadings and the projected data points to the direc-
tion of loading vectors Z = XΘ are called scores or latent variables which
are mutually uncorrelated. Λ is a diagonal matrix containing eigenvalues of
corresponding eigenvectors, Λ = diag{λ1 , ..., λm }. Λ is also interpreted as the
variance of each score vector zi indicating the importance of corresponding
principal component (θi ), λ1 >= λ2 >= ... >= λm . It has been assumed that
the data are mean-centered and variance scaled. Matrix X with rank d is then
a linear combination of loading vectors (principal components) weighted by
scores (transformed data (z))
X = z1 θ1T + z2 θ2T + ... + zd θdT = ZΘT . (3.22)
PCA provides many advantages. For example, the principal components

with largest corresponding eigenvalues include the data with largest signal to
noise ratio and principal components with small corresponding eigenvalues can
be interpreted as noise. This is one way of mitigating the noise from the data
so that the dimension of data is reduced by selecting only l<d PCs
X ≈ z1 θ1T + z2 θ2T + ... + zl θlT . (3.23)
41
It is important to note that X is an approximation of the original X because

after dropping (d − l) of eigenvectors, some information (noisy data) is lost.
Hyötyniemi (2001) provides a complete and detailed description of multivariate
regression techniques including different approaches of dimension reduction.
Eigenvectors of a data matrix are mutually orthogonal and this is a signif-
icant property of eigenvectors and a fundamental reason for using space of
eigenvectors as a proper space for data analysis. The steps of deriving eigen-
values and eigenvectors are explained in details by Strang and Borre (1997).
An alternative method for PCA is singular value decomposition known as
SVD. This technique enables decomposing every kind of matrix into orthogonal
components. SVD decomposes the matrix Xn×m into
Xn×m = Ξn×n Σn×m ΨTm×m (3.24)
where Ξn×n and ΨTm×m are unitary or orthogonal matrices and Σn×m is a
diagonal matrix with non-negative real numbers on the diagonal. There is
direct relation between PCA and SVD. Given matrix decomposition formula
of SVD defined by equation 3.24, we have
1 T 1 1 1
R= X X = ΨΣT ΞT ΞΣΨT = ΨΣT ΣΨT = Ψ ΣT ΣΨT
n n n n
⎡ ⎤⎛ ⎞
T T
⎢σ1 ⎥ ⎜ ψ1 ⎟ (3.25)
1⎢
⎢ ..
⎥⎜
⎥ ⎜ .. ⎟
⎟
= (ψ1 |...|ψm ) ⎢ . ⎥⎜ . ⎟
n⎢
⎣
⎥⎜
⎦⎝
⎟
⎠
T
σm T
ψm
where ΞT Ξ = In and ΨT Ψ = Im due to orthogonality. Ultimately, comparing

equation 3.25 with equation 3.21 provides the relation between SVD and PCA
1 2
θi = ψi , λi = σ , Z = ΞΣ. (3.26)
n i
Therefore, matrix Ψ is derived from eigenvalue decomposition of matrix
X T X and correspondingly it can be shown that matrix Ξ is derived from
eigenvalue decomposition of XX T . When m is larger than n for Xn×m , SVD
is computationally less expensive than PCA because the matrix of eigenvalues
(Σ) and the matrix of score vectors (Z = ΞΣ) are calculated by computing
only Ξ. Since XX T has less dimension than X T X, the computation of XX T
will be faster. This property of SVD has resulted in application of SVD more
than PCA.
42
3.5 Partial least squares regression
Partial least squares (PLS)(Wold, 1985) regression is a common technique in

chemometrics. PLS is the main regression approach used in publications IV,
V, VI of this research. PLS is a well-known statistical method introduced by
Wold (1985) and can be considered as the extension of principal component
regression (PCR). PCR is a suitable alternative for linear regression when
there are large number of predictor variables and those predictors are highly
correlated. The basic linear regression follows the form
Y = XF + E (3.27)
where E is the error term and F is the matrix of parameters to be identified

(equivalent with hypothesis function described by Figure 3.4). Using the least
squares solution, the formula for calculating matrix F is as follows
F = (X T X)−1 X T Y. (3.28)
When there is correlation between input variables the solution of multivari-

ate linear regression is usually not optimal because some variables do not
contain more information than others and those dimensions just increase the
complexity of model. In this case, PCR is introduced as a solution. PCR is
combination of PCA and linear regression. It solves the problem of collinearity
between input variables by first projecting the data into PCA subspace where
score variables are mutually uncorrelated. Given the dimensionality reduction
property of PCA, the dimension of new space can be even reduced. Then from
the new space it predicts the output using score variables. PCR performs the
regression by taking only the variance of input variables into account. How-
ever, when the purpose is to predict multiple outputs (e.g. concentration of
multiple minerals), the correlation between output variables might be impor-
tant to be considered ( e.g. presence of one mineral can affect concentration
of other minerals). Neither multivariate linear regression nor PCR can gener-
ate a model which considers the correlation between both input and output
variables. PLS is capable of overcoming the collinearity between independent
variables and it takes the correlation between both inputs and outputs into ac-
count. It is a useful technique when the target is to predict a set of dependent
variables from a large set of independent variables. The PLS model makes
43
a linear mapping, B , between the independent variables, X , and dependent

variables, Y
Y = XBpls + F (3.29)
where F is the residual term. The purpose is to compute matrix Bpls so that
the structure of both data matrices X and Y are taken into account. Both
matrices are interpreted as a product of scores and loadings, which are further
utilized in the model estimation
Y = U ΦT + G (3.30)
X = ZΘT + E (3.31)
where U and Z are the scores and Φ and Θ are the loading vectors of Y and
X respectively. PLS uses a subset of latent variables (l < m ) that we can
represent them by matrix Z to approximate the predictor data X. Therefore,
the scores Z are related to the predictor data with n × l weights specified in
matrix W ∗
Z = XW ∗ . (3.32)
Furthermore, it is assumed that the scores of X can be used as predictors of

Y
Y = Z ΦT + F (3.33)
which gives the original mapping as in equation 3.29
Y = XW ∗ ΦT + F = XBpls + F. (3.34)
It can be seen that equation 3.34 is equivalent with general structure of

multivariate linear regression represented by equation 3.27. So far, the PLS
regression was explained using eigenvalue-based approach. However, in prac-
tice the latent variables are computed by an iterative algorithm called non-
linear iterative partial least squares (NIPALS) (Geladi and Kowalski, 1986).
NIPALS is faster than eigenvalue-based computation of PLS and it computes
only the first few required latent variable dimensions rather than computing
the whole components of a high-dimensional structure. The idea behind NI-
PALS algorithm is to use a matrix of weights W as the projection of X so that
Z has the maximum correlation with U
44
1. Initialize Ui . Use one of the columns of Y randomly as initial guess of U
X T Ui
2. wi = (UiT Ui )
. Calculate X weight
3. Zi = Xwi . Calculate X score
Y T Zi
4. φi = (ZiT Zi )
. Calculate Y loading
Y T φi
5. Ui = (φT
. Calculate Y score
i φi )
6. repeat steps 2. - 5 until U is converged and there is no more change on

that.
7. deflate X and Y matrices: X ← X − Zi θiT , Y ← Y − Zi φTi
3.6 Model selection and validation
When training a model, it is important to avoid the overfitting of the model.

Overfitting happens when a complex model with large number of parameters
is trained to the dataset. In case of PLS model, the complexity is defined by
the number of latent variables in the PLS model. Selecting a large number of
latent variables causes the overfitting of the model. Therefore, the regulariza-
tion of the model is performed using the idea of early stopping which means
selecting the minimum number of latent variables suitable for modeling. Part
of the dataset is used as test set or validation data and the rest as training
data. Leave-one-out cross validation is a reliable method commonly used for
model selection. At each iteration of the algorithm, one spectrum is left out
for validation and the rest of spectra are employed for training. The algorithm
runs with the number of training data and the mean squared error (MSE) of
both training and validation data are calculated for each mineral separately as
a criterion for model selection. At the end, the average error of training data
(called training error) and validation data (called validation error) are calcu-
lated for each dependent variable (minerals in this work). When the training
error and validation error are plotted for each number of latent variables, we
45
see a graph like Figure 3.6. As the model complexity increases, the training
error goes to zero. However, the validation error follows an elbow shape, it
improves to a point and then becomes progressively worse. The optimal num-
ber of latent variables is selected when the validation error is around a local
(or global) minimum at the same time that training error starts having steady
decrease.
0.5
.FBOTRVBSFFSSPS
Training error
Test error
0.0
0 1 2
/VNCFSPGMBUFOUWBSJBCMFT
Figure 3.6. Model selection
Estimation error is simply defined by the difference between true and esti-
mated values
E = Y − XF (3.35)
where F is the matrix of trained parameters. Mean square error is defined by

the average of the estimation errors
1 T 1 n
M SE = E E= (y(n) − F T x(n))2 . (3.36)
n n i=1
The quality of a trained model is usually expressed by coefficient of deter-

mination or R2 value
(Y − XF )(Y − XF )T
R2 = 1 − (3.37)
(Y − Ȳ )(Y − Ȳ )T
where Ȳ denotes the mean value of output. Generally, R2 value explains the
relative amount of variance explained by the model.
46
4. Results and discussion
The results of this thesis include mainly two achievements 1- rapid mineral
identification from the surface of drill core samples, 2- online and quantitative
mineral identification of slurry streams of an iron ore concentrator. The core
topic of this thesis is data-fusion of different available measurement techniques
in order to improve the process of mineral identification. This work introduces
knowledge extraction by integration of spectra obtained by various measure-
ment techniques. There is no need for further instrumentation or investment
in device development.
4.1 Drill core surface scanning and mineral map generation
The purpose of Publication I was to evaluate the possibility of rapid abun-

dance calculation of fluorescent minerals on the surface of drill core samples.
The sample set under study are segments of rock drill core collected from Ke-
vitsa mine of FQM Kevitsa Mining Oy, Finland. Kevitsa deposit holds nickel,
copper, cobalt, platinum, palladium and gold. Kevitsa intrusion is comprised
of ultramafic olivine pyroxenites and peridotites in the northeast, gabbro in
the west and central sections and predominantly granophyre in the south.
This work proposed for the first time a combined Raman spectroscopy and
LIF-imaging technique for rapid surface scanning of the drill core samples.
It turned out that Raman spectroscopy provides precise mineral identifica-
tion but since it is very slow measurement technique and has week emissions
specially for iron-bearing minerals, it was used for calibration of LIF-imaging
technique.
Although LIF-imaging does not enable concurrent identification of broad
range of minerals, it enables identification of some fluorescent minerals impor-
47
Results and discussion
tant for geologists. The abundance of albite, magnesite, quartz and a group
of amphiboles were calculated rapidly. The merger of Raman spectroscopy
and LIF image analysis demonstrated the ability to detect and locate the
constituent minerals of rocks on the round-side surface of the drill core sam-
ples. The method facilitates direct scanning of the drill core without sample
preparation (except for cleaning of the drill core surface).
Publication II continued and further validated the approach presented in
publication I by testing the method on the new drill core samples. The new
collected drill core samples were three pieces of drill cores from Kevitsa mine,
each one two meters length from the depths of 67, 200 and 340 meters below the
ground surface. Actinolite, augite, dolomite, chalcopyrite, enstatite and albite
were the minerals that could be qualitatively identified with combined Raman
and LIF-imaging technique. The mineralogy of the drill core samples were
provided by the laboratory of Kevitsa mine and using X-ray diffraction (XRD)
technique. Considerable amount of albite was identified by LIF-imaging in drill
core sample 7517802 while it was not reported by the XRD measurement of
the samples. Existence of albite in Kevitsa ore samples had been confirmed
by geologists of Kevitsa mine. Consequently, it can be deduced that LIF-
imaging can give complementary information to other assaying techniques such
as XRD.
Generally, geologists are enthusiastic about a mineral map of the drill core
samples for logging purposes. Moreover, all the mines or geological survey
companies are interested to provide a comprehensive database of minerals
for future analysis. Therefore, the advantage of developing techniques like
combined LIF and Raman which lead to faster scanning of the samples and
occasionally provides complementary information is intuitive. An important
achievement of this work presented within Publication II was the generation
of a mineral map from the round side surface of the drill core samples. Such
mineral maps provide coordinates of the constituent minerals at different levels
below the ground surface. The surface scanning technique was automated
using a logistic regression technique for mineral classification.
4.1.1 Positives, negatives and significance of the work compared to

the state-of-the-art
There are several works reporting mineral exploration using surface scanning.
Chatterjee (2012) reports image processing techniques for classification of rock
48
types. However, their approach is case-specific for a limestone deposit and

is not a general approach applicable to different types of mineral deposits.
Kuhn et al. (2016) reports surface scanning of drillcore samples using LIBS
technique and they validate their approach by comparing their results with an
energy-dispersive X-ray fluorescence. However, LIBS technique was slow as it
is a point-wise measurement technique. Although X-ray fluorescence covered
larger area of the surface, both LIBS and XRF provide elemental information
and X-ray was used in laboratory conditions. Yi et al. (2017) introduces a
combined LIBS and LIF setup for measuring Pb element in soil plasma. To
eliminate the interference of the LIBS lines of other present elements, a LIF
excitation process was added to the LIBS system. The plasma was further
illuminated by a LIF system resulting in great enhancement of the LIBS tran-
sition lines. However, the technique was applied to identify only one element
(Pb) and was slow for large throughput analysis.
The main advantage of the combined LIF-imaging and Raman spectroscopy
compared to the other works is the high speed of scanning and analysis. More-
over, the technique is providing on-site mineralogical information which helps
geologists and improves drillcore sorting or rock classification. While many
laser-based scanning techniques have poor performance on the round-side sur-
face of the drillcore samples or they are susceptible when the surface is not
smooth, the proposed LIF-imaging with high energy of the laser and longer
exposure time of camera is less susceptible to the surface condition.
It is important to note that prior to LIF-imaging measurement, Raman spec-
troscopy should be performed sporadically on some samples to provide some
reference data for calibrating LIF-imaging. However, Raman spectroscopy has
a complex measurement technique and it is usually applied in laboratory con-
ditions. Therefore, calibration of LIF-imaging with Raman spectroscopy is
a slow measurement procedure. Additionally, LIF-imaging enables identifica-
tion of only fluorescent minerals and does not provide information about broad
range of minerals. Therefore, the mineralogy of the samples will influence the
usability of LIF-imaging technique.
49
4.2 Qualitative mineral identification using LIBS technique
Section 2.2.4 discussed the increasing trend of LIBS application in multiple

engineering fields. The wide range of the elements which can be detected by
LIBS and the simplicity of the measurement setup are the main reasons for
broad application of LIBS specially for fast scanning purposes. The current
phase of this research focused on the mineral interpretation from the elemental
contents of LIBS using statistical methods.
There are a variety of approaches in the literature for data processing of
LIBS spectra. The data processing techniques are divided into supervised and
unsupervised as explained in section 3.3. Conventional methods for LIBS data
processing are unsupervised learning techniques. Principal Component Anal-
ysis (PCA) is very commonly used for LIBS data processing. However, this
process of data analysis is usually performed off-line so that a large set of the
measured spectra are stored into a matrix and the data are clustered using
PCA technique. Then, each cluster is interpreted in terms of the present ele-
ments. The number of clusters, interpretation of clusters and their correlation
with the present elements are usually performed with uncertainties because
there is not any standard way for interpretation of the clusters (Rosen-Gooding
et al., 2014). Moreover, for rapid sorting of the drill core samples the process
of scanning should be fast and automated and this research work sought to
address this problem.
According to the report of Gaft et al. (2007), elemental contents of the
same samples usually have constant ratios of LIBS line intensities and this
fact initiated the idea of mineral interpretation from the available elemental
LIBS lines. Publication III proposed a novel application of singular value
decomposition (SVD) for automated scanning of the drill core samples. The
proposed approach was a supervised learning application of SVD. First, a
reference database of the present minerals was produced so that purred (or
high-grade) minerals of interest were measured by LIBS technique. For each
target mineral, an average of multiple measured spectra was evaluated and
the most representative, significant and repetitive peaks were selected as the
characteristic of the target mineral. This process of database generation was
considered as part of the calibration and was time-consuming. However, it
was done prior to automated scanning of the samples. The next step was to
evaluate the spectrum of each new measured pulse and to determine whether
50
it contained any of those minerals in the database.

The reference spectra were stored into a matrix. Every new measured spec-
trum was appended to the end of the matrix and the next step was to realize
whether the measured spectrum contains any of the target minerals. Using
SVD the dimension was reduced to three (it was sufficient for analysis and
enabled visualization too) and each reference and measured spectrum was
represented as a point in the three dimension space produced by SVD. When
a measured spectrum contains one or multiple minerals of the database of
reference spectra, the corresponding points of those spectra in the 3-D space
will place in the close vicinity of the measured spectrum (see Figure 4.1). This
approach was used to identify the presence of minerals at each measured point.
References
Measured Spectrum
Dolomite
20
15
Score 3
10 Quartz
0 Hematite Magnesite
-20
-10
30
20
Chalcopyrite 0
10 10
0 Score 2
Score 1
Figure 4.1. Scatter plot of the first three scores. Chalcopyrite is the nearest neighbor of
the measured spectrum
Using the discussed approach, publication III demonstrated how qualitative

mineral identification of LIBS technique can provide complementary informa-
tion for LIF-imaging. Chalcopyrite was a very important mineral for copper
identification in Kevitsa mine of Finland and this mineral was not responsive
by either of LIF-imaging and Raman spectroscopy. The proposed LIBS min-
eral identification enabled reliable detection of chalcopyrite on the surface of
the drillcore sample.
Another achievement reported by publication III was the detection of mul-
tiple minerals in one measurement. Due to the overlapping of the elemental
LIBS lines, usually each mineral is studied and identified separately by select-
ing the relevant LIBS lines (Gaft et al., 2007). However, the proposed method
in this work enabled multi-mineral identification within one spectrum. LIBS
along with the proposed mineral identification approach was applied on two
51
pieces of high- and low-grade hematite drill core samples. For the mine under
study, the criterion for separating the drill core samples was the abundance
of hematite compared to the amount of quartz. A high-grade ore had at least
60% of hematite and a low-grade ore contained less than 40% of hematite.
Hematite and quartz were the most abundant and dominant minerals on the
surface of the drill core samples and their grains were almost mixed. After
scanning the surface of the drillcore samples, the measured LIBS spectra con-
tained mixed elemental lines of quartz and hematite. Using the proposed data
processing technique, quartz and hematite were identified in all spectra and
this led to successful sorting of the drill core samples.

The older research works were performing spectral analysis mainly by selecting
relevant elemental LIBS lines in the spectra and identifying the presence of
each element by its LIBS lines or by using dimension reduction techniques
like PCA to achieve a general view about the collection of elements in the
clustered data ((Rosen-Gooding et al., 2014) & (Gaft et al., 2008) & (Gaft
et al., 2014) & (Haavisto et al., 2013a)) . The more recent works have reported
more of supervised learning techniques for quantitative LIBS spectral analysis
((Washburn, 2015) & (Kuhn et al., 2016) & (Zhang et al., 2015)). Despite
all advances and successful reports for LIBS spectral analysis, there is still
a lack of mineral identification from the elemental LIBS lines. Majority of
the works have focused on the accurate elemental identification techniques.
This research proposed a supervised learning technique for qualitative mineral
identification from the LIBS spectra. It exploits the information in elemental
LIBS lines and extends the knowledge about the presence of minerals.
The advantage of the proposed SVD-based LIBS analysis method compared
to the other reported approaches is the suitability for rapid surface scanning
and providing mineralogical knowledge for drill core sorting. The technique
also makes the identification of multiple minerals in one spectrum possible.
Moreover, preparing the reference spectra is performed by the same LIBS
setup and there is no need for measuring the samples in laboratories using
other techniques.
However, when applying a point-wise measurement technique like LIBS for
drill core sorting, it would be extremely important that the minerals are evenly
52
distributed so that scanning a small volume on the surface of the sample

can be representative of larger volume. Preparation of high-grade samples
for reference spectra can be a challenging and time-consuming process and
usually the knowledge of geologists is required. It is also important to note
that the proposed technique is qualitative identification of minerals which
enables abundance calculation on the surface. Therefore, the technique is not
suitable for mineral identification of low-grade samples with sparse distribution
of minerals. Additionally, the smoothness of samples surface will influence
the quality LIBS spectra. Rugged samples cause variations on the distance
between the hitting point on the surface and the laser source and this will
attenuate the quality of the measured spectra.
4.3 Quantitative mineral identification of tailing slurries using

LIBS technique
After drill core surface scanning and qualitative mineral identification, this
research focused on quantitative mineral identification specially in slurry flow
streams. Publication IV discusses the importance of online mineral identifica-
tion for optimized process control. The samples under study were 14 tailing
slurry samples collected from an iron ore concentrator. The process of separa-
tion was magnetic and gravity followed by a reverse cationic flotation of quartz.
The purpose was to measure the grade of valuable minerals prior to flotation in
order to bypass the flotation circuit if the grade is high enough. Moreover, the
flotation properties of quartz is different from other mixed silica, and therefore
the online measurement of Si-bearing minerals was advantageous.
Partial least squares (PLS) regression was the analytical technique used for
mineral identification. The training dataset for learning the PLS model was
provided by mineral liberation analyzer (MLA) method in laboratory. A draw-
back of LIBS technique was the inability to distinguish between hematite and
magnetite due to their identical elemental contents. Therefore, the collection
of them was measured as Fe-oxides. Quartz and Fe-oxides were the most
abundant minerals and their concentrations were estimated with coefficients
of determination 70% and 82% respectively. The PLS model for quartz and
Fe-oxides were trained with 2 latent variables which explains the robustness
of the model (very unlikely to be overfit). Biotite, glauconite, ferrorichterite
and stilpnomelane were other minerals that could be identified by PLS regres-
53
sion. However, the concentration of those minerals were less than quartz and
Fe-oxides, and therefore the trained models for them were more complex and
less robust.
For K-feldspar, augite, actinolite and a group of unknown minerals, the
training of PLS model failed. K-feldspar was justified by the fact that the
two major elemental LIBS lines of potassium occur at 766 nm and 769 nm
wavelengths and those two wavelengths were not in the detection range of the
applied spectrometer. The rest of the minerals were not identified due to their
very low concentrations .
The novelty in Publication IV is the feasibility to identify multiple minerals
in the accumulated LIBS spectra. NIPALS-based PLS algorithm generates a
model employing linear combination of the accumulated intensities per LIBS
wavelengths so that the spectral changes are correlated with the mineral con-
centrations. For implementation in real world applications, the model training
should be done carefully by assuring that model is not overfit (e.g. large num-
ber of latent variables).

On-stream measurement of the aqueous samples is challenging since particles

are moving, surface has ripples and humidity of water attenuates the plasma
temperature which causes weak signal-to-noise ratios. Several studies have
reported approaches to overcome the problems generated by wet or aqueous
samples. Eseller et al. (2010) presents LIBS application on dried slurry sam-
ples prepared on PVC coated slides. Ayyalasomayajula et al. (2011) converts
liquid measurement to almost solid-sample measurement and applies a new
spin-on-glass sampling method. However, those techniques are performed on
the dried samples which needs time-consuming sample preparations. Oh et al.
(2007) reports a circulating sampling system in order to enable direct mea-
surement of liquid samples. However, there was the problem of LIBS intensity
variations which results from the surface fluctuations. To enable in-situ and
on-stream analysis of slurry samples, a proper sampling method is required
to significantly improve sensitivity and repeatability. Publication IV of this
study applied the circulating sampling system discussed in section 2.3.5. The
slurry was scanned by the LIBS system while slurry was flowing through a ver-
tical surface and a fixed point on the surface was measured using LIBS system.
54
This approach reduced the surface fluctuations and made the measurements
more stable.
Each slurry sample was circulated into a LIBS measurement cell and sev-
eral LIBS spectra per slurry sample were produced. The spectra were then
accumulated and the process was repeated three times. Then, the three ac-
cumulated spectra were averaged and formed into a matrix of independent
variables for 14 samples. Considering the hypothesis that the noise of natural
processes and measurement setups have the characteristics of the white noise,
it was expected that the accumulation and averaging of the spectra should ma-
jorly reveal the elemental peaks and diminish the noisy magnitudes and the
results validated the hypothesis. Later, some other research works reported
almost similar approach for slurry sampling and measurement (Cheng et al.,
2017). However, they focused only on the elemental analysis, and the online
and quantitative mineral identification has not been reported.
This study successfully reported quantitative identification of the important
minerals for better process control of an iron-ore concentrator under study.
However, distinguishing some important minerals such as hematite and mag-
netite could not be achieved. Moreover, the identification of ferrorichterite
as another important iron-bearing mineral was less reliable than other iron-
bearing minerals (small coefficient of determination (69%) and larger number
of variables for the PLS model). These problems were addressed within the
continuation of this work and reported in Publication V. Application of LIBS
technique on slurry samples should be resistant to the humidity generated by
water. This problem is roughly resolvable by applying an air vacuum around
the measurement point.
4.4 Online mineral identification by data fusion of LIBS,

reflectance spectroscopy and XRF measurement techniques
Publication IV discussed online mineral identification of slurry samples using

LIBS technique. The main drawback of LIBS technique for mineral identifica-
tion was incapability to distinguish between minerals like magnetite (Fe3 O4 )
and hematite (Fe2 O3 ) due to their identical elemental contents.
Publication V is the extended work of Publication IV. It focuses on the
capabilities of different slurry analyzers widely applied in the industry such as
LIBS, reflectance spectroscopy and XRF. Each slurry analyzer provides online
55
identification of some important elements and minerals. The hypothesis of

publication V is that instead of physical combination of the measurement
techniques, linear combination of their measured spectra leads to faster and
more efficient identification of minerals.
One of the particularities of reflectance spectroscopy technique is the ability
to measure hematite so that the characteristic spectrum of hematite is well-
represented in the reflectance spectroscopy technique (magnetite is less repre-
sented than hematite),(Brown et al., 2006). This research work presents data
fusion of LIBS and reflectance spectroscopy and then discusses the data fusion
of reflectance spectroscopy and X-ray fluorescence (XRF) measurement tech-
niques operating on the same slurry samples. The results will show that such
data integrations enable on-stream and quantitative identification of slurry
mineral contents specially for hematite, magnetite, quartz and ferrorichterite
which are important minerals in iron ore beneficiation.
The samples set under study were the same samples used in Publication
IV. Table 4.1 shows the coefficients of determination for different minerals
and measurement techniques. Publication IV concluded that LIBS enabled
reliable estimation of total Fe content. Therefore, data fusion of LIBS and re-
flectance spectroscopy was evaluated. The data fusion of LIBS and reflectance
spectroscopy improved the estimation of all minerals compared to LIBS tech-
nique alone. Using reflectance spectra alone, hematite could be identified with
better coefficient of determination (73%) compared to the combined LIBS and
reflectance spectra (61%). However, the model of hematite using the combined
LIBS and reflectance spectra was selected more reliably (model was trained
with less number of variables and therefore was more robust).
In this work XRF spectra of the measured slurry samples contained few
peaks mainly related to iron, and therefore the on-stream mineralogical iden-
tification of slurries using XRF alone was not feasible. The two XRF iron
lines were appended to the measured reflectance spectra. Clear PLS models
for hematite, magnetite and quartz could be trained. It was demonstrated that
the combined XRF and reflectance spectroscopy technique was outperforming
other techniques for hematite, magnetite and quartz estimation, Table 4.1.
The results were justified so that XRF technique is able to accurately mea-
sure the total Fe content using the iron energy lines of XRF spectra around
6.40 keV and 7.06 keV. Magnetite is strongly correlated with total iron content.
On the other hand, reflectance spectroscopy enabled accurate measuring of
56
hematite contents. Therefore, the linear combination of XRF and reflectance

spectra was expected to estimate the hematite and magnetite contents accu-
rately. The concentration of minerals assayed by MLA technique in laboratory
showed that quartz and total iron contents are highly correlated. Therefore,
the proposed analysis certainly benefits from the strong inverse correlation
between iron and quartz.
Table 4.1. Summary of the results including the R-squared values of minerals content
estimations
Mineral LIBS LIBS+reflectance Reflectance XRF+reflectance
R2 (%) R2 (%) R2 (%) R2 (%)
Quartz 69 88 85 95
Magnetite 78 92 87 94
Hematite 23 61 73 85
Ferrorichterite 52 83 79 59
It should be noted that LIBS technique is able to identify more minerals than
XRF as it provides broader range wavelengths. Since ferrorichterite contains
light elements and it was discussed in Section 2.2 that LIBS has better perfor-
mance than XRF when measuring elements lighter than titanium, therefore
combined LIBS and reflectance spectra enabled more accurate measuring of
ferrorichterite than other techniques .
Generally, the selection of suitable analyzers or the right fusion of the mea-
surement techniques should be performed based on the mineralogy of the slurry
samples and available analyzers in the plant. Publication V presented some
potentials of different measurement techniques and advantages of their data
fusion.

It was discussed in multiple sections of this report that the main novelty of
this research study is developing data analysis and measurement techniques
enabling online identification of minerals. There are not many research works
reporting online identification of minerals. McMillan et al. (2007) presents
qualitative mineral identification of carbonate and silicate families using cor-
57
relation coefficients between LIBS spectra of similar minerals. However, they

reported miss-identification of similar minerals due to their similar elemen-
tal contents such as Ca-Mg carbonates and Ca-Mg silicates. Moreover, their
approach enabled only rough estimations about presence of some minerals.
This research work could overcome such problems using spectral fusion of
reflectance spectroscopy with LIBS or XRF techniques.
Although there are some research works reporting feasibility of fast and
quantitative mineral analysis using reflectance spectra obtained from visible
and near infrared to shortwave infrared (VNIR and SWIR) wavelengths (see
(Haest et al., 2012a) and (Haest et al., 2012b)), their approach was accom-
plished and tested for solid samples and not for slurries. This research work
exploits the information provided by reflectance spectroscopy alongside LIBS
and XRF spectra and provides valuable information about present minerals
in the slurries.
To the best knowledge of the author, insufficient amount work has been
taken place for the data fusion of different measurement techniques or for
the purpose of online mineral identification. There are outdated works such
as Wiens et al. (2002) reporting application of reflectance spectroscopy and
laser-induced breakdown spectroscopy together. However, their approach is
a non-automated or offline comparison of results in order to acquire comple-
mentary mineralogical information. Previous works of our research group such
as (Haavisto and Hyötyniemi, 2009) have presented VNIR reflectance spec-
trum measurements as a supplemental assaying method for XRF analyzers.
However, the focus has been on the elemental analysis of the slurry samples
and not minerals. What distinguishes this work from other works is the way
that spectra of each measurement technique are pre-processed and combined
with the spectra of other measurement techniques for mineral identification.
Instead of separate analysis of spectra and comparing the results to gain sup-
plementary information, we combine all the spectra and train a model which
generates a linear combination of all variables. However, the spectra should
be pre-processed properly so that all variables have similar importance (See
Publication V).
It is important to mention that the proposed data fusion techniques in this
research provided reliable estimations of abundant minerals and failed to iden-
tify low-concentration minerals.
The measurement issue related to reflectance spectroscopy is the suscepti-
58
bility to the particle size and solids content of the samples. The solids content
and particle size vary depending on the feed ore properties and operating point
of the process. For large throughput analysis, the reflectance spectroscopy re-
quires recalibration of the model on some time intervals.
When XRF is applied on slurry samples, water in the slurry, protective
windows and air in the spectrometer causes attenuation of the signal and
this results in missing some elemental peaks specially elements below Ca.
Furthermore, the measurement time per sample of XRF is long and this leads
to missing detection of sudden grade changes.
59
60
5. Conclusions
This thesis concentrated on the rapid mineral identification using data-based

analytical techniques. The main objective was to determine whether the data
of currently available measurement techniques could be utilized to extract
mineralogical information. The target was qualitative and quantitative min-
eral identification with smallest amount of investment or modification of the
instruments.
The research started by surface scanning of the drill core samples in order to
make the process of drill core logging faster. LIF-imaging was calibrated with
Raman spectroscopy, and mineral maps from the round side surface of the drill
core samples were produced. The detected minerals, however, were limited to
fluorescent minerals. The process of drill core scanning was automated using
a logistic regression. Mineral groups were detected based on their pixel colors
in digital images. However, application of LIF-imaging for surface scanning
is dependent to the type of ore and the important fluorescent minerals to be
identified.
A novel application of singular value decomposition (SVD) was proposed for
mineral detection from LIBS elemental spectra. It was shown that multiple
minerals in one spectrum could be qualitatively identified. The approach pro-
vided detection of chalcopyrite on the surface of drill core samples while it was
not detected using combined LIF-imaging and Raman spectroscopy. Finally,
the proposed analytical approach could be applied for rapid sorting of hematite
drill core samples. However, the drill core samples should contain composite
minerals of interest so that small fraction of the sample is representative of
the large volume.
Quantitative mineral identification of slurry samples using the elemental
spectra of LIBS technique was demonstrated. Partial least squares (PLS)
61
Conclusions
regression was the analytical technique used for mineral identification and
NIPALS was the algorithm used for learning the PLS model. LIBS was un-
able to distinguish between hematite and magnetite due to their identical
elemental contents. Quartz and total iron-oxides (magnetite and hematite)
were estimated with good accuracy. Identifying multiple minerals from the
accumulated spectra of several measurements was a novel approach proposed.
Additionally, the proposed PLS-based technique enabled identification of other
minerals such as biotite, glauconite, ferrorichterite and stilpnomelane.
The research continued by suggesting that linear combination of LIBS or
iron channels of XRF with reflectance spectra could improve the accuracy of
mineral identification. The linear combination was performed using PLS re-
gression. Reflectance spectroscopy mainly detects hematite due to hematite’s
characteristic spectrum and LIBS could measure the total iron-oxides. The
linear combination of their spectra using PLS model led to the estimation
of magnetite, quartz, ferrorichterite and hematite. On the other hand, XRF
could measure the total Fe content accurately. It was shown that linear com-
bination of two XRF iron channels with reflectance spectra performs better
than other techniques. Finally, quartz correlated well with the total iron con-
tent, so that the accurate Fe measurement by XRF improved also the quartz
estimates significantly.
In a more general perspective, the results of this thesis show that there are
capacities for further knowledge extraction from the available data generated
by different measurement techniques. This thesis discusses some potential
applications of data fusion for online mineral analysis. The right data fusion
or selection of suitable analyzers should be done with respect to the type of
ore and mineralogy of the samples.
62
6. Future work
After discussing the results and achievements of this thesis work, some lines
of research are arising which are worth pursuing.
Firstly, the work on the surface scanning of the drill core samples using
LIF-imaging reported a qualitative identification of minerals from the round
side surface of the drill cores. Raman spectroscopy provided some reference
information about the present minerals and those references were used for
calibration of pixel colors in LIF digital images. It is beneficial to apply
analytical techniques in order to use the intensities of the color pixels as a
criteria for more quantitative analysis of the minerals. Moreover, Raman
spectroscopy has a complex measurement setup and it very slow for online
mineral identification. Data fusion of LIF-imaging with other rapid surface
scanning methods such as reflectance spectroscopy should provide faster and
more quantitative mineralogical analysis.
A second topic of research which follows Section 5.2, is an automated in-
tegration of information using advanced mathematical techniques. Publica-
tion III reported that LIBS technique could identify chalcopyrite which was
an important mineral for the mine under study, while it was not detected
by LIF-imaging or Raman spectroscopy. Although LIBS technique provided
complementary knowledge to LIF-imaging results, there is still a need to inte-
grate all the information in an automated way rather than just reporting the
acquired knowledge separately. Moreover, applying advanced probabilistic
methods may yield more conclusive results for the samples that their surface
are not representative of the whole volume.
It was discussed in Section 2.3.5 that reflectance spectroscopy measurement
is affected by the solids content and particle size of the slurry. Therefore, a
constant calibration model does not necessarily hold for long periods of time.
63
Future work
Consequently, the requirement of frequent calibration hinders the application

of reflectance spectroscopy as a stand-alone online measurement technique.
Following from Publications IV and V, one of the advantages of combined
LIBS and reflectance spectroscopy is the possibility to calibrate the model of
reflectance spectroscopy with the results of online LIBS. However, as hematite
was not identified by LIBS alone while it was an important mineral for the
iron ore concentrator, the calibration of reflectance spectroscopy was left for
the extended work. Probably more elemental lines of XRF could assist the
calibration of reflectance spectroscopy.
Finally, there exist plenty of information in different sectors of mines and
those information should be further exploited. However, those information,
such as knowledge of geologists, should be collected in a structured format.
Then, the information should be quantified and after that they can be ap-
plied alongside other measured spectra as new variables. There are significant
works going on the advanced probabilistic methods such as Bayesian infer-
ence. These techniques use other sources of information as prior knowledge
for better training of the models.
64
References
Adeyemi, A., Jarad, F., Pender, N., and Higham, S. (2006). Comparison of quantita-
tive light-induced fluorescence (QLF) and digital imaging applied for the detection
and quantification of staining and stain removal on teeth. Journal of Dentistry,
34(7):460–466.
Arkadiev, V., Knüpfer, W., and Langhoff, N. (2006). X-ray sources. In Beckhoff,
B., Langhoff, N., Kanngießer, B., Wedell, R., and Wolff, H., editors, Handbook of
Practical X-ray Fluorescence Analysis, pages 36–53. Springer.
Ayyalasomayajula, K. K., Dikshit, V., Yueh, F. Y., Singh, J. P., and Smith, L. T.
(2011). Quantitative analysis of slurry sample by laser-induced breakdown spec-
troscopy. Analytical and bioanalytical chemistry, 400(10):3315–3322.
Ayyalasomayajula, K. K., Yu-Yueh, F., Singh, J. P., McIntyre, D. L., and Jain, J.
(2012). Application of laser-induced breakdown spectroscopy for total carbon quan-
tification in soil samples. Applied Optics, 51(7):B149–B154.
Basilevsky, A. T. (2009). Statistical factor analysis and related methods: Theory and
applications, volume 418. John Wiley & Sons.
Beckhoff, B., Kanngießer, B., Langhoff, N., Wedell, R., and Wolff, H. (2007). Hand-
book of Practical X-ray Fluorescence Analysis. Springer Science & Business Media.
Brech, F. and Cross, L. (1962). Optical microemission stimulated by a ruby laser.

Appl. Spectrosc, 16(2):59.
Broicher, H. (1999). Ore and waste identification and quality control by means of
laser induced fluorescence. CIM Bulletin, 92(1034):59–63.
Brown, D. J., Shepherd, K. D., Walsh, M. G., Mays, M. D., and Reinsch, T. G.
(2006). Global soil characterization with VNIR diffuse reflectance spectroscopy.
Geoderma, 132(3):273–290.
Carr-Brion, K. (1989). X-ray analysers in process control. Kluwer Academic Pub.
Chatterjee, S. (2012). Vision-based rock-type classification of limestone using multi-

class support vector machine. Applied Intelligence, 39:14–27.
Chatterjee, S. (2013). Vision-based rock-type classification of limestone using multi-

class support vector machine. Applied Intelligence, 39(1):14–27.
65
References
Chen, D. and Plemmons, R. J. (2009). Nonnegativity constraints in numerical anal-

ysis. The Birth of Numerical Analysis, 10:109–140.
Cheng, X., Yang, X., Zhu, Z., Guo, L., Li, X., Lu, Y., and Zeng, X. (2017). On-stream
analysis of iron ore slurry using laser-induced breakdown spectroscopy. Applied
optics, 56(33):9144–9149.
Clark, R. N. et al. (1999). Spectroscopy of rocks and minerals, and principles of

spectroscopy. Manual of Remote Sensing, 3:3–58.
Cremers, D. A. and Knight, A. K. (2006). Laser-Induced Breakdown Spectroscopy.

Wiley Online Library.
Das, R. S. and Agrawal, Y. (2011). Raman spectroscopy: Recent advancements,

techniques and applications. Vibrational Spectroscopy, 57(2):163–176.
De La Vergne, J. (2003). Hard Rock Miner’s Handbook. McIntosh Engineering.
Death, D., Cunningham, A., and Pollard, L. (2008). Multi-element analysis of iron
ore pellets by laser-induced breakdown spectroscopy and principal components re-
gression. Spectrochimica Acta Part B: Atomic Spectroscopy, 63(7):763–769.
Díaz Pace, D. M., Gabriele, N. A., Garcimuño, M., D’Angelo, C. A., Bertuccelli,
G., and Bertuccelli, D. (2011). Analysis of minerals and rocks by laser-induced
breakdown spectroscopy. Spectroscopy Letters, 44(6):399–411.
Escárate, P., Bailo, D., Guesalaga, A., and Albertini, V. R. (2009). Energy disper-
sive X-ray diffraction spectroscopy for rapid estimation of calcite in copper ores.
Minerals Engineering, 22:566–571.
Eseller, K. E., Tripathi, M. M., Yueh, F.-Y., and Singh, J. P. (2010). Elemental
analysis of slurry samples with laser induced breakdown spectroscopy. Applied
optics, 49(13):C21–C26.
Filippov, L., Severov, V., and Filippova, I. (2014). An overview of the beneficiation of
iron ores via reverse cationic flotation. International Journal of Mineral Processing,
127:62–69.
Finlayson, G. and Süsstrunk, S. (2002). Optimization for hue constant RGB sensors.
In Proc. IS&T/SID 10th Colour Imaging Conference.
Ford, A. and Roberts, A. (1998). Colour space conversions. Westminster University,

London, 1998:1–31.
Fortes, F. and Laserna, J. (2010). The development of fieldable laser-induced break-

down spectrometer: No limits on the horizon. Spectrochimica Acta Part B, 65:975–
990.
Fricke-Begemann, C., Jander, P., Wotruba, H., and Gaastra, M. (2010). Laser-based
online analysis of minerals. ZKG International, 63(10):65–70.
Gaft, M., Nagli, L., Groisman, Y., and Barishnikov, A. (2014). Industrial online
raw materials analyzer based on laser-induced breakdown spectroscopy. Applied
spectroscopy, 68(9):1004–1015.
66
References
Gaft, M., Reisfeld, R., and Panczer, G. (2015). Modern Luminescence Spectroscopy
of Minerals and Materials. Springer.
Gaft, M., Reisfeld, R., Panczer, G., Blank, P., and Boulon, G. (1998). Laser-induced
time-resolved luminescence of minerals. Spectrochimica Acta Part A: Molecular and
Biomolecular Spectroscopy, 54(13):2163–2175.
Gaft, M., Reisfeld, R., Panczer, G., and Dimova, M. (2008). Time-resolved laser-
induced luminescence of UV–VIS emission of Nd 3+ in fluorite, scheelite and barite.
Journal of Alloys and Compounds, 451(1):56–61.
Gaft, M., Sapir-Sofer, I., Modiano, H., and Stana, R. (2007). Laser induced breakdown
spectroscopy for bulk minerals online analyses. Spectrochimica Acta Part B: Atomic
Spectroscopy, 62(12):1496–1503.
Geladi, P. and Kowalski, B. R. (1986). Partial least-squares regression: a tutorial.

Analytica Chimica Acta, 185:1–17.
Goncharov, A. F. and Struzhkin, V. V. (2003). Raman spectroscopy of metals, high-

temperature superconductors and related materials under high pressure. Journal
of Raman Spectroscopy, 34(7-8):532–548.
Gondal, M., Hussain, T., Yamani, Z., and Ahmed, Z. (2007). Determination of
toxic metals in petroleum, cultivated land and ore samples using laser-induced
breakdown spectroscopy. Bulletin of Environmental Contamination and Toxicology,
78(3-4):270–274.
Gorain, B., Oravainen, H., Allenius, H., Peaker, R., Weber, A., and Tracyzk, F.
(2007). Mechanical froth flotation cells. Froth Flotation: A Century of Innovation,
pages 637–680.
Grant, K. J., Paul, G. L., and O’Neill, J. A. (1991). Quantitative elemental analysis of
iron ore by laser-induced breakdown spectroscopy. Applied Spectroscopy, 45(4):701–
705.
Haavisto, O. (2009). Reflectance spectrum analysis of mineral flotation froths and

slurries. PhD thesis, School of Electrical Engineering, Aalto University, Finland.
Haavisto, O. and Hyötyniemi, H. (2009). Recursive multimodel partial least squares

estimation of mineral flotation slurry contents using optical reflectance spectra.
Analytica Chimica Acta, 642(1):102–109.
Haavisto, O., Kaartinen, J., and Hyotyniemi, H. (2006). Optical spectrum based
estimation of grades in mineral flotation. In Industrial Technology, 2006. ICIT
2006. IEEE International Conference on, pages 2529–2534. IEEE.
Haavisto, O., Kaartinen, J., and Hyötyniemi, H. (2008). Optical spectrum based mea-
surement of flotation slurry contents. International Journal of Mineral Processing,
88(3):80–88.
Haavisto, O., Kauppinen, T., and Häkkänen, H. (2013a). Laser-induced breakdown

spectroscopy for rapid elemental analysis of drillcore. IFAC Proceedings Volumes,
46(16):87–91.
67
References
Haavisto, O., Kauppinen, T., and Khajehzadeh, N. (2013b). Rapid rock drillcore
analysis, page 6. Finnish Automation Association, Helsinki. VK: T40102.
Haest, M., Cudahy, T., Laukamp, C., and Gregory, S. (2012a). Quantitative min-
eralogy from infrared spectroscopic data. i. validation of mineral abundance and
composition scripts at the rocklea channel iron deposit in Western Australia. Eco-
nomic Geology, 107(2):209–228.
Haest, M., Cudahy, T., Laukamp, C., and Gregory, S. (2012b). Quantitative mineral-
ogy from infrared spectroscopic data. ii. three-dimensional mineralogical character-
ization of the Rocklea channel iron deposit, Western Australia. Economic Geology,
107(2):229–249.
Harmon, R. S., DeLucia, F. C., McManus, C. E., McMillan, N. J., Jenkins, T. F.,
Walsh, M. E., and Miziolek, A. (2006). Laser-induced breakdown spectroscopy–
an emerging chemical sensor technology for real-time field-portable, geochemical,
mineralogical, and environmental applications. Applied Geochemistry, 21(5):730–
747.
Hunt, R. W. G. and Pointer, M. R. (2011). Measuring colour. John Wiley & Sons.
Huntington, J., Whitbourn, L., Mason, P., Berman, M., and Schodlok, M. C. (2006).
Hylogging - voluminous industrial-scale reflectance spectroscopy of the earth’s sub-
surface. In Proceedings of ASD and IEEE GRS; Art, Science and Applications of
Reflectance Spectroscopy Symposium, volume 2.
Hyötyniemi, H. (2001). Multivariate regression: Techniques and tools. Helsinki Uni-

versity of Technology.
Kauppinen, T., Haavisto, O., and Häkkänen, H. (2013). Optimisation algorithms in

the case of mineral detection using Raman analysis. IFAC Proceedings Volumes,
46(16):40–44.
Keeling, J., Mauger, A., and Huntington, J. (2004). Spectral core logger update
premilinary results from the barns gold prospect. MESA Journal, 33:32–36.
Khater, M. A. (2013). Laser-induced breakdown spectroscopy for light elements de-

tection in steel: state of the art. Spectrochimica Acta Part B: Atomic Spectroscopy,
81:1–10.
Kinsey, J. L. (1977). Laser-induced fluorescence. Annual Review of Physical Chem-

istry, 28(1):349–372.
Kruse, F. A. (2012). Mapping surface mineralogy using imaging spectrometry. Geo-

morphology, 137(1):41–56.
Kuhn, K., Meima, J. A., Rammlmair, D., and Ohlendorf, C. (2016). Chemical map-
ping of mine waste drill cores with laser-induced breakdown spectroscopy (libs)
and energy dispersive x-ray fluorescence (edxrf) for mineral resource exploration.
Journal of Geochemical Exploration, 161:72–84.
Lamprecht, G., Human, H., and Snyman, L. (2007). Detection of diamond in ore using
pulsed laser Raman spectroscopy. International Journal of Mineral Processing,
84(1):262–273.
68
References
Li, C., Sun, H., Bai, J., and Li, L. (2010). Innovative methodology for comprehensive
utilization of iron ore tailings: Part 1. the recovery of iron from iron ore tailings
using magnetic separation after magnetizing roasting. Journal of Hazardous Ma-
terials, 174(1):71–77.
McCanta, M., Dobosh, P., Dyar, M., and Newsom, H. (2013). Testing the veracity
of libs analyses on mars using the libssim program. Planetary and Space Science,
81:48–54.
McManus, C. E., McMillan, N. J., Harmon, R. S., Whitmore, R. C., De Lucia Jr,
F. C., and Miziolek, A. W. (2008). Use of laser induced breakdown spectroscopy
in the determination of gem provenance: beryls. Applied Optics, 47(31):G72–G79.
McMillan, N. J., Harmon, R. S., De Lucia, F. C., and Miziolek, A. M. (2007). Laser-
induced breakdown spectroscopy analysis of minerals: carbonates and silicates.
Spectrochimica Acta Part B: Atomic Spectroscopy, 62(12):1528–1536.
Miller, L. M., Ruppel, M. E., Ott, C. H., Smith, R. J., and Lanzirotti, A. (2005).
Development and applications of an epifluorescence module for synchrotron X-ray
fluorescence microprobe imaging. Review of Scientific Instruments, 76.
Ng, A. (2000). Cs229 lecture notes. CS229 Lecture Notes, 1(1):1–3.
Ng, A. Y. (2004). Feature selection, L 1 vs. L 2 regularization, and rotational in-

variance. In Proceedings of the Twenty-First International Conference on Machine
Learning, page 78. ACM.
Nienhaus, K. and Bayer, A. (2003). Innovative systems for horizon control of mining
machines by means of laser induced fluorescence (lif). In 12 th International Sym-
posium on Mine Planning and Equipment Selection (MPES), Kalgoorlie, Australia.
Oh, S. Y., Miller, T., Yueh, F. Y., and Singh, J. P. (2007). Comparative study of
laser-induced breakdown spectroscopy measurement using two slurry circulation
systems. Applied optics, 46(19):4020–4025.
Osborne, B. G., Fearn, T., Hindle, P. H., et al. (1993). Practical NIR spectroscopy
with applications in food and beverage analysis. Longman scientific and technical.
Outotec (accessed January 14, 2018). Outotec Courier

SL On-Line Analyzers, Brochure. Outotec Co. Finland.
http://www.outotec.com/globalassets/products/analyzers-and-
automation/ote_courier_8sl_online_analyzer_eng.pdf.
Rakovskỳ, J., Musset, O., Buoncristiani, J., Bichet, V., Monna, F., Neige, P., and
Veis, P. (2012). Testing a portable laser-induced breakdown spectroscopy system
on geological samples. Spectrochimica Acta Part B: Atomic Spectroscopy, 74:57–65.
Reisfeld, R., Gaft, M., Boulon, G., Panczer, C., and Jørgensen, C. (1996). Laser-
induced luminescence of rare-earth elements in natural fluor-apatites. Journal of
Luminescence, 69(5-6):343–353.
Rosen-Gooding, A., Ollila, A., Gordon, S., Newsom, H., Williams, A., Martinez,
R., Wiens, R., and Clegg, S. (2014). Laser-induced breakdown spectroscopy as a
tool to differentiate compositions of iron-bearing minerals. In Eighth International
Conference on Mars, volume 1791, page 1174.
69
References
Rosenwasser, S., Asimellis, G., Bromley, B., Hazlett, R., Martin, J., Pearce, T., and
Zigler, A. (2001). Development of a method for automated quantitative analysis of
ores using libs. Spectrochimica Acta Part B: Atomic Spectroscopy, 56(6):707–714.
Sangwine, J. and Horne, R. (1998). Colour Space Conversions. Chapman & Hall.
Scheinost, A., Chavernas, A., Barrón, V., and Torrent, J. (1998). Use and limitations
of second-derivative diffuse reflectance spectroscopy in the visible to near-infrared
range to identify and quantity fe oxide minerals in soils. Clays and Clay Minerals,
46(5):528–536.
Sharma, S. K., Lucey, P. G., Ghosh, M., Hubble, H. W., and Horton, K. A. (2003).
Stand-off Raman spectroscopic detection of minerals on planetary surfaces. Spec-
trochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 59(10):2391–
2407.
Strang, G. and Borre, K. (1997). Linear algebra, geodesy, and GPS. Siam.
Sun, Q., Tran, M., Smith, B., and Winefordner, J. (2000). Determination of mn
and si in iron ore by laser-induced plasma spectroscopy. Analytica Chimica Acta,
413(1):187–195.
Svoboda, J. and Fujita, T. (2003). Recent developments in magnetic methods of

material separation. Minerals Engineering, 16(9):785–792.
Tappert, M., Rivard, B., Giles, D., Tappert, R., and Mauger, A. (2011). Automated
drill core logging using visible and near-infrared reflectance spectroscopy: a case
study from the olympic dam iocg deposit, south australia. Economic Geology,
106:289–296.
Van der Meer, F. D., Van der Werff, H. M., van Ruitenbeek, F. J., Hecker, C. A.,
Bakker, W. H., Noomen, M. F., van der Meijde, M., Carranza, E. J. M., de Smeth,
J. B., and Woldai, T. (2012). Multi-and hyperspectral geologic remote sensing: A
review. International Journal of Applied Earth Observation and Geoinformation,
14(1):112–128.
Washburn, K. E. (2015). Rapid geochemical and mineralogical characterization of

shale by laser-induced breakdown spectroscopy. Organic Geochemistry, 83:114–
117.
Waychunas, G. A. (1988). Luminescence, X-ray emission and new spectroscopies.

Reviews in Mineralogy and Geochemistry, 18(1):639–698.
Wiens, R., Maurice, S., Lasue, J., Forni, O., Anderson, R., Clegg, S., Bender, S.,
Blaney, D., Barraclough, B., Cousin, A., et al. (2013). Pre-flight calibration and
initial data processing for the chemcam laser-induced breakdown spectroscopy in-
strument on the mars science laboratory rover. Spectrochimica Acta Part B: Atomic
Spectroscopy, 82:1–27.
Wiens, R. C., Arvidson, R. E., Cremers, D. A., Ferris, M. J., Blacic, J. D., Seelos,
F. P., and Deal, K. S. (2002). Combined remote mineralogical and elemental identi-
fication from rovers: Field and laboratory tests using reflectance and laser-induced
breakdown spectroscopy. Journal of Geophysical Research: Planets, 107(E11).
70
References
Wold, H. (1985). Partial least squares. Encyclopedia of Statistical Sciences.
Yi, R., Li, J., Yang, X., Zhou, R., Yu, H., Hao, Z., Guo, L., Li, X., Zeng, X.,
and Lu, Y. (2017). Spectral interference elimination in soil analysis using laser-
induced breakdown spectroscopy assisted by laser-induced fluorescence. Analytical
Chemistry, 89(4):2334–2337.
Zhang, T., Wu, S., Dong, J., Wei, J., Wang, K., Tang, H., Yang, X., and Li, H. (2015).
Quantitative and classification analysis of slag samples by laser induced breakdown
spectroscopy (libs) coupled with support vector machine (svm) and partial least
square (pls) methods. Journal of Analytical Atomic Spectrometry, 30(2):368–374.
71
References
72
Publication I
Kauppinen, T., Khajehzadeh, N. and Haavisto, O. Laser-induced fluores-

cence images and Raman spectroscopy studies on rapid scanning of rock
drill core samples. International Journal of Mineral Processing, 132, 26-33,
https://doi.org/10.1016/j.minpro.2014.09.003, November 2014.

c 2014 Elsevier.
Reprinted with permission.
73
International Journal of Mineral Processing 132 (2014) 26–33
Contents lists available at ScienceDirect
International Journal of Mineral Processing

journal homepage: www.elsevier.com/locate/ijminpro
Laser-induced fluorescence images and Raman spectroscopy studies on

rapid scanning of rock drillcore samples
Tommi Kauppinen ⁎,1,2, Navid Khajehzadeh ⁎,1,2, Olli Haavisto
Aalto University School of Electrical Engineering, Department of Automation and Systems Technology, P.O. Box 15500, 00076 AALTO, Finland
a r t i c l e i n f o a b s t r a c t
Article history: There exist many different ways of analysing rock drillcore samples. Here the interest is on the faster exploration
Received 11 November 2013 activities of mining. A capable, rapid and online measurement technique is presented for analysing the mineral
Accepted 7 September 2014 composition of rock drillcore samples. Raman analysis is an accurate but slow method of analysing rock drillcore
Available online 16 September 2014
samples. On the other hand, laser-induced fluorescence (LIF) can be used as a rapid method to produce a colour
map of minerals based on their fluorescence. These two methods can be combined by calibrating LIF measure-
Keywords:
Raman analysis
ments with Raman results. This paper presents a combination of Raman analysis with LIF image analysis to
Laser-induced fluorescence form a rapid online mineral exploration technique used on the rock drillcore samples. To the best knowledge
Drillcore of the authors, the combination of LIF and Raman in the study of rock drillcore has not been published. The sam-
Image processing ple set under study is segments of rock drillcore collected from a Ni–Cu multi-metal mine situated in Kevitsa,
Non-Negative Least Squares Finland. Scanning drillcore samples using LIF technique produces images which present a map of colours based
on emissions of fluorescent minerals. Some parts of the sample set, based on the exposed colours, are measured
through Raman spectroscopy and the minerals are identified using reference spectra from a public source. Iden-
tification of minerals through Raman analysis is done using Non-Negative Least Squares (NNLS) method. Results
show that colours of LIF images correspond to minerals identified through Raman analysis. The combination
of LIF and Raman is then used to produce mineral maps of the drillcore samples and finally an estimation of
the mineral abundance per units of drillcore area is developed.
© 2014 Elsevier B.V. All rights reserved.
1. Introduction Huntington et al., 2006; Tappert et al., 2011)) are the predominant
methods for drillcore analysis in the industrial field. There are other
Drillcore samples are the key source of information in mineral explo- spectroscopic and laser-based methods mostly applied in the laboratory
ration. Custom mineral exploration procedures are rather slow. Geolo- but with a high potential for industrial applications, e.g. Laser-Induced
gists log the drillcore samples manually by visual inspection, the Breakdown Spectroscopy (LIBS) which detects elemental contents of
samples are then halved and one half of them are sent to laboratories the samples (Fortes and Laserna, 2010), Raman spectroscopy for miner-
for further analytical measurements. Exact results are obtained in al detection already applied in the field (Kauppinen et al., 2013), X-ray
laboratories using slow but precise techniques. This process of sample Diffraction (XRD) applied for mineral analysis of the pulverised samples
preparation, drillcore shipment and laboratory analysis is rather expen- in the laboratory (XRD) (Escárate et al., 2009) and Time-Resolved Laser-
sive, needs a lot of workforce and may cause delays in receiving final Induced Fluorescence Spectroscopy (TR-LIFS) for quantitative and qual-
analysis results. itative interpretation of luminescent centres (Gaft et al., 2005).
Currently available drillcore scanners are used to measure the Although results of the spectroscopic methods are usually unfailing
elemental or mineralogical composition of the samples. X-ray Fluores- since they operate quantitatively or seek distinct patterns in data, the
cence (XRF) analysis which detects elemental contents of the sample process of measuring and data analysing is slow and far from the goal
(Arkadiev et al., 2006), visible and near-infrared (VNIR) and the short- of rapid analysis. On the other hand, geologists are enthusiastic about
wave infrared (SWIR) reflectance spectroscopies which detect indica- the abundance of the minerals as well as mineral map of the drillcore
tions of the minerals in the drillcore (see e.g. (Keeling et al., 2004; samples. The diameter of the radiation, e.g. laser beam, in most of the
point-wise measurement techniques (LIBS, XRD, Raman, TR-LIFS) is
about micrometres. Calculating the abundance of the minerals adequate
⁎ Corresponding author.
for industrial scales necessitates radiating substantial number of spots
E-mail addresses: tommi.kauppinen@aalto.fi (T. Kauppinen),
navid.khajehzadeh@aalto.fi (N. Khajehzadeh), olli.haavisto@alumni.aalto.fi (O. Haavisto).
to cover broader area of the drillcore surface. Compared to point-wise
1
These authors have contributed equally on the article. measurement techniques, XRF analysis enables relatively rapid scan-
2
All the authors can be reached through the same address. ning of the drillcore samples, but it explores the elemental contents of
http://dx.doi.org/10.1016/j.minpro.2014.09.003
0301-7516/© 2014 Elsevier B.V. All rights reserved.
T. Kauppinen et al. / International Journal of Mineral Processing 132 (2014) 26–33 27
Table 1
Drillcore samples measured with Raman and LIF from Kevitsa mine.
Drillcore Start (m) Length (m) No. of Raman LIF
KV386 L2 18.1 0.1 400 Yes

KV291 L4 26.05 0.25 200 Yes
KV357 L34 190.4 0.25 600 Yes
KV386 L16 90.0 0.2 250 No
KV157 L20 116.7 0.15 500 No
the rocks. VNIR and SWIR reflectance spectroscopy using line scanning
hyperspectral imaging systems is considerably rapid and whole visible
surface of the drillcore can be analysed (Haavisto et al., 2013). However,
there are still noticeable number of minerals not detectable by Infrared
Reflectance spectroscopy techniques.
Having discussed about the slowness of the spectroscopic tech-
niques, there is a need to develop new techniques which speed up the
process of rock and ore drillcore scanning. This paper aims to propose
a faster way to determine mineral contents of rock and ore samples. It
focuses on scanning of the round-side of the drillcore which is the
Fig. 2. Schematic diagram of the laser induced fluorescence (LIF) imaging set-up.
main sample received during exploration. A new method used for
rapid scanning of the drillcore which benefits from both the promptness
of laser induced fluorescence (LIF) imaging and advances of digital paper presents calibration of the error-prone but rapid LIF image analy-
image processing is presented. On the other hand, different minerals sis using the precise but slow Raman spectroscopy technique and even-
may fluoresce with the same colours under UV light. Moreover, similar tually shows how Raman can provide complementary information to
fluorescent minerals may fluoresce with dissimilar colours in different LIF analysis.
mines due to diverse activators (Miller et al., 2005). Consequently, LIF
analysis requires a reference which links colours to the minerals. This 2. Materials and methods
2.1. Laser-induced fluorescence
A luminescent mineral is composed of a lattice and luminescent

centres called activators. Different activators are the cause of different
colour radiations of the luminescent minerals (Gaft et al., 2005). Activa-
tors are translated as impurities in minerals and they are mainly transi-
tion metals or rare earth element ions (Waychunas, 1988).
Colourful emissions of the fluorescent minerals signify the idea of
mineral detection using those colours. Laser-induced fluorescence
(LIF) is a sensitive, moderately easy to implement and well-studied
measurement technique. Conventional methods of LIF analysis are
based on assessing the measured fluorescence spectrum of minerals
Fig. 3. Measurement no. 80 of the KV357 L34 drillcore. Sample spectrum (blue) and a
Fig. 1. Measurement apparatus for Raman measurements. linear combination of the reference spectra (red).
28 T. Kauppinen et al. / International Journal of Mineral Processing 132 (2014) 26–33
lakes. Chatterjee (2012) applies image processing techniques for classi-

fication of rock types. Adeyemi et al. (2006) presented a high correlation
between light-induced fluorescent images and digital images applied
for tooth stain removal.
2.2. Raman analysis
Raman analysis is a well-known method of mineral composition

analysis in ores and rocks. Materials have a characteristic Raman scat-
tering when excited by light. Measuring of this phenomenon is called
Raman spectroscopy (Das and Agrawal, 2011). Raman scattering causes
shifts in the wavelength of radiation from the material, and through
Raman analysis this can be used to deduce chemical contents of the ma-
terial. Raman analysis has been used in ore analysis in many parts of the
world, e.g. by Pinzaru and Onac (2009) and by De Biasio et al. (2012).
Here the Raman analysis is used on rock drillcore. The analysis is
based on optimisation algorithm called the Non-Negative Least Squares
(NNLS) method. The NNLS method is basically a linear least squares ap-
proach but only with positive values. Dörfer et al. (2010) criticize a
method of linear least squares in the context of Raman analysis if all
Fig. 4. Average of a hundred measurements of the KV357 L34 drillcore. Average sample
the real values are given as possible solutions. Here we argue that it is
spectrum (blue) and an average of the linear combination of the reference spectra (red). possible to use linear least squares approach if standard preprocessing
techniques and only positive values are accepted. The NNLS method of
analysis is further described by Kauppinen et al. (2013). The results of
the NNLS method give us a chance to review the widely accepted
that are excited by a UV laser. Extensive studies have been done on the point-count method (Haskin et al., 1997). NNLS method gives similar,
fluorescence spectrum of minerals in the laboratory conditions (see e.g. yet more accurate results.
(Gaft et al., 2008; Waychunas, 1988; Reisfeld et al., 1996)). Bozlee et al. Let us define our optimisation problem as follows. There exists a vec-
(2005) presented a remote LIF spectral analysis of natural minerals and tor c = (c1, …, cn)T, which defines the values at which the different ref-
rocks to provide complementary information to Raman spectroscopy. erence spectra (1,…,n) are included into the solution. The studied
Due to broad bandwidth of the fluorescent emission spectra of glow sample spectrum is given by a vector x of measured Raman counts as
centres (Nienhaus and Bayer, 2003), LIF spectra are mainly processed a function of wavenumber. The intensities of the reference spectra as a
qualitatively, e.g. searching for alterations like peaks and plateaux function of wavenumber are included into the columns of matrix R.
which are mineral specific (Broicher, 1999). Therefore, conventional The NNLS problem is therefore given as:
LIF analysis techniques are still point-wise, slow and eventually inap-
c 2
propriate for the idea of rapid scanning of the drillcore samples in the mincNNLS F NNLS ðcNNLS Þ ¼ mincNNLS jjxsample −Rref NNLSjj ; cNNLS ≥0; ð1Þ
industrial scale.
Advances in photography and digital image processing have resulted where ||. || denotes the L2 norm. It should also be noted that the main in-
in a wide usage of digital cameras in real world applications. In compar- terest here is to solve c which defines the coefficients at which different
ison with conventional spectroscopic techniques, digital imaging and reference spectra are summed up (to form the sample spectrum).
image processing have the advantages of being contactless, economical, Linear optimisation was used because some of the sample spectra
widely available and rapid (Adeyemi et al., 2006). Given high tenden- displayed a property of having multiple reference spectra merged to-
cies of mining companies on rapid scanning and analysing of the rock gether, in order to compose a linear combination of reference spectra
and ore samples, utilization of colour-based mineral detection through which better matched the given sample spectrum.
LIF images taken from the drillcore samples is acknowledged. Laser It should be noted that when discussing traditional point-count
induced fluorescent imaging and digital image processing have been ex- method, NNLS method identifies the mineral which is best presented
tended to various fields of study. Storrie-Lombardi and Sattler (2009) by given spectrum as follows. From every spectrum only the mineral
demonstrate laser induced fluorescent emission imaging as a non- which has the largest value in vector c is included into the point-
destructive detection of microbial life in the ice covers of Antarctic count. This effectively means that every spectrum can present only
Fig. 5. (a) Ordinary close-up photograph of a drillcore surface from Kevitsa mine, depth: 18 m, (b) laser-induced fluorescence (LIF) image of the same region.
(a) (b)
(c) (d)
Magnesite
Bluish White hue range

Albite
LIF
Red hue range
LIF
0 2 4 6 8 10 0 2 4 6 8 10
Raman Raman
p.c
p.c
0 2 4 6 8 10 0 2 4 6 8 10
Position (1 mm) Position (1 mm)
Fig. 6. (a) Drillcore sample from Kevitsa mine, Finland, depth 18 m. Specified region was irradiated by UV laser, (b) LIF image of the same region, Raman spectroscopy scanned 1 cm line on
the drillcore surface. Albite (red) and Magnesite (bluish white). (c) Position of Albite detected by Raman and LIF along the same scanned line,(d) position of Magnesite detected by Raman
and LIF along the same scanned line.
one mineral at the time. However, there is more information to be ex- The Raman measurements were performed with a home-built
tracted from the values of vector c than just the largest value. This infor- Raman spectrometer in a backscattering geometry. Samples were excit-
mation is to be discussed further in coming papers. Here it is only worth ed with laser at a wavelength of 532 nm (Changchun New Industries
to note that NNLS method is backward compatible with point-count Optoelectronics Tech. Co., Ltd, MLL-III-532 nm–200 mW). The excita-
method. tion light was attenuated to 100 mW and focused into a sample with
a microscope objective (Zeiss, 10×, 0.30 N.A.) and the backscattered
light was collected with the same objective.
2.3. Drillcore samples
Rayleigh scattering was attenuated with a 532 nm Razor Edge
ultrasteep long-pass edge filter (Semrock) allowing recording of
The sample set under study is segments of rock drillcore collected
Raman spectrum down to 100 (1/cm). Raman scattering was dispersed
from Kevitsa mine of FQM Kevitsa Mining Oy, Finland. Kevitsa deposit
with 1200 g/mm grating in a 0.5 m imaging spectrograph (Acton,
holds nickel, copper, cobalt, platinum, palladium and gold. Kevitsa in-
SpectraPro 2500i). The signal was detected with a CCD camera
trusion is comprised of ultramafic olivine pyroxenites and peridotites
(Andor, Newton DU940N-BV).
in the northeast, gabbro in the west and central sections and predomi-
nantly granophyre in the south (see Table 1).
2.5. LIF measurement setup
The samples studied with both LIF and Raman can be shortly de-
scribed as follows: KV386 L2 is disseminated copper, KV291 L4 is false
Laser-induced fluorescence measurements were done in the Univer-
ore and KV357 L34 is mafic tuff and barren.
sity of Jyvaskyla. The light source was a Krypton Fluoride Excimer laser
emitting at 248 nm wavelength. Laser power for LIF measurements was
2.4. Raman measurement setup 10 mJ/pulse at 20 Hz and the pulse length was 8 ns. Fig. 2 demonstrates
a schematic diagram of the LIF setup and the imaging system. The laser
Parts of the sample set are measured through Raman spectroscopy beam passes through a right-angle prism redirecting the beam to the
and the minerals are identified using reference spectra from a public surface of the specimen perpendicularly and the beam is then passed
source. In total, 1950 Raman spectra were analysed from Kevitsa mine. through a Plano-convex lens that causes slight dispersion of the laser
The actual drillcore samples and depths are presented in Table 1. beam on the surface of the sample. It enables analysis of larger surface
A laser operating at 532 nm and with a power of 100 mW was used area of the samples and considerably improves the pace of the measure-
for Raman measurements (see Fig. 1). The measurements were per- ment. Dispersing the laser beam might cause marked decreases in the
formed by translating the sample vertically in front of the measurement fluorescent intensity, hence the power of the laser should be amplified
area of the spectrograph. accordingly. Digital LIF images and ordinary close-up photographs of
(a) (b)
(c) Albite
(d)
Magnesite

LIF
Red hue range
LIF
0 1 2 3 4 5 6 7 8
0 1 2 3 4 5 6 7 8
Raman Raman
p.c
p.c
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
Fig. 7. (a) Drillcore sample from Kevitsa mine, Finland, depth 26 m. (b) LIF image of the specified region, Raman spectroscopy scanned 1 cm line on the drillcore surface. Albite (red) and
Magnesite (bluish-white). (c),(d) Positions of Albite and Magnesite detected by Raman and LIF along the same scanned line.
the drillcore surfaces were taken using a Sony DSLR-A390 digital cam- appearance like lightness or chroma. However, in this study another
era. For LIF images, the camera was set on manual mode (M) with a method in CIELAB colour space is employed to eliminate subsidiary col-
3.2 s exposure time and F-stop number of f/5.6. ours. CIELAB is a complete and three-dimensional colour space widely
applied for colour and object comparisons. Three axes of the CIELAB
2.6. Image analysis colour space represent lightness of the colour (L∗), redness–greenness
of the colour (a∗) and yellowness–blueness of the colour (b∗)
RGB colour model is a space holding three channels of red, green and (Schanda, 2007). In this work, CIELAB coordinates help to eliminate
blue colours. Colour specification in RGB space is well-known and colour the redundant pixels. An area containing solid colour of interest is
detection through RGB space entails inspecting three channels of data for selected and RGB values are recorded. These RGB values are then
each colour pixel. There are other colour spaces containing components converted to CIELAB coordinates and average value of them is produced.
which comfort the process of colour detection (Ford and Roberts, 1998). The average values are considered as the standard values. The total
Hue is a significant property of colour. It refers to pure colour and colour difference value is then calculated:
is an important perceptual correlate of colour. This attribute provides
h i
an easy way to describe a colour. Hue is widely applied in image 2 2 2 0:5
ΔE ¼ L−L þ a−a þ b−b ð2Þ
segmentation, object recognition and colour detection (Finlayson
and Susstrunk, 2002). Hue-based representations of RGB images can
be achieved using colour transformations like HSV, HLS and HIS L, a and b are the CIELAB coordinates of each pixel from the sample
(Sangwine and Horne, 1998). These colour spaces are linear transforms and L∗, a∗ and b∗ are the standard coordinates. Tolerances are set for
of RGB (Ford and Roberts, 1998). It is very straightforward to detect a the delta values so that the delta values out of the tolerance range indi-
colour using the hue component. Knowing hue value of the target col- cate significant difference of the sampled pixel and the standard values.
our, a tolerance is selected and colours are detected through specified It is important to note that ΔE is calculated for pixels only of the desired
hue ranges. The tolerance value can be selected either manually, by colour (target mineral) and subsidiary colour (e.g. dust) appeared in the
observing the histogram of the hue values or by assigning times of stan- same hue ranges and the rest of the colours are eliminated.
dard deviation of hue values which can be used to make an automated
process. 3. Results
In the context of mineral exploration using colour detection, some-
times hue component is not sufficient. For instance, dusts on the surface 3.1. Raman fitting
of the drillcore samples fluoresce, on occasion, with colours in the same
hue ranges of the minerals but usually shinier or duller. In such cases, Because some of the sample spectra displayed a property of having a
colours can be distinguished with other correlatives of colour multiple reference spectra merged together, optimisation was used to
(a) (b)
(c) (d)
Albite Quartz

LIF LIF
Red hue range
0 1 2 3 4 5 6 0 1 2 3 4 5 6
Raman Raman
p.c
p.c
0 1 2 3 4 5 6 0 1 2 3 4 5 6
Fig. 8. (a) Drillcore sample from Kevitsa mine, Finland, depth 190 m. Specified region was irradiated by UV laser, (b) LIF image of the same region, Raman spectroscopy scanned 6 mm line
on the drillcore surface. Albite (red) and Quartz (bluish-white). (c),(d) Positions of Albite and Quartz detected by Raman and LIF along the same scanned line.
find a linear combination of reference spectra which better matched the As shown in Fig. 3, the linear combination of the reference spectra
given sample spectrum. follows the sample spectrum. There are minor deviations around wave
Let us examine a case of combining the reference spectra to form a number 550 cm−1 and 600 cm−1, but the clear Quartz peak shown
single sample spectrum more closely. Fig. 3 represents a measurement around 450 cm−1 is followed by the linear combination.
from drillcore KV357 L34. Sample spectrum (blue) of measurement Fig. 4 demonstrates a more varied sample spectrum, because of the
no. 80 and a linear combination of reference spectra (red) are shown. averaging. There are many peaks visible, most notably a double peak
In Fig. 4, an average of one hundred measurements (blue) is given, as of Albite around 450 cm−1 and 500 cm−1, mixed with the single peak
well as an average of the linear combinations of reference spectra for of Quartz around 450 cm− 1. In addition there are two peaks related
these one hundred measurements (red). These one hundred measure- with Albite 250 cm−1 and a weak carbonate peak around 1100 cm−1.
ments have been conducted on specific area of the drillcore as a line When the individual measurements (100 pieces) are studied, every
scanning with a step size of 0.1 mm, resulting in a line scan of 1 cm. sample spectrum clearly presents one, two or three different minerals.
(a) (b) (c)
Fig. 9. Colour detection using hue: (a) Original LIF image, 9 cm2 of the drillcore surface, depth 26 m (b) Albite (c) Quartz.
(a) (b) (c)
Fig. 10. Refined image using ΔE value: (a) Masked image of Quartz (b) dust (c) pure Quartz.
This is due to fitting of reference spectra (see (Kauppinen et al., 2013)). Next the incorporation of Raman spectroscopy results into LIF im-
If we use point-count method, the mineral most clearly presented, or ages is presented. Fig. 6a shows one piece of the drillcore sample from
having the highest multiplier as a result of the optimisation, is the one a depth of 18 m below the ground level. Fig. 6b is a photograph of the
that is chosen to present the whole data in that specific point of mea- same area specified in Fig. 6a after being illuminated by UV laser.
surement. In following, we denote the results of point count method Raman method scanned 1 cm of the specimen. Each Raman spot size
as p.c. is 10 μm and each pixel of LIF image is about 15 μm. According to
It can also be argued why the NNLS method is used and not the more Raman results, the red colour was associated to Albite and the bluish-
simple point-count method. The answer is three-fold. First, the NNLS white colour to Magnesite (MgCO3). Fig. 6c is representing the position
method is fast enough to form the results needed for any experimental of Albite detected by Raman and the position of red pixels specified
setup on industrial or smaller scale. Second, NNLS method is backward based on their hue values in LIF image along the scanned line. Exposed
compatible with point-count method as we discussed previously. colours of the fluorescent minerals in the LIF images are noticeably fol-
Third, the results achieved through NNLS method, although here used lowing the Raman spectroscopy results since Albite has been detected
only for point-count, can be further analysed to present more varied by Raman technique at the same position in which associated LIF colour
Raman data on the samples measured. was revealed. Fig. 6d represents the same results for Magnesite and
It can be noted that the accuracy of this method is dependent on the similarly elucidates the correspondence between Raman results and
number of reference spectra. If a certain peak (or a combination of cer- LIF colours.
tain peaks) is not presented by the reference spectra, it cannot be fitted. Compatibility of Raman spectroscopy results and LIF analysis has
The reference spectra used were 77 in total and have been collected been investigated through multiple drillcore samples from various
from RRUFF database (Downs, 2006). depths. Fig. 7a to 7d represent the result of the same LIF and Raman
measurements performed on another piece of the drillcore sample
from a depth of 26 m below the ground level. Similarly the red and
3.2. LIF colours bluish-white colours were associated to Albite and Magnesite respec-
tively. The test results (like the previous results, Fig. 6) illustrate the
In this section we focus on the information that luminescence positions of Albite and Magnesite detected by Raman spectroscopy
exposes about the minerals. Fig. 5a is showing an ordinary close-up and seen in the LIF images and likewise confirm the compatibility of
photograph of one piece of the drillcore sample. Fig. 5b displays laser- Raman results and LIF colours.
induced fluorescence (LIF) image of the specified region on the surface Due to the diverse activators residing in minerals, similar minerals
of the drillcore (white rectangle in Fig. 5a). This simply implies that may fluoresce with dissimilar colours in different mines. Furthermore,
in the UV illumination, luminescent minerals show strongly coloured different minerals may have fluorescence with identical colours because
fluorescence and establishes the idea of mineral exploration using LIF of analogous activators. Given in Fig. 8, Raman and LIF measurements
colours. were performed on another piece of the drillcore sample from the
LIF technique needs to be calibrated using Raman spectroscopy. depth of 190 m below the ground level. The bluish-white and the red
Small-scale and sporadic Raman measurements are performed on the colours were attributed to Quartz and Albite respectively. According to
regions containing large quantities of distinct minerals based on their Fig. 6b and Fig. 7b, Magnesite fluoresced with the same bluish-white
exposed fluorescent colours. After scanning the drillcore samples with colour as Quartz. Moreover, both Quartz and Magnesite pixels are in
Raman spectroscopy, the dominant red colours were associated to Albite the same hue ranges, 0.5 to 0.8. Based on this, we can see the bearing
(NaAlSi3O8), green colours to anthophyllite ((Mg, Fe)7Si8O22(OH)2) and of Raman spectroscopy in the analysis of LIF images since LIF analysis
bluish white colours to Quartz (SiO2). technique is incapable of distinguishing between the minerals which
have the same colour. Fig. 8 is also denoting the adaptability of Raman
spectroscopy results with LIF images.
Table 2
Mineral abundance estimated from Figs. 6b, 7b and 8b.
3.3. Abundance
Mineral Abundance % Fluorescent
colour
Fig. 6b Fig. 7b Fig. 8b When calibrated with Raman spectroscopy, LIF technique contains a
Albite 49.0 22.5 29.0 Red reference linking colourful display with corresponding minerals. As
Amphibole 0.120 0.0700 0 Green discussed in Section 1, average hue values of the known areas of interest
Magnesite 4.00 15.0 0 Bluish-white in LIF images are calculated. Having an average value of the target
Quartz 0 0 11.0 Bluish-white
colour, a tolerance is specified in hue dimension and finally a mask is
designed for the desired colour. After masking LIF images on the target References
colours, the abundance can be easily estimated by:
Adeyemi, A., Jarad, F., Pender, N., Higham, S., 2006. Comparison of quantitative light-
induced fluorescence (qlf) and digital imaging applied for the detection and quanti-
fication of staining and stain removal on teeth. J. Dent. 24, 460–466.
#Colourful pixels of masked image
100 ð3Þ Arkadiev, V., Knüpfer, W., Langhoff, N., 2006. X-ray sources. In: Beckhoff, B., Langhoff, N.,
#Total pixels Kanngieer, B., Wedell, R., Wolff, H. (Eds.), Handbook of Practical X-ray Fluorescence
Analysis. Springer, pp. 36–53.
Bozlee, B.J., Misra, A.K., Sharma, S.K., Ingram, M., 2005. Remote Raman and fluorescence
studies of mineral samples. Spectrochim. Acta A 61, 2342–2348.
Fig. 9 exemplifies the idea of masking the original image. Using hue Broicher, H.F., 1999. Ore and waste identification and quality control by means of laser
component of the HSV colour space, original LIF image is masked for red induced fluorescence. CIM Bull. 92 (1034), 59–63.
and bluish-white colours associated with Albite (Fig. 9b) and Quartz Chatterjee, S., 2012. Vision-based rock-type classification of limestone using multi-class
support vector machine. Appl. Intell. 39, 14–27.
(Fig. 9c) respectively. The original LIF image covers 9 cm2 of the drillcore Das, R.S., Agrawal, Y., 2011. Raman spectroscopy: recent advancements, techniques and
surface. As discussed in Section 2.6, some minerals or dusts on the sur- applications. Vib. Spectrosc. 57, 163–176.
face of the drillcore may fluoresce with colours similar to the target min- De Biasio, M., Arnold, T., McGunnigle, G., Leitner, R., Tortschanoff, A., Fietz, N.,
Weitkämper, L., Balthasar, D., Rehrmann, V., 2012. Industrial raman mapping
erals and with the same hue ranges. Given in Fig. 10a, shiny spots of blue spectroscopy for mining applications. Proceedings of SPIE, p. 83740S.
colour mainly visible in the same regions of Quartz are most probably Dörfer, T., Schumacher, W., Tarcea, N., Schmitt, M., Popp, J., 2010. Quantitative mineral
dust. Using CIELAB colour space and total colour difference (ΔE) value, analysis using raman spectroscopy and chemometric techniques. J. Raman Spectrosc.
41, 684–689.
those secondary colours (dust) are eliminated (see Figs 10b and 10c).
Downs, R.T., 2006. The rruff project: an integrated study of the chemistry, crystallography,
After masking and refining the images (Figs. 9b and 10c), the abun- raman and infrared spectroscopy of minerals. Program and Abstracts of the 19th General
dance is estimated by counting the number of coloured pixels (see (3)). Meeting of the International Mineralogical Association in Kobe, Japan, pp. O03–O13.
Escárate, P., Bailo, D., Guesalaga, A., Albertini, V.R., 2009. Energy dispersive x-ray diffraction
Table 2 demonstrates the abundance percentage of the detected
spectroscopy for rapid estimation of calcite in copper ores. Miner. Eng. 22, 566–571.
minerals estimated from Figs. 6b, 7b and 8b. Finlayson, G., Susstrunk, S., 2002. Optimization for hue constant rgb sensors. Proc. IS&T/
SID 10th Colour Imaging Conference.
Ford, A., Roberts, A., 1998. Colour Space Conversions. Westminster Univ, London, U.K.
4. Conclusions Fortes, F., Laserna, J., 2010. The development of fieldable laser-induced breakdown spec-
trometer: no limits on the horizon. Spectrochim. Acta B 65, 975–990.
Gaft, M., Reisfeld, R., Panczer, G., 2005. Modern Luminescence Spectroscopy of Minerals
The merger of Raman spectroscopy and LIF image analysis demon- and Materials. Springer, Berlin; New York.
strated the ability to detect and locate the constituent minerals of Gaft, M., Reisfeld, R., Panczer, C., Dimova, M., 2008. Time-resolved laser induced lumines-
rocks on the round-side surface of the drillcore samples. This leads us cence luminescence of uv–vis emission of nd3+ in fluorite, scheelite and barite. J. Alloys
Compd. 451, 56–61.
to believe that combined Raman spectroscopy and LIF image analysis Haavisto, O., Kauppinen, T.K., Khajehzadeh, N., 2013. Rapid rock drillcore analysis.
is a promising way to provide mineral maps of the drillcore samples. Automaatio XX seminar, Helsinki: Finland.
Dispersing and amplifying the laser beam provides illumination of larg- Haskin, L.A., Wang, A., Rockow, K.M., Jolliff, B.L., Korotev, R.L., Viskupic, K.M., 1997. Raman
spectroscopy for mineral identification and quantification for in situ planetary surface
er areas on the drillcore surface for LIF analysis. Raman spectroscopy
analysis: a point count method. J. Geophys. Res. 102, 19293–19306.
produces a reference linking fluorescent colours to the related minerals. Huntington, J., Whitbourn, L., Mason, P., Berman, M., Schodlok, M.C., 2006. Hylogging —
This step is termed as calibration of LIF image analysis with Raman spec- voluminous industrial-scale reflectance spectroscopy of the earth subsurface.
troscopy. Combining LIF and Raman seems to enable scanning of rock Proceedings of ASD and IEEE GRS. Art, Science and Applications of Reflectance
Spectroscopy Symposium vol. 2.
drillcore in large scale. Compared to more traditional laser-based Kauppinen, T., Haavisto, O., Häkkänen, H., 2013. Optimisation algorithms in the case of
methods, faster measurement speeds can be achieved. The method mineral detection using Raman analysis. Proceedings of IFACMMM vol. 15, pp. 40–44.
facilitates direct scanning of the drillcore without sample preparation Keeling, J., Mauger, A., Huntington, J., 2004. Spectral core logger update premilinary
results from the barns gold prospect. Mesa J. 33, 32–36.
(except for cleaning of the drillcore surface). In the mining industry Miller, L.M., Ruppel, M.E., Ott, C.H., Smith, R.J., Lanzirotti, A., 2005. Development and appli-
fast and objective mineral analysis is important for exploration and for cations of an epifluorescence module for synchrotron X-ray fluorescence microprobe
grade control purposes. This paper proposes a light method giving imaging. Rev. Sci. Instrum. 76.
Nienhaus, K., Bayer, A.K., 2003. Innovative systems for horizon control of mining
rough estimates of the mineral abundance through LIF images taken machines by means of laser induced fluorescence (lif). Mine Planning and Equipment
from the surface of the drillcore samples. Selection. Kalgoorlie, WA, pp. 1–7.
The method for calculating the abundance is concise and uncompli- Pinzaru, S.C., Onac, B.P., 2009. Raman study of natural berlinite from a geological
phosphate deposit. Vib. Spectrosc. 49, 97–100.
cated. In order to have more genuine results practically, more automat- Reisfeld, R., Gaft, M., Boulon, G., Panczer, C., Jrgensen, C.K., 1996. Laser-induced lumines-
ed vision techniques and high-quality images are required. cence of rare-earth elements in natural fluorapatites. J. Lumin. 69, 343–353.
For typical industrial scale applications, the spatial resolution of Sangwine, J., Horne, R., 1998. Colour Space Conversions. Chapman & Hall.
Schanda, J., 2007. Colorimetry: Understanding the CIE System. Wiley, John & Sons,
Raman spectroscopy is unnecessarily high and speed is too slow. How-
Incorporated.
ever, since Raman is more accurate, it can be nicely used to calibrate the Storrie-Lombardi, M., Sattler, B., 2009. Laser-induced fluorescence emission (l.i.f.e.): in
faster but less accurate LIF image analysis method. situ nondestructive detection of microbial life in the ice covers of ant-arctic lakes.
Astrobiology 9, 460–466.
Tappert, M., Rivard, B., Giles, D., Tappert, R., Mauger, A., 2011. Automated drill core
Acknowledgements logging using visible and near-infrared reflectance spectroscopy: a case study from
the olympic dam iocg deposit, south australia. Econ. Geol. 106, 289–296.
Waychunas, G.A., 1988. Luminescence, X-ray emission and new spectroscopies. In:
This work is part of the Green Mining Programme funded by the Hawthorne, F.C. (Ed.), Spectroscopic Methods in Mineralogy and Geology, Reviews
Finnish Funding Agency for Technology and Innovation (Tekes) (Grant in Mineralogy. Mineralogical Society of America vol. 18, pp. 639–698.
Number: 2834/31/2011).
Publication II
Khajehzadeh, N., Kauppinen, T., Häkkänen, H. Laser-induced fluores-

cence imaging: detection of fluorescent minerals and estimation of abun-
dance. In International Mineral Processing Congress, www.gecaminpublications.com/imp
October 2014.

c 2014 IMPC.
83

"3&2*/%5$&% '-502&3$&/$& *."(*/( %&4&$4*0/ 0'

'-502&3$&/4.*/&2"-3"/% &34*."4*0/0' "#5/%"/$&
"6*% )"+&)8"%&)0..* "511*/&/
! ! !

&*,,*9,,9/&/
! ! "!" ! "!"

*96C62C62G2C:6EJ@7>6E9@5DDF:E23=6E@2?2=JD66=6>6?E2=2?5>:?6C2=@8:42=4@?E6?ED@7C@4<5C:==4@C6
D2>A=6D FCC6?E 56>2?5D 2C6 @? 72DE 6IA=@C2E:@? 2?5 @3;64E:G6 2?2=JD:D @7 @C6 D2>A=6D *9:D A2A6C
AC@A@D6D2 4@>3:?2E:@?@7"2D6C?5F465=F@C6D46?46"2?5:>286AC@46DD:?86?E:E=65":>28:?8
2D2 A@E6?E:2=>6E9@5 7@C2FE@>2E65>:?6C2=2?2=JD:D@7E965C:==4@C6D2>A=6D
E :D D9@H? E92E ":>28:?8 42? AC@5F46 2 >:?6C2= >2A @7 E96 7=F@C6D46?E >:?6C2=D 6I:DE:?8 @? E96
DFC7246 @7 E96 5C:== 4@C6 D2>A=6D *96 E6DE 42D6D 2C6 5C:== 4@C6 D2>A=6D 7C@> 2 $:F >:?6 D:EF2E65 :?
!6G:ED2 :?=2?5 *96 >2:? AC6C6BF:D:E6 @7 ":>28:?8 E649?:BF6 :D 2 C676C6?46 52E2 =:?<:?8 4@=@FCD E@
E96>:?6C2=D *9:D DEF5J 56D4C:36DE9642=:3C2E:@?@7":>28:?8 E649?:BF6 FD:?8AC64:D63FED=@H(2>2?
DA64EC@D4@AJ +D:?8 C68F=2C:K65=@8:DE:4C68C6DD:@?2>@56= :D @3E2:?65E92E4=2DD:7:6DE96>:?6C2=D32D65
@? E96:C 6>:EE65 7=F@C6D46?E 4@=@FCD :? E96 5:8:E2= :>286D *96 23F?52?46 @7 6249 >:?6C2= 8C@FA :D
6DE:>2E65 3J A:I6= 4@F?E:?8 *96 C6DF=ED 2C6 4@>A2C65 H:E9 .(2J :77C24E:@? =23@C2E@CJ 2DD2JD
A6C7@C>65@?E96D2>6D2>A=6D
*96C6DF=EDD9@HE92E E96 :?E6CAC6E2E:@?@7E967=F@C6D46?E>:?6C2=DE9C@F89":>286D 4@CC6DA@?5DE@E96
.( C6DF=ED FCE96C>@C6 E96 FE:=:K2E:@? @7 E96 =@8:DE:4 C68C6DD:@? E@ :?E6CAC6E E96 6I:DE:?8 7=F@C6D46?E
>:?6C2=D 8C62E=J:>AC@G6D E96 E649?:BF6

FCC6?E=J >:?:?8 2?5 6IA=@C2E:@? 4@CA@C2E:@?D 2C6 2G:5 7@C >@C6 C2A:5 2?2=JD:D E649?:BF6D @7 E96 @C6
D2>A=6D >AC@G652?572DE2?2=JD:D>6E9@5DC65F46E96E:>6@76IA=@C2E:@? 92G6 =6DD 64@=@8:42= :>A24E
2?54@?D6BF6?E=J564C62D6E964@DED
*C25:E:@?2= >:?6C2= 6IA=@C2E:@? 24E:G:E:6D 2C6 >@56C2E6=J D=@H C:== 4@C6 492C24E6C:K2E:@? AC@G:56D
:?7@C>2E:@?23@FEE966=6>6?E2=>:?6C2=@8:42= 2?586@A9JD:42=AC@A6CE:6D@7C@4<6=2,6C8?6
-96? G:DF2==J :?DA64E65 3J 86@=@8:DED 5C:== 4@C6 D2>A=6D 2C6 92=G65 2?5 @?6 92=7 @7 E96> :D D6?E E@
=23@C2E@C:6D 7@C 56E2:=65 2?2=JD6D I24E C6DF=ED 2C6 @3E2:?65 :? =23@C2E@C:6D E9C@F89 D=@H 3FE C6=:23=6
>62DFC6>6?E E649?:BF6D @C =23@C2E@CJ 2?2=JD6D D2>A=6D ?665 E@ 36 AC6A2C65 7:?6=J 8C@F?5 2?5
D@>6E:>6D E6DE65 :? G24FF> 4@?5:E:@?D *96 H9@=6 AC@46DD @7 D2>A=6 AC6A2C2E:@? D9:A>6?E 2?5
=23@C2E@CJ 6I2>:?6D :D C2E96C 6IA6?D:G6 =23@C:@FD 2?5 E65:@FD *96 7:?2= C6DF=ED @7 E96 =23@C2E@CJ >2J
E2<6 >@?E9D E@36C625J
%?=:?62?52FE@>2E65D42??:?8@75C:== 4@C63@5JFD:?872DE4@?E24E=6DD2?5?@?56DECF4E:G6E649?:BF6D
AC@G:56D 4@>A=6>6?E2CJ :?7@C>2E:@? 7@C =23@C2E@CJ 2?2=JD6D )42??:?8 E96 5C:== 4@C6 D2>A=6D 5:C64E=J
7C@>E965C:==4@C63@I6D EJA:42==J ?665D?@D2>A=6AC6A2C2E:@?6I46AE7@CDFC7246 4=62?:?8
*96C66I:DED6G6C2=2FE@>2E655C:==4@C6D42??:?8E649?:BF6D .C2J=F@C6D46?46.( C<25:6G6E2=
2?5 C67=64E2?46DA64EC@D4@AJ:?E96 G:D:3=62?5:?7C2C65 C2?86D D66!66=:?86E2= F?E:?8E@?
6E 2= *2AA6CE 6E 2= 2C6 E96 AC6G2:=:?8 D42??:?8 E649?:BF6D :? :?5FDECJ *96 7@C>6C :D
6=6>6?E2= 2?5 E96 =2EE6C :D >:?6C2=@8:42= 2?2=JD:D >6E9@5 ? 255:E:@? E@ .( 2?5 C67=64E2?46
DA64EC@D4@AJ E96C6 2C6 @E96C DA64EC@D4@A:4 2?5 =2D6C32D65 E649?:BF6D 2AA=:423=6 E@ 6=6>6?E2= 2?5
>:?6C2=@8:42= :?E6CAC6E2E:@? @7 E96 @C6 D2>A=6D DF49 2D "2D6C ?5F465 C62<5@H? )A64EC@D4@AJ ")
7@C 6=6>6?E2= 2?2=JD:D @CE6D 2?5 "2D6C?2 (2>2? DA64EC@D4@AJ 2D 2 A@H6C7F= E649?:BF6 7@C
>:?6C2= 2?2=JD:D.C2J :77C24E:@? .( 24@>>@?=23@C2E@CJ>6E9@5H:56=J2AA=:657@C>:?6C2=@8:42=
2?2=JD:D D4QC2E6 6E 2= 2?5 "2D6C?5F465 =F@C6D46?46 " 56E64E:?8 6I:DE:?8 7=F@C6D46?E
>:?6C2=D @7 E96 @C6 D2>A=6D 27E 6E 2= =E9@F89 E96 23@G6>6?E:@?65 E649?:BF6D 2C6 >2:?=J
<?@H? 2D=23@C2E@CJ >6E9@5DE96J92G69:89A@E6?E:2=E@362AA=:652D2FE@>2E65 2?5 C2A:55C:==4@C6
2?2=JD:D E649?:BF6D
#:?6C2= >2AD @7 E96 5C:== 4@C6 D2>A=6D 2D H6== 2D <?@H=6586 23@FE E96 23F?52?46 @7 >:?6C2=D 2C6
EC6>6?5@FD=J 96=A7F= 7@C 86@=@8:DED 7@C =@88:?8 AFCA@D6D @H6G6C @3E2:?:?8 @3;64E:G6 C6DF=ED
?646DD:E2E6D D42??:?8H:562C62D@7E965C:==4@C6DFC7246D *96C63JA@:?EH:D6>62DFC6>6?EE649?:BF6D
=:<6 ") (2>2? 2?5 4FDE@>2CJ " >6E9@5D 2C6 ?@E DF:E23=6 7@C DF49 72DE D42??:?8 AFCA@D6D :?
:?5FDEC:2= D42=6D @H6G6C 4@>3:?:?8 5:776C6?E 4@?E24E=6DD E649?:BF6D 42? =625 E@ 72DE6C 2?5 36EE6C
492C24E6C:K2E:@? @7 5C:== 4@C6 D2>A=6D C646?E DEF5J 92D D9@H? 9@H ") E649?:BF6 AC@G:56D
4@>A=6>6?E2CJ:?7@C>2E:@?7@C9JA6CDA64EC2=:>286D 22G:DE@6E2= @C E96C6 92D 366? 2 DEF5J @?
36EE6C2?2=JD:DA6C7@C>2?46D 249:6G23=6 3J4@>3:?:?8")2?5(2>2?)A64EC@D4@AJ E@86E96C )92C>26E
2=
4@>3:?65"2D6C?5F465=F@C6D46?46 2?5(2>2?DA64EC@D4@AJ >6E9@5 7@C>:?6C2= :56?E:7:42E:@? 92D
366?:?EC@5F465 3J 0!92;69K25696E2= 1 2?5E96 G:23:=:EJ @7E962AAC@249 92D366?566A=J5:D4FDD65
?E9:DDEF5J E964@>3:?65"2D6C?5F465=F@C6D46?462?5(2>2?DA64EC@D4@AJ6>A=@J:?82DFA6CG:D65
4=2DD:7:42E:@? >6E9@5 :?@C56C E@ 492C24E6C:K6 2?5 4=2DD:7J >:?6C2= 4@?E6?ED @7 E96 @C6 D2>A=6D 6?92?46D
E96E649?:BF6 #:?6C2=23F?52?46D @7E965C:==4@C6D2>A=6D @7 G2C:@FD56AE9D2C66DE:>2E65 :?2==J E96
677:4:6?4J@7E96 2AAC@249 :DG6C:7:65 3J4@>A2C:?8E96C6DF=EDH:E9 E96C6DF=ED @7 A@AF=2C .( >6E9@5
@3E2:?65 7C@> =23@C2E@CJ2DD2JD
!
2&".1-&3
*96AC6D6?EDEF5J>256FD6@7>6E6CD@7C@4<5C:==4@C6D2>A=6D4@==64E657C@>!6G:ED2>:?6:?=2?5
!6G:ED2 56A@D:E 9@=5D ?:4<6= 4@AA6C 4@32=E A=2E:?F> A2==25:F> 2?5 8@=5 *96 4@==64E65 5C:== 4@C6
D2>A=6D 2C6 E9C66A:646D@7 E96 5C:==4@C66249@?6:D EH@ >6E6CD7C@> E96 56AE9D@7 2?5 >6E6CD
36=@HE968C@F?5=6G6=
.&"352&.&/4"11"2"453
"2D6C:?5F4657=F@C6D46?46>62DFC6>6?EDH6C65@?6:?+?:G6CD:EJ@7 JGPD<J=2:?=2?5 *96=:89ED@FC46
H2D2!CJAE@?=F@C:56I4:>6C=2D6C6>:EE:?8 2E ?>H2G6=6?8E9 "2D6C 6?6C8J 7@C">62DFC6>6?ED
H2D > AF=D6 2E K 2?5 E96 AF=D6 =6?8E9 H2D ?D :8FC6 56>@?DEC2E6D E96 " D6EFA 2?5 E96
:>28:?8DJDE6> *96=2D6C362>A2DD6DE9C@F892C:89E2?8=6AC:D>C65:C64E:?8E96362>E@E96DFC7246@7
E96DA64:>6?A6CA6?5:4F=2C=J2?5E96362>:DE96?A2DD65E9C@F892&=2?@4@?G6I=6?DE92E42FD6DD=:89E
5:DA6CD:@?@7E96=2D6C362>@?E96DFC7246@7E96D2>A=6 E6?23=6D2?2=JD:D@7=2C86CDFC72462C62@7E96
D2>A=6D2?54@?D:56C23=J:>AC@G6DE96A246@7E96>62DFC6>6?E :DA6CD:?8E96=2D6C362>>:89E42FD6
>2C<65564C62D6D:?E967=F@C6D46?E:?E6?D:EJ96?46 E96 :?E6?D:EJ @7E96=2D6C 362> D9@F=536 6?92?465
244@C5:?8=J :8:E2=":>286D2?5@C5:?2CJ4=@D6FAA9@E@8C2A9D@7E965C:== 4@C6DFC7246DH6C6E2<6?
FD:?82)@?J)"( 5:8:E2=42>6C2 @C":>286DE9642>6C2H2DD6E@?>2?F2=>@56#H:E9
EH@ D64@?5D 6IA@DFC6E:>62?5DE@A?F>36C@77
*(52& ":>28:?8 2AA2C2EFD

*."(*/( $"-*#2"4*0/
=F>:?6D46?E>:?6C2=4@?D:DED@72=2EE:462?5=F>:?6D46?E46?E6CD 42==6524E:G2E@CD *96A96?@>6?@?@7
=F>:?6D46?46C6DF=ED7C@>6I:DE6?46@724E:G2E@CD:?E96>:?6C2=D 27E6E2= -96? :==F>:?2E653J
+,=:89E5:DD:>:=2C>:?6C2=D>2J 6>:E D:>:=2C7=F@C6D46?E 4@=@FCD %?E96@E96C 92?5D:>:=2C>:?6C2=D:?
5:776C6?E>:?6D42? 6>:E 5:DD:>:=2C 4@=@FCD D:?46 E96J>:89E92G6 5:776C6?E 24E:G2E@CD #:==6C6E2=
@?D6BF6?E=J":>28:?8 E649?:BF6?665DE@3642=:3C2E65 FD:?82C676C6?46 D@E92E 7=F@C6D46?E 4@=@FCD
2C6=:?<65 E@ E962DD@4:2E65>:?6C2=D ?E9:DDEF5J(2>2?DA64EC@D4@AJ2D2AC64:D6 3FED=@H>6E9@57@C
>:?6C2=56E64E:@?92D366?FD65E@ 42=:3C2E6 ":>28:?8 E649?:BF6
:8:E2= :>286D C6AC6D6?E:?8 7=F@C6D46?E >:?6C2=D2C6E96@FEAFE@7E96":>28:?8E649?:BF6 :8FC6 2
C6AC6D6?ED24=@D6FAA9@E@8C2A9@7@?6A:646 @7E96 5C:==4@C6D2>A=6 @7 E96 56AE9@7 >6E6CD 36=@H E96
8C@F?5=6G6=:==F>:?2E653J25:DA6CD65+,=2D6C362> :8FC6 3 D9@HD E96":>286@7E96:==F>:?2E65
2C62 >:IEFC6@7H9:E62?5C65 7=F@C6D46?E 4@=@FC :DG:D:3=6@?E96G6:?2?5?@E23=62>@F?E@752C<3=F6
4@=@FC :D@3D6CG23=6 DAC625 @? E96 DFC7246 @7 E96 5C:==4@C6 (2>2? DA64EC@D4@AJ D42?D 4> @7 E96 5C:==
4@C6 24C@DD E96 G6:? :8FC6 4 C6AC6D6?ED E96 4@?46?EC2E:@? @7 >:?6C2=D 2?5 E96:C A@D:E:@? DA64:7:65 3J
(2>2? E649?:BF6 :G6? (2>2? C6DF=ED E96 >:IEFC6 @7 C65 2?5 H9:E6 4@=@FC @? E96 G6:? :D 4@>3:?65
@=@>:E62?592=4@AJC:E62?5E9623F?52?E 52C<3=F6 4@=@FC :D4E:?@=:E6
4
*(52& " C:==4@C6D2>A=67C@>!6G:ED2>:?656AE9 > :CC25:2E653J+,=2D6C # ":>286@7E96
D2>6 D2>A=6(2>2?DA64EC@D4@AJD42??65 4>@?E965C:== 4@C6DFC7246 $ @?46?EC2E:@?@7 E96
7=F@C6D46?E>:?6C2=D >62DFC65 3J(2>2?
"."/1&$4203$017
(2>2? 2?2=JD:D:D2A@H6C7F=>6E9@57@C>:?6C2=56E64E:@?:?@C6D2>A=6D -96?>2E6C:2=D2C66I4:E653J
=:89E E96J D9@H 2 492C24E6C:DE:4 (2>2? D42EE6C:?8 *9:D 492C24E6C:DE:4 >2<6D D9:7ED :? E96 C25:2E:@?
H2G6=6?8E9D @7 E96 >2E6C:2=D 2?5 >62DFC:?8 E9:D A96?@>6?@? :D 42==65 (2>2? DA64EC@D4@AJ 2D 2?5
8C2H2= )2>A=6D >62DFC65 H:E9 (2>2? 2C6 :?E6CAC6E65 FD:?8 C676C6?46 DA64EC2 7C@> 2 AF3=:4
D@FC46 @H?D (2>2? 2?2=JD6D E96 D2>A=6D FD:?8 2 $@?$682E:G6 "62DE )BF2C6D $$")
@AE:>:K2E:@?2=8@C:E9> !2FAA:?6?6E2=
51&26*3&%$-"33*'*$"4*0/0' '-502&3$&/4.*/&2"-3
(68:@?D @7 E96 5C:== 4@C6 DFC7246 4@?E2:?:?8 =2C86 BF2?E:E:6D @7 5:DE:?4E >:?6C2=D 32D65 @? E96:C 6IA@D65
7=F@C6D46?E4@=@FCD 2C6D42??65 3J(2>2? >6E9@5 *96C67@C62 DF3D6E@7 A:I6=D 2C6=236==65 2D E96J2C6
=:?<65E@E96>:?6C2=D FD:?8(2>2? *9:DDF3D6E :D DF77:4:6?EE@ 36FD652D 2 EC2:?:?852E2D6E2?56DE:>2E6
E96 >:?6C2=D 2DD@4:2E65H:E9 E96 C6DE@7E96 4@=@FC A:I6=D ?@FC6IA6C:>6?E6249(2>2?DA@ED:K6:D
M>2?5 D2>A=6DH6C6 AC@5F465 3J(2>2?:?E@E2= 249A:I6=D:K6@7E96:>286:D2AAC@I:>2E6=J
M> *96C67@C6(2>2?=236==6523@FE @7E96A:I6=D
DFA6CG:D65 4=2DD:7:42E:@? 32D65 @? =@8:DE:4 C68C6DD:@? 92D 366? FD65 E@ 4=2DD:7J 7=F@C6D46?E >:?6C2=D
244@C5:?8E@E96:C 6>:EE65 7=F@C6D46?E 4@=@FCD :? E96 5:8:E2=:>286D *@AC6G6?E2?J @G6C7:EE:?8AC@3=6>
C68F=2C:K2E:@? 92D 366? FD65 ? C68F=2C:K2E:@? 2 A6?2=EJ E6C> :D 2AA=:65 =625:?8 DF> @7 E96 23D@=FE6
G2=F6D@7E96@AE:>:K2E:@?A2C2>6E6CDE@36D>2== ?5C6H
*96 EC2:?:?8 D6E :?4=F56D E9C66 5:>6?D:@?D D:?46 6249 A:I6= 4@?D:DED @7 E9C66 4@=@FC 4@>A@?6?ED (65
C66? 2?5 =F6 @C :>286 AC@46DD:?8 AFCA@D6D ( 4@=@FC DA246 :D EC2?D7@C>65 E@ ), 4@=@FC DA246
H96C6 ) 2?5 , 2C6 F6 )2EFC2E:@? 2?5 ,2=F6 4@>A@?6?ED @7 4@=@FC C6DA64E:G6=J @=@FC :>286
AC@46DD:?8FD:?8F632D65C6AC6D6?E2E:@?D@7:>286D:D>@C6DEC2:89E7@CH2C5E92?( :>286D @7
EC2:?:?8 52E2 AC@G:565 3J (2>2? 2C6 FD65 E@ =62C? A2C2>6E6CD 2?5 E96 C6>2:?:?8 2C6 FE:=:K65 2D
4C@DD G2=:52E:@? E@ G6C:7J E96 >@56= *96 EC2:?:?8 D6E 4@?D:DED @7 >@C6 E92? EH@ 4=2DD6D 2?5 6249 4=2DD
:?5:42E6D @?6 7=F@C6D46?E>:?6C2= *96C67@C6E96AC@3=6>E@36D@=G65:D2>F=E:4=2DD4=2DD:7:42E:@? %?6
G6CDFD2== :D 2 D:>A=6 3FE 25G2?E286@FD DEC2E68J C64FCC6?E=J 2AA=:65 E@ 4=2DD:7:42E:@? >6E9@5D %?6
G6CDFD2== D496>692D366?56D:8?6532D65 @?3:?2CJ 4=2DD:7:42E:@? @C28:G6?52E2D6EH:E9$5:776C6?E
4=2DD6D$5:776C6?E =@8:DE:4C68C6DD:@?4=2DD:7:6CD 2C6EC2:?65 *96? 7@C6249?6H:?AFEI$4=2DD:7:6CD2C6
2AA=:652?5E964=2DD92G:?8>2I:>F>G2=F6:DA:4<65
*96 8@2= @7 E9:D DEF5J :D E@ G6C:7J 2 ?6H >6E9@5 2AAC@AC:2E6 E@ C@F89=J 6DE:>2E6 E96 23F?52?46 @7 E96
7=F@C6D46?E>:?6C2=D7C@> E96 C@F?5D:56DFC7246@7E965C:==4@C6D2>A=6D 7E6C4=2DD:7J:?8 E96 >:?6C2=D
FD:?8=@8:DE:4C68C6DD:@? E965:8:E2=":>286:D>2D<65@? E96 E2C86E>:?6C2=4@=@FC2?5 E96 23F?52?46
42?3662D:=J6DE:>2E65 3J

:8FC6 56>@?DEC2E6D2?6I2>A=6@7 2 >:?6C2=>2A AC@5F4653J ":>28:?8E649?:BF6 7C@>E96C@F?5
D:56DFC7246@7E965C:==4@C6D2>A=6 92=4@AJC:E6 92D 366? 56E64E65 >:I65 H:E9 @=@>:E6 2?5 E96>:IEFC6
H2D C6AC6D6?E653J >286?E2 4@=@FC :?E96>:?6C2=>2A &FC6 4E:?@=:E6 E9623F?52?E>:?6C2= :DD9@H?
H:E9 8C66? 4@=@FC 2?5AFC65@=@>:E6H:E9C65
249 4=@D6FA A9@E@ H9:E6 C64E2?8=6 :? :8FC6 2 4@G6CD 4> 4> 2C62 @7 E96 5C:== 4@C6
DFC7246 :G6?E96>:?6C2=>2A":>28:?8E649?:BF6 86?6C2E6D 4@@C5:?2E6D @7E967=F@C6D46?E >:?6C2=D
2 3
*(52& #:?6C2=>2A86?6C2E653J":>28:?8
*23=6 :D C6AC6D6?E:?8 E96 23F?52?46 @7 E96 >:?6C2=D 42=4F=2E65 3J .( E649?:BF6 A6C7@C>65 @? E96
A@H56CD2>A=6D :?E96=23@C2E@CJ 249 .( >62DFC6>6?E=2DE65 >:?A6C EH@ >6E6CD @7 AF=G6C:K65
D2>A=6D =E9@F89E96A@H56C.(>62DFC6>6?EE649?:BF642?364@?D:56C652D2AC64:D62?5C6=2E:G6=J
72DE E649?:BF6 :E ?665D D:8?:7:42?E 2>@F?E @7 D2>A=6 AC6A2C2E:@? 6 8 AF=G6C:K:?8 2?5 E96C67@C6 ?@E
DF:E23=67@C C2A:5 @?=:?6D42??:?8AFCA@D6D
"#-& 3F?52?4642=4F=2E653J .(
&14) ".1-& )- -$ $4 *8 $- /3 5( 0 $ 0- / )"- "( 722 20
.

*23=6 56>@?DEC2E6D E96 7=F@C6D46?E >:?6C2=D 56E64E65 3J ":>28:?8 2?5 E96:C 23F?52?46 6DE:>2E65
FD:?8 A:I6=4@F?E:?8 >6E9@5 D66 E:D:>A@CE2?EE@?@E6E92EE9623F?52?46A6C46?E286D9@H?:? *23=6
:D:?5:776C6?ED42=67C@> *23=6 *23=6 6IAC6DD6DE9623F?52?46A6C46?E2862DE96C2E:@36EH66?E96
?F>36C@7 A:I6=D@72 E2C86E4@=@FC2?5 E96 E@E2= ?F>36C@7 A:I6=D :? 6249:>286 @H6G6C6249G2=F6:?
*23=6 :D E9623F?52?46@72>:?6C2=:?AC@A@CE:@? E@E9623F?52?46@7 E96 H9@=6 56E64E65 >:?6C2=D :?
E96D2>6D2>A=62?5E96F?56E64E65A2CE:4=6D2C6?@E:?4=F565
:G6? *23=6 D2>A=6D 2?5 2C6C:49:? 4E:?@=:E6 F8:E6 :DE96>@DE23F?52?E>:?6C2=@7
D2>A=6 G:56?E=J E96 23F?52?E >:?6C2=D 56E64E65 3J ":>28:?8 2C6 >2E4965 H:E9 .(
C6DF=ED ? D2>A=6 F8:E6 H2D 4@>3:?65 H:E9 4E:?@=:E6 2?5 92=4@AJC:E6 92D 366? 2AA62C65
>:I65 H:E9 @=@>:E6 244@C5:?8 E@ (2>2? :?E6CAC6E2E:@? 2?5 7=F@C6D46?E 4@=@FCD :? " :>286D *96
9:89=:89E65>:?6C2=D@7 *23=6 DA64:7J E96>:?6C2=D:56?E:7:653J":>28:?8E649?:BF6
*96 4@>A2E:3:=:EJ @7 E96 7:?5:?8D =625D FD E@ 36=:6G6 E92E ":>28:?8 E649?:BF6 AC@G:56D 2 C6=:23=6
6DE:>2E:@? @7 E96 23F?52?46 @7 7=F@C6D46?E >:?6C2=D E9C@F89 E96 DFC7246 @7 E96 5C:== 4@C6 D2>A=6D
#@C6@G6C4@?D:56C23=62>@F?E@7=3:E692D366?:56?E:7:65 3J":>28:?8 :?D2>A=6 D66 *23=6
H9:=6 =3:E6 H2D ?@E 56E64E65 3J .( >62DFC6>6?E E649?:BF6 I:DE6?46 @7 =3:E6 :? !6G:ED2 @C6
D2>A=6D 92D 366? 4@?7:C>65 3J 86@=@8:DED @7 !6G:ED2 >:?6 @?D6BF6?E=J :E 42? 36 565F465 E92E "
:>28:?842?8:G64@>A=6>6?E2CJ:?7@C>2E:@?E@ @E96C 2DD2J:?8E649?:BF6D DF492D.(
"#-& G6C286@7>:?6C2=D 23F?52?466DE:>2E653JA:I6=4@F?E:?8
&14). ".1-& $4 $4 5( 0- 0-)"- 5( /3 -#*4&

*9:D DEF5J :?EC@5F465 ":>28:?8 E649?:BF6 2D 2 >@56C2E6=J C2A:5 >6E9@5 86?6C2E:?8 2 >2A @7 E96
7=F@C6D46?E >:?6C2=D 7C@> E96 C@F?5D:56 DFC7246 @7 5C:== 4@C6 D2>A=6D E 2==@HD D42??:?8 E96 5C:== 4@C6
D2>A=6D :? E96 3@I6D H:E9 ?@ ?665 7@C D2>A=6 AC6A2C2E:@? 6I46AE 7@C DFC7246 4=62?:?8 *96C67@C6 :E
6?23=6D >@C6 677:4:6?E 6IA=@C2E:@? =6DD 6?6C8J 4@?DF>AE:@? =6DD 6?G:C@?>6?E2= :>A24E =6DD =23@C:@FD
H@C< 2?54@DE67764E:G6?6DD *96EC25:E:@?2=" D6EFA H2D>@5:7:65D@E92E E96=2D6C362> H2D5:DA6CD65
2?5 E96:?E6?D:EJ@7E96=2D6C362>:?4C62D65 *9:D >@5:7:42E:@? 6?23=6572DE6CD42??:?8 @7E96D2>A=6D
#:?6C2=D H6C6 4=2DD:7:65 32D65 @? E96:C 7=F@C6D46?E 4@=@FCD FD:?8 2 C68F=2C:K65 =@8:DE:4 C68C6DD:@?
E649?:BF6 2?5E96 23F?52?46H2DC@F89=J6DE:>2E653JA:I6=4@F?E:?8 :?E965:8:E2=:>286D *96 DA665@7
E96>62DFC6>6?E 6IEC6>6=J 56A6?5D@?E966IA@DFC6E:>6D6E@?E965:8:E2=42>6C2 ?E9:D6IA6C:>6?E
EH@ D64@?5D @7 6IA@DFC6 E:>6 H2D DF77:4:6?E E@ G:DF2=:K6 E96 7=F@C6D46?E >:?6C2=D :? E96 5:8:E2= :>286D
FCE96C>@C6 6I4:E:?8 E96 >:?6C2=D H:E9 9:896C :?E6?D:EJ @7 E96=2D6C 362> 564C62D6D E96 6IA@DFC6 E:>6
*96 C6BF:C65 6IA@DFC6E:>6 2=D@ G2C:6D 32D65@?E96 4@?46?EC2E:@?@7E967=F@C6D46?E>:?6C2=D:?5:776C6?E
>:?6D
?86?6C2=":>28:?8E649?:BF6:D2DF:E23=6>6E9@57@C:?5FDEC:2= 2AA=:42E:@?D @H6G6C E96 ?F>36C
@7E96D2>A=6D6I2>:?6596C6:DD>2== :? 4@>A2C:D@? H:E9 =2C86 BF2?E:E:6D C6BF:C65 :? E96 :?5FDECJ *96
7=F@C6D46?E >:?6C2=D :56?E:7:65 3J ":>28:?8 E649?:BF6 8:G6 255:E:@?2= :?7@C>2E:@? E@ E96 C6DF=ED @7
@E96CE649?:BF6D %?E96@E96C92?5":>28:?8E649?:BF6 :D=:>:E65E@ E96 6I:DE:?87=F@C6D46?E>:?6C2=D
2?52=H2JD?665DE@3642=:3C2E653J@E96CAC64:D6E649?:BF6D

*9:DDEF5J92D366?5@?67@CE96)%(.AC@;64E2A2CE@7E96C66?#:?:?8AC@8C2>7F?565>2:?=J3J
*!)*96:??:D9F?5:?886?4J7@C*649?@=@8J2?5??@G2E:@? -6H@F=52=D@=:<6E@6IE6?5@FC
E92?<D 7@C E96:C G2=F23=6 2DD:DE2?46 E@ 8?:4@28=6 :?=2?5 %J '# !6G:ED2 #:?:?8 %J #
?8:?66C:?8"E5%J#:?6%?":?6)6CG:46%J2?5)A64EC2=>28:?8"E5

&9=&9=@8@A:E6*=4*2=44E4E:?@=:E6":K":K2C5:E64==:?@49=@C6??DE2E:E6F8F8:E6
@@CDE6C:E642=4:E6@@=@>:E6??@CE9:E692=92=4@AJC:E6#28#28?6E:E6&JCC9
&JCC@9E:E6 *C@*C@:=:E6

?5C6H/ $8 62EFC6D6=64E:@?" GD " C68F=2C:K2E:@?2?5C@E2E:@?2=:?G2C:2?46 DE?E6C?2E:@?2=
@?76C6?46@?#249:?6"62C?:?82?2?252
C<25:6G, !?FA76C- "2?89@77$ .C2JD@FC46D ?64<9@77 "2?89@77$ !2??8:66C -656==(

-@=77 5D 2?53@@<@7AC24E:42=.C2J7=F@C6D46?462?2=JD:D )AC:?86CAA S
6=2,6C8?6 2C5(@4<#:?6CTD 2?53@@<C55:E:@? #4?E@D9 ?8:?66C:?8
@H?D( * *96CCF77 AC@;64E2?:?E68C2E65DEF5J@7E96496>:DECJ4CJDE2==@8C2A9J(2>2?2?5:?7C2C65

DA64EC@D4@AJ@7>:?6C2=D ?&C@8C2>2?53DEC24ED@7E96 E96?6C2=#66E:?8@7E96?E6C?2E:@?2=#:?6C2=@8:42=
DD@4:2E:@?:?!@36 2A2? AA % S
D4NC2E6& 2:=@ F6D2=282 =36CE:?:, ( ?6C8J5:DA6CD:G6IC2J5:77C24E:@?DA64EC@D4@AJ7@CC2A:5

6DE:>2E:@?@742=4:E6:?4@AA6C@C6D #:?6C2=D?8:?66C:?8 S
:?=2JD@? )DDECF?<) %AE:>:K2E:@?7@C9F64@?DE2?E(D6?D@CD ?&C@4 )* ) E94@=@FC>28:?8

@?76C6?46
@CE6D "2D6C?2 *9656G6=@A>6?E@77:6=523=6=2D6C:?5F4653C62<5@H?DA64EC@>6E6C$@=:>:ED@?E96

9@C:K@? )A64EC@49:>:424E2&2CES
27E# (6:D76=5( &2?4K6C #@56C?"F>:?6D46?46)A64EC@D4@AJ@7#:?6C2=D2?5#2E6C:2=D 6C=:?$6H

/@C<)AC:?86C
22G:DE@% !2FAA:?6?* P<<P?6? "2D6C?5F465C62<5@H?)A64EC@D4@AJ7@C(2A:5=6>6?E2=

?2=JD:D@7C:==4@C6 ### FE@>2E:4@?EC@= 7@C)FDE2:?23=66G6=@A>6?E S &(%$)AA
F?E:?8E@? -9:E3@FC?" #2D@?& 6C>2?# )49@5=@<# J"@88:?8 ,@=F>:?@FD:?5FDEC:2=
D42=6C67=64E2?46DA64EC@D4@AJ@7E9662CE9TDDF3DFC7246 ?&C@4665:?8D@7)2?5 ()CE)4:6?462?5
AA=:42E:@?D@7(67=64E2?46)A64EC@D4@AJ)J>A@D:F> ,@=
!2FAA:?6?* 22G:DE@% L2<<L2?6? %AE:>:D2E:@?2=8@C:E9>D :?E9642D6@7>:?6C2=56E64E:@?FD:?8

(2>2?2?2=JD:D ?&C@4665:?8D@7 ### ,@= A
!66=:?8 #2F86C F?E:?8E@? )A64EC2=4@C6=@886CFA52E6AC6=:>:?2CJC6DF=ED7C@>E9632C?D8@=5

AC@DA64E #) @FC?2= S
!92;69K2569$ !2FAA:?6?* 22G:DE@% @:?EAF3=:42E:@?H:E9!2FAA:?6?* E@2AA62C W"2D6C?5F465

=F@C6D46?46 :>286D2?5(2>2?DA64EC@D4@AJDEF5:6D@?C2A:5D42??:?8@7C@4<5C:==4@C6D2>A=6DW ?E6C?2E:@?2=
@FC?2=@7#:?6C2=&C@46DD:?8
#:==6C" # (FAA6=# %EE )>:E9( "2?K:C@EE: 6G6=@A>6?E2?52AA=:42E:@?D@72?

6A:7=F@C6D46?46>@5F=67@CDJ?49C@EC@?.C2J7=F@C6D46?46>:4C@AC@36:>28:?8 (6G:6H@7)4:6?E:7:4?DECF>6?ED
) ! )92C>2& "F46J# 9@D9 - F33=62?5! @CE@?U)E2?5@77(2>2?DA64EC@D4@A:456E64E:@?@7

>:?6C2=D@?A=2?6E2CJDFC7246DW)A64EC@49:>:424E2&2CE#@=64F=2C2?5:@>@=64F=2C)A64EC@D4@AJG@= ?@
AA
*2AA6CE# (:G2C5 :=6D *2AA6CE( #2F86C FE@>2E655C:==4@C6=@88:?8FD:?8G:D:3=62?5?62C

:?7C2C65C67=64E2?46DA64EC@D4@AJ242D6DEF5J7C@>E96@=J>A:452>:@4856A@D:ED@FE9FDEC2=:2 4@?@>:4
6@=@8J S
Publication III
Khajehzadeh, N., Kauppinen, T. Fast mineral identification using elemen-

tal LIBS technique. In 4th IFAC Workshop on Mining, Mineral and Metal
Processing MMM, 48, 17, 119-124, https://doi.org/10.1016/j.ifacol.2015.10.089,
October 2015.

c 2015 IFAC.
95
Available online at www.sciencedirect.com
ScienceDirect
IFAC-PapersOnLine 48-17 (2015) 119–124
Fast mineral identification using elemental

LIBS technique
Navid Khajehzadeh ∗ Tommi K. Kauppinen ∗∗
∗
Department of Electrical Engineering and Automation, Aalto
University School of Electrical engineering, Helsinki, Finland (e-mail:
navid.khajehzadeh@aalto.fi).
∗∗
Department of Electrical Engineering and Automation, Aalto
University School of Electrical engineering, Helsinki, Finland (e-mail:
tommi.kauppinen@aalto.fi).
Abstract: Rapid and on-line scanning of rock and drillcore samples gives fast results that can
be used to ease the decision-making process during exploration and to guide the future drilling
activities without delays. Recently, faster and more efficient ore characterization by combining
various laser-based and contactless measurement techniques has drawn tremendous attention in
research. However, complexity of different measurement setups and the difference between the
sources of light make it non-economic and complicated for industry. Considering the wide range
of the elements which can be detected by Laser-Induced Breakdown Spectroscopy and bearing
in mind that LIBS is a very simple spectroscopic technique, the importance of applying LIBS for
fast scanning purposes is certified. This study proposes a simple statistical analysis technique
leading to mineral identification from the elemental results of LIBS. It is shown that LIBS can
be used for calibrating and giving complementary information to other fast scanning techniques
like Laser-Induced Fluorescence imaging. The application of the point-wise LIBS measurement
technique for online and fast estimation of the minerals abundance from the surface of the rock
and drillcore samples is discussed.
© 2015, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
Keywords: Laser-induced breakdown spectroscopy; drillcore sample; mineral detection;
principal component analysis; singular value decomposition; calibration; mineral abundance
1. INTRODUCTION for reflectance spectroscopy. The wide range of the ele-

ments which can be detected by LIBS and the simplicity
This study addresses the issue of rapidly finding and lo- of the measurement setup provides enough motivation to
cating minerals from the surface of rock and drillcore sam- apply LIBS for fast scanning purposes. The main challenge
ples using Laser-Induced Breakdown Spectroscopy (LIBS). is to interpret the existing minerals from the elemental
LIBS is a well-known and widely applied elemental mea- contents provided by LIBS. Thereby, one major phase
surement technique. Developing and applying such fast of this research is about mineral interpretation from the
scanning methods make the costly and time-consuming elemental contents using statistical methods.
laboratory analysis secondary in decision making of the
exploration phase. This in turn eases both the environmen- Combining different contactless measurement techniques
tal and economic pressure in exploration. Prior knowledge can lead to faster and better characterization of drill
about mineralogical contents of the rock and drillcore core samples and it has been investigated and proved in
samples is essential for geologists and makes the process various research works. For example, it has been shown
of logging more effective, thus saving time and manpower. how LIBS technique provides complementary information
It also alleviates the negative economic and social effects for hyperspectral images (Haavisto et al., 2013). Also, a
of the exploration phase of mining. combined LIF-imaging and Raman spectroscopy technique
has been introduced as a reliable technique generating
X-ray Fluorescence (XRF) (Arkadiev et al., 2006) and a mineral map from the surface of the drillcore samples
reflectance spectroscopy in the visible and infrared ranges (Kauppinen et al., 2014). Although the merger of different
(see Huntington et al., 2006; Tappert et al., 2011) are measurement techniques enhances the capability of fast
the prevailing online surface scanning techniques in in- scanning techniques, it should be taken into account that
dustry. However, XRF is an elemental technique and not the complexity of the laser setups and the difference
appropriate when the target is to detect light elements between the sources of light makes it non-economic and
(usually with atomic number less than 20). Furthermore, complicated for industry. Since LIBS and LIF techniques
various elements may appear with similar characteristic can apply the same source of excitation light at similar
emission lines which makes the decision-making process wavelengths (UV range) and considering the simplicity
uncertain. Reflectance spectroscopy is also a widely ap- of LIBS measurement setup, the combination of LIF and
plied online surface scanning technique, however there are LIBS can be a favorable approach.
still considerable number of mineral groups left unknown
2405-8963 © 2015, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
Peer review under responsibility of International Federation of Automatic Control.
10.1016/j.ifacol.2015.10.089
120 Navid Khajehzadeh et al. / IFAC-PapersOnLine 48-17 (2015) 119–124
This study concentrates on the topic of mineral detection 4

x 10
using elemental results of the LIBS measurement tech-

nique. A method of data processing is demonstrated lead- 2
521
Hematite
ing to application of LIBS technique as a capable method Chalcopyrite
in order to calibrate and give complementary information 515
Intensity [a.u]
1.5
to other surface scanning techniques. Furthermore, abun- 495 510

dance estimation from the surface of the rock and drillcore 491
1
516
samples using LIBS is discussed. 488 513 519
0.5
486
2. PROBLEM STATEMENT AND OBJECTIVES
0
For LIBS, a large number of analytical methods for data 460 470 480 490 500 510 520 530
processing is available (Tucker et al., 2010). Principal nm

Component Analysis (PCA) is very commonly used for
LIBS data processing. However, applying PCA for LIBS is Fig. 1. Important atomic transition lines selected as indi-
usually performed off-line so that a large set of measured cators for Hematite and Chalcopyrite
spectra is stored into a database. PCA takes the space
tance metric function, the reference spectrum most similar
between data points and rotates it such that the maximum
to the new measurement is identified. Accomplishment of
variability is visible. Then data clustering is performed on
this work demonstrates the feasibility of applying LIBS
the rotated data (Basilevsky, 2009). Finally the spectra of
technique for rapid mineral detection.
the selected clusters are reproduced in the original space.
Those clusters represent the spectra with the most influen- 2.1 Developing reference spectra
tial and correlated peaks. However, the clustering step of
the analysis is an ambiguous process and usually cannot be
As explained before, interpretation of the minerals from
performed automatically (see the report of Rosen-Gooding
the elemental information produced by LIBS technique ne-
et al. (2014)). The peaks are translated using a reference
cessitates having a database containing minerals reference
database of atomic emission lines.The most regularly used
spectra. The spectra can be produced either manually by
database is the Atomic Line Database by The National
referring to a reference database of atomic emission lines or
Institute of Standards and Technology (NIST) (Ralchenko,
by direct measuring of high grade samples of interest. This
2005). The interpretation of the LIBS lines using reference
section explains the procedure of building the reference
database is usually performed indecisively and based on
spectra by direct measuring. The LIBS spectra are broad
personal impressions since multiple elements may emit at
bands from 350 nm to 800 nm. Figure 1 represents atomic
similar lines. Moreover, the intensity of the emission lines
transition lines selected as indicators for Hematite (Fe2 O3 )
can be affected by many factors such as the intensity of the
and Chalcopyrite (CuFeS2 ) minerals. Both minerals have a
laser light, the distance of the sample from the laser, the
number of similar transition lines (e.g. 486 nm,488nm ,491
shape of the sample and other factors like the moisture
nm , 495 and etc.) due to having iron in common. However,
or external lights. Therefore, such conventional analysis
the difference reveals at wavelengths such as 510 nm, 515
methods restrict LIBS technique for online and fast min-
nm, 521 nm for Chalcopyrite and 513 nm, 516 nm and 519
eral or elemental identification purposes. Nonetheless, the
nm for Hematite. Baseline correction has been performed
ratios of the line intensities of the similar samples are
using an asymmetric least squares smoothing algorithm
usually constant(Gaft et al., 2007) and this fact is the key
(Eilers and Boelens 2005) to highlight the peaks.
point of this research enabling fast mineral detection using
LIBS results. A small wavelength range around each peak is selected and
the rest are set to zero. Reference spectra of other desired
Conventional methods of LIBS data processing can be
minerals are produced similarly. Figure 2 demonstrates
considered as unsupervised learning algorithms. This work
the generated reference spectra of Dolomite, Chalcopyrite,
proposes a supervised learning algorithm resulting in on-
Hematite, Magnesite and Quartz minerals. Each spectrum
line application of LIBS to mineral detection. Since the
is normalized to unit variance. It is important to note
term supervised stems from utilization of a known dataset
that the reference spectra are achieved by measuring
(called the training dataset) to make predictions, in this
the high grade or purred minerals of interest. Otherwise
work target minerals are studied in advance so that purred
impurities of the samples add to the number of irrelevant
(or high-grade) minerals of interest are measured by LIBS
peaks significantly and the intensities of the peaks may
technique. For each target mineral, average of multiple
be affected which makes the spectrum inappropriate for
measured spectra is evaluated and the most representative,
mineral detection.
significant and repetitive peaks are selected as the charac-
teristic of the target mineral. The lines are not selected Expertise is required to select the proper peaks of the
based on the intensity values of the associated elements minerals. Selection of irrelevant peaks adds to the com-
and it makes this method more reliable since human- plexity and influences the accuracy of the results. For
based and perceptual interpretations of the LIBS lines are instance, it is essential to select emission lines of Copper
eliminated. When a database of minerals reference spectra (Cu) for Chalcopyrite (CuFeS2 ) mineral. Otherwise, many
is manually produced, each new spectrum measured by Iron-bearing minerals such as Hematite (Fe2 O3 ) will show
LIBS is inserted into the algorithm alongside the reference characteristic spectrum similar to Chalcopyrite. The tran-
spectra. Using a suitable statistical technique (e.g. PCA or sition lines selected for reference spectra of five minerals
Singular Value Decomposition (SVD)) and a suitable dis- are listed in Table 1. The intensity of the emission lines are
Navid Khajehzadeh et al. / IFAC-PapersOnLine 48-17 (2015) 119–124 121
Table 1. Atomic transition lines used in this equations is less than number of variables. It provides a
study as characteristics of the reference miner- good approximation by substituting a modified version of
als A into the problem where small singular values have been
Mineral Line (nm)
eliminated.
Dolomite 374, 383, 393, 397, 422, 430, 518 The matrix of the score values (U ×T ) represents the spec-
Chalcopyrite 487, 489, 492, 495.5, 510.5, 515, 521.5 tra in n dimensions (note: when n<m, rectangular matrix
Magnesite 383, 385, 518, 568.5, 589 decomposition is performed (Petersen et al., 2008)). Con-
Quartz 385,390, 500, 634.5, 636.5
Hematite 487, 489, 492, 495.5, 500, 513.5, 516.5,
sidering the number of reference spectra, dimensions can
519, 522.5, 527 be further reduced to l. Dimensionality reduction mitigates
the impact of the noise coming from measurement setup
in arbitrary unit. Intensity values of the reference spectra and environment. In this study, the first three principal
are considered as reference values for identification. component directions were sufficient to interpret the data
(l = 3). Therefore, each LIBS spectrum can be represented
2.2 Analytical Approach as a point in a three-dimensional space. The next step is
to realize whether the measured spectrum contains any
Having a database containing reference spectra of the of the target minerals. A threshold t is defined and the
minerals of interest, the next step is to evaluate the new measured point is tracked so that its l-dimensional
spectrum of each new measured pulse and determine euclidean distance with other points is calculated and
whether it contains any of those minerals in the database. stored into vector d:
The easiest way of data processing is line identification,
which is based on straight-forward identification of the
peaks in the wavelength spectra. However, straight line di = (vi,l − cl )2 , i = 1, ..., k and l = 1, 2, 3 (2)
identification is time-consuming, inaccurate and based l
on human perception which may be influenced by many
peripheral factors. c is the point associated with the new measured spectrum
and vi is representative for the ith reference spectrum
Suppose X denotes an n × m matrix of reference spectra of the original space. Nearest neighbors of the point c
where n=k+1, k corresponding to the number of reference satisfying the condition d < t are selected as the existing
spectra and m the number of wavelengths (variables). The minerals in the measured spectrum. It is important to note
last row of the matrix holds the new measured spectrum to that one spectrum may contain more than one mineral.
be evaluated. When the problem is about fast or real-time Therefore, there will be two or more neighbors with
identification of the minerals from LIBS spectra, it is very approximately similar distances from the target point.
plausible that the number of samples is much less than the
number of wavelengths, (n<m). Therefore, applying PCA
and working with covariance matrix is very undesirable 2.3 LIBS Measurement Setup
because of being very large and singular. Singular Value
Decomposition (SVD) can be performed as an alternative The laser-induced plasma was created using an ULTRA
for PCA specially when the number of samples is less 100 Q-switched Nd:YAG laser operating at 266 nm with
than variables. There are numerous sources describing the 25 mJ/pulse and up to 20 pulses/s repetition rate. The
relation between SVD and PCA analysis (Madsen et al., pulse duration was 8 ns. An optical fiber used to collect the
2004). It decomposes the matrix Xn×m (n<m) into: radiation from the plasma and the spectrum was obtained
by a USB4000 Fiber Optic Spectrometer covering from
X = Un×n Tn×m V t m×m (1)
350 nm to 1000 nm. In order to avoid continuum radiation,
The product of U and T contains the scores and V t usually the detector is triggered with a delay. Another way
demonstrates the loading values. Generally, SVD is a to avoid continuum radiation is to detect the emission at
powerful solution for least square problems of any rank a few millimeters above the sample as the plasma plume
or shape. More importantly, it can be used for rank- expands and cools down. In this study, the second method
deficient problems like AX = b where the number of was sufficient.
14 3. RESULTS
12
Dolomite
Chalcopyrite 3.1 LIBS for calibrating other techniques
Hematite
Intensity [a.u]
10
Magnesite
8
Quartz A combined Laser-Induced Fluorescence and Raman spec-
troscopy method for mineral identification has been in-
6
troduced by Kauppinen et al. (2014) and the viability of
4 the approach has been deeply discussed. However, there
2
are still a number of important minerals non-responsive
with both LIF-imaging and Raman spectroscopy tech-
0
400 450 500 550 600 niques. For example, important sulfides like Chalcopyrite
nm and Pentlandite or iron-bearing minerals like Hematite are
neither fluorescent nor easily traceable by Raman spec-
Fig. 2. Generated reference spectra troscopy. Using the proposed analytical approach for LIBS
mineral identification, this work demonstrates rapid iden-

tification of Chalcopyrite, Hematite and Quartz minerals
from the surface of the rock and drillcore samples.
Fig. 4. Scatter plot of the first three scores. Chalcopyrite

is the nearest neighbor.
contains peaks at other wavelengths associated with other
elements (e.g. 500 nm, 527 nm, 568 nm, 570 nm, 589 nm
and etc) , the algorithm represented Chalcopyrite as the
most compatible mineral because the variation of the data
is chiefly following Chalcopyrite reference spectrum.
The proposed analytical method implies the feasibility of
applying LIBS technique for rapid calibration of other sur-
Fig. 3. (a) Ordinary close-up photograph of a drillcore face scanning methods specially those functioning based on
surface from Kevitsa mine,depth: 18 meters, (b) laser- image processing of the drillcore surface.
induced fluorescence (LIF) image of the drillcore
surface
3.2 Multiple mineral detection when mineral grains are
Figure 3-a is showing an ordinary close-up photograph mixed
of one piece of the drillcore sample from the depth of
18 m collected from Kevitsa mine, Finland. Figure 3-b When the laser beam hits more than one mineral grain at
represents laser-induced fluorescence (LIF) image of the the same time, the spectrum holds a mixture of multiple
specified region in Figure 3-a (white circle) on the surface lines of minerals. Numerous analytical methods applied
of the drillcore . LIF imaging was performed using the to online mineral detection processes are incapable of
same Nd:YAG laser utilized for LIBS technique. The red identifying such spectra. Here we discuss the capability of
fluorescent colour is Albite and the bluish-white colour is the proposed method for identification of more than one
Magnesite. However, in figure 3-a the golden grain (high- mineral in a spectrum. Furthermore, online application
lighted by arrow) did not fluoresce and they were not iden- of LIBS technique for sorting of the drillcore samples
tified by Raman spectroscopy. The position of the golden is discussed. Figure 6 represents one piece of high-grade
color was measured by 10 pulses of LIBS. The analysis (>60%) Hematite drillcore sample from the depth of 168
was performed to mimic the online or real time analysis m below the ground level. The dark-gray, brownish and
of each new pulse. After baseline correction and scaling white color veins represent high-level Hematite, low-level
to unit variance, the average of the measured spectra was Hematite and Quartz respectively. The white arrow points
appended to the matrix of the reference spectra. Using
SVD and the first three principal components, data were
analyzed in a new 3-D space. Chalopyrite Reference
6
Average of measured
Figure 4 shows the scatter plot of the first three scores spectra
where each data point stands for a full spectrum in the 5
Intensity [a.u]
original space. The red point corresponds to the spectrum 4

measured from the golden spot on the surface of the
drillcore. The euclidean distance of the red point and the 3
rest of the points was calculated. A threshold of t=10

2
was defined to recognize possible candidate minerals. The
distance between the average of the measured spectra and 1
Chalcopyrite was less than ten and the distance with

other reference minerals was larger than the threshold. 380 400 420 440 460 480 500 520 540 560 580
Therefore, it can be clearly deduced that the golden nm

spot represents Chalcopyrite mineral. Figure 5 shows the
spectra corresponding with the red point and Chalcopyrite Fig. 5. Chalcopyrite reference and spectrum of the new
in the original space. Although the measured spectrum measured point
Navid Khajehzadeh et al. / IFAC-PapersOnLine 48-17 (2015) 119–124 123
Fig. 6. A high-grade (>60%) Hematite drillcore sample

from the depth of 168 m
to the position where Hematite and Quartz grains are Fig. 8. Scatter plot of the first three scores. Quartz and
mixed. Hematite are the nearest neighbors.
Figure 7 illustrates the spectrum measured from the po-
sition specified by white arrow in Figure 6 along with position was scanned by ten pulses and the average of
Hematite and Quartz reference spectra. Double-arrows the spectra was taken as representative of each scanned
indicate the wavelength range where measured spectrum position. In total 18 averaged spectra were produced and
is following Hematite and the single arrows represent the inserted into the algorithm when generated. 11 out of
wavelengths where peaks of Quartz occur. Although the 18 spectra were associated to Hematite and 5 spectra
measured spectrum contains peaks compatible with other to Quartz. Two spectra consisted of both Quartz and
reference minerals, the variations of the data and the ratio Hematite. As in this case Hematite is the main criterion
of the peaks are mostly compatible with Hematite and for ore grade, it can be roughly estimated that the sample
Quartz. This sentiment is authenticated in figure 8 where is from a high-grade (>60%) ore zone.
both Hematite and Quartz minerals situate in the close
neighborhood of the measured spectrum (red point) in the Figure 9 shows another 10 cm piece of low-level (∼ 30%)
new 3-D space. Hematite drillcore sample from the depth of 205 m below
the ground level. The sample was scanned through the red
In three-dimensional Euclidean space, the distance be- line. 10 out 18 spectra were detected as Quartz, 6 out of
tween the measured spectrum and Dolomite, Chalcopyrite, 18 as Hematite and 2 spectra contained both Hematite
Hematite, Magnesite and Quartz were 18.4, 16.03, 6.8, and Quartz. Therefore, by online scanning of the drillcore
39.29 and 8.47 respectively. sample the grade of ore can be roughly approximated
to 39%. Increasing the scanning resolution makes the
3.3 LIBS for online drillcore sorting. abundance approximation more precise but it reduces the
speed of measurement.
Point-wise measurement techniques such as LIBS can be
used for sorting the drillcore samples containing composite
minerals of interest so that small fraction of the sample is
representative of the large volume. The red line in Figure
6 shows the scanning line of LIBS. Mineral detection from
LIBS elemental data was performed on a 10 cm piece of
drillcore and the scanning resolution was 5 mm . Each
4
Hematite Reference
3.5 Quartz Reference
Measured spectrum
3
Fig. 9. A low-grade (<30%) Hematite drillcore sample
Intensity [a.u]
2.5 from the depth of 205 m

2
1.5
It is important to emphasize that for online abundance
1 estimation of the minerals using LIBS technique, selecting
0.5
the samples containing mixed materials so that the surface
is indicator of the whole volume is of paramount impor-
400 450 500 550 600 tance. Future research will focus on reporting the fast
nm
abundance estimation from larger volumes of the samples.
The proposed analytical approach can be extended to
Fig. 7. New measured spectrum versus Hematite and measure the concentration of the minerals and it provides
Quartz reference spectra. the backbone of the future work.
4. CONCLUSION Rosen-Gooding, A., Ollila, A., Gordon, S., Newsom, H.,

Williams, A., Martinez, R., Wiens, R., and Clegg, S.
This paper presented an analytical method leading to (2014). Laser-induced breakdown spectroscopy as a tool
rapid mineral identification from the elemental contents to differentiate compositions of iron-bearing minerals.
of the LIBS technique. Using a statistical method (SVD), LPI Contributions, 1791, 1174.
the minerals content of each spectrum measured by LIBS Tappert, M., Rivard, B., Giles, D., Tappert, R., and
technique was evaluated. This evaluation was performed to Mauger, A. (2011). Automated drill core logging using
imitate online and rapid mineral identification. Analysis visible and near-infrared reflectance spectroscopy: a
was performed in a reduced dimension of the rotated case study from the olympic dam iocg deposit, south
space of the data points. A similarity search procedure australia. Economic Geology, 106, 289–296.
was performed between the averaged spectrum of each Tucker, J., Dyar, M., Schaefer, M., Clegg, S., and Wiens,
measured position and the reference spectra produced R. (2010). Optimization of laser-induced breakdown
manually. The results show that the proposed method can spectroscopy for rapid geochemical analysis. Chemical
give complementary information to other surface scanning Geology, 277(1), 137–148.
techniques like LIF-imaging. The feasibility of detecting
multiple minerals in one spectrum was presented. Ulti-
mately, application of online LIBS technique for sorting
purposes was discussed shortly.
ACKNOWLEDGEMENTS
We would like to acknowledge Aalto University for sup-

porting us and providing us with the working environment
to write this paper. We would also like to extend our
thanks to the geologists of the Kevitsa and Kittila mines,
IMA Engineering Ltd Oy and Mine On-Line Service Oy
for their great support.
REFERENCES
Arkadiev, V., Knüpfer, W., and Langhoff, N. (2006). X-
ray sources. In B. Beckhoff, N. Langhoff, B. Kanngieer,
R. Wedell, and H. Wolff (eds.), Handbook of practical
X-ray fluorescence analysis., 36–53. Springer.
Basilevsky, A.T. (2009). Statistical factor analysis and
related methods: theory and applications, volume 418.
John Wiley & Sons.
Gaft, M., Sapir-Sofer, I., Modiano, H., and Stana, R.
(2007). Laser induced breakdown spectroscopy for bulk
minerals online analyses. Spectrochimica Acta Part B:
Atomic Spectroscopy, 62(12), 1496–1503.
Haavisto, O., Kauppinen, T., and Häkkänen, H. (2013).
Laser-induced breakdown spectroscopy for rapid ele-
mental analysis of drillcore. In IFAC MMM2013 Au-
tomatic Control for Sustainable Development - PRO-
CEEDINGS, 87–91.
Huntington, J., Whitbourn, L., Mason, P., Berman, M.,
and Schodlok, M.C. (2006). Hylogging - voluminous
industrial-scale reflectance spectroscopy of the earths
subsurface. In Proceedings of ASD and IEEE GRS; Art,
Science and Applications of Reflectance Spectroscopy
Symposium, volume 2.
Kauppinen, T., Khajehzadeh, N., and Haavisto, O. (2014).
Laser-induced fluorescence images and raman spec-
troscopy studies on rapid scanning of rock drillcore
samples. International Journal of Mineral Processing,
132, 26–33.
Madsen, R.E., Hansen, L.K., and Winther, O. (2004).
Singular value decomposition and principal component
analysis. Neural Networks, 1, 1–5.
Petersen, K.B., Pedersen, M.S., et al. (2008). The matrix
cookbook. Technical University of Denmark, 450, 7–15.
Ralchenko, Y. (2005). Nist atomic spectra database. Mem-
orie della Societa Astronomica Italiana Supplementi, 8,
96.
124
Publication IV
Khajehzadeh, N., Haavisto, O. and Koresaar, L. On-stream and quantita-

tive mineral identification of tailing slurries using LIBS technique. Minerals
Engineering , 98, 101-109, https://doi.org/10.1016/j.mineng.2016.08.002,
August 2016.

c 2016 Elsevier Ltd.
103
Minerals Engineering 98 (2016) 101–109
Minerals Engineering
journal homepage: www.elsevier.com/locate/mineng
On-stream and quantitative mineral identification of tailing slurries

using LIBS technique
Navid Khajehzadeh a,⇑, Olli Haavisto b, Lauri Koresaar b
a
Aalto University School of Electrical Engineering, Department of Electrical Engineering and Automation, Finland
b
Outotec (Finland) Oy, Rauhalanpuisto 9, 02230 Espoo, Finland
a r t i c l e i n f o a b s t r a c t
Article history: Mineral flotation processes are controlled by monitoring the grade of the present minerals. The perfor-
Received 19 April 2016 mance of flotation can be greatly optimized by online assaying of the minerals in the slurry flows.
Revised 1 August 2016 However, online and quantitative mineral identification of the slurries is challenging. The major focus
Accepted 2 August 2016
of this research is about measuring mineral contents from the elemental concentrations acquired by
Available online 8 August 2016
an on-stream slurry analyzer operating based on laser-induced breakdown spectroscopy (LIBS). A multi-
variate statistical method called partial least squares regression is employed to perform the elemental to
Keywords:
mineral conversion. In this work, the samples under study come from an iron concentrator where knowl-
Laser-induced breakdown spectroscopy
Slurry
edge about the grade of silicates and iron-oxides are important for controlling the flotation plant. In total,
Tailings the concentrations of six out of ten minerals were successfully estimated. This accomplishment provides
On-stream quantitative knowledge about mineral contents from the nearly real-time assay data without any
Mineral interpretation modification on the measurement setup or further instrumentation. This can lead to many benefits such
Partial least squares as rapid control of concentrate quality, enhanced recovery and savings in money, time, energy and
manpower. The proposed technique is applicable to all types of slurry samples.
Ó 2016 Elsevier Ltd. All rights reserved.
1. Introduction requires advanced data analyzing to extract mineralogical knowl-

edge quantitatively.
Identification of mineral composites and quantitative measure Given the needs of the mineral-processing plants for rapidly
of mineral contents provide valuable information for mining. Infor- measuring of the mineral contents, it is advantageous to identify
mation about mineral contents not only gives broad knowledge the mineral quantities from the ordinary elemental assays
about source, history and type of the rock bodies, but also today collected from the streams of the process circuit. The concept of
is necessary for better process control and more efficient use of mineral estimation from the chemical assays is usually expressed
energy and raw materials. The rapid and on-line scanning of rock by the term elemental-to-mineral conversion. A few studies have
and slurry samples gives fast results that can be used to ease the been published on this topic as elemental analysis is simpler,
decision-making process, however the majority of the online and cheaper and more available in comparison with the process of
on-stream analytical techniques are based on elemental analysis quantitative mineral analysis. Whiten (2007) calculates the con-
or performed off-line when applied to mineral identification. centration of minerals from elemental assays using singular value
Many flotation processes contain minerals that can affect the decomposition. Lamberg and Vianna (2007) estimate mineral
flotation performance but cannot be quantified using elemental quantities from elemental contents using minimization of a set of
analysis methods. On-line mineral assaying provides useful infor- mass balancing equations. However, their techniques are still slow
mation for process control and e.g. chemical consumption opti- for on-stream measurement of the minerals and are not considered
mization. Therefore, the importance of rapid and quantitative as automated techniques.
mineral analysis is intuitive. On the other hand, developing new Lund et al. (2013) discusses that mineral analysis using tradi-
measurement setups only for the purpose of mineral identification tional optical microscopy along with point counting method is a
is costly, needs proper space for installation in the plant and slow technique and subjective. While the image analysis is improv-
ing the speed, this technique is unsuccessful for very fine grained
⇑ Corresponding author. samples. It is also discussed that techniques based on scanning
E-mail addresses: navid.khajehzadeh@aalto.fi (N. Khajehzadeh), olli.haavisto@ electron microscopy and images analysis such as QUEMSCAN,
outotec.com (O. Haavisto), lauri.koresaar@outotec.com (L. Koresaar). MLA and PTA have improved the speed of quantitative mineralogy
http://dx.doi.org/10.1016/j.mineng.2016.08.002
0892-6875/Ó 2016 Elsevier Ltd. All rights reserved.
102 N. Khajehzadeh et al. / Minerals Engineering 98 (2016) 101–109
and they have turned out to be supreme alternative for optical Khajehzadeh and Kauppinen (2015), a statistical method leading
microscopy and XRD mineralogy. However, Lund et al. (2013) man- to online and qualitative interpretation of the mineral contents
ifests some drawbacks of the technique such as standard-less, from the elemental results of LIBS has been discussed. Additionally,
time-consuming sample preparation, problem of resolution for the method was demonstrated to succeed on sorting hematite drill
very fine particles and difficulty in the interpretation of similar core samples. However, the proposed technique is considered as
minerals. They introduce a practical way of elemental to mineral standard calibration method where reference spectra of the pure
conversion leading to development of an ore classification system. or high grade samples of present minerals are required. Moreover,
There are direct mineralogical measurement techniques avail- the technique was performed on the dry (drill core) samples.
able like Raman spectroscopy, X-ray diffraction (XRD) and hyper- McMillan et al. (2007) demonstrated identification of carbonates
spectral imaging, but with substantial limitations for online and silicates by using comparisons of correlation coefficients
usage. Raman spectroscopy (Das and Agrawal, 2011) has relatively between averaged LIBS spectra and a database of the LIBS spectra
long measurement time and it is susceptible to auto-fluorescence of the same minerals. However, as discussed before most of the
emission of the minerals. Moreover, in order for a sample to be successful works and published results about elemental and min-
Raman active, it must have vibrations (or rotations) that result in eralogical analysis have been performed on the dry samples or in
a change in the polarizability of the molecule or material. There- laboratory conditions while on-stream and quantitative miner-
fore, Raman scattering in metallic substances is very weak, and alogical analysis of the aqueous samples and slurries is still a big
cannot be detected without special conditions (Goncharov and challenge for the mining sector.
Struzhkin, 2003). XRD (Warren, 1969) is an analytical method for LIBS technique has become popular for on-stream elemental
mineralogical identification and quantification but its online appli- analysis of the slurry samples. It gives real-time on-line assays
cation is limited. Saarhelo et al. (1988) reported that online XRD for monitoring and control of the mineral processing plant. In addi-
measurement is successful where the mineralogy is simple e.g. tion to the elemental contents, it would be very beneficial to mea-
potash and talc concentrators. Hyperspectral imaging (Chang, sure the mineralogical contents of the slurries from the LIBS
2003) is a method widely applied for surface scanning but there spectra with no modification of the measurement setup. Given
is considerable number of mineral groups that do not give hyper- the large amount of information in each LIBS spectrum, a data-
spectral response and hence cannot be identified. Therefore, base approach for interpreting the measured data is a natural
despite the great advances in the current laser-based and non- choice.
contact scanning techniques, there is a need for novel data inter- This work presents application of a multi-variable statistical
pretation techniques leading to nearly real-time mineralogical technique called Partial Least Squares (PLS) along with Non-
analysis. Linear Iterative Partial Least Squares (NIPALS) algorithm (Geladi
LIBS (Cremers and Radziemski, 2006) and XRF (Beckhoff et al., and Kowalski, 1986) enabling estimation of the mineral contents
2007) measurement techniques both have high capabilities for from the elemental LIBS spectra produced by an on-stream slurry
measuring elements. Nowadays, both techniques are widely analyzer. The method has been applied on the tailing samples com-
applied as cross-belt analyzers of the crushed ore and slurries of ing from an iron concentrator. While the conventional methods
the flotation processes (Outotec, 2015). On one hand, LIBS is con- used for LIBS data analyzing mainly consider only a few of the sig-
sidered as a point-wise measurement technique and less reliable nificant LIBS lines, the proposed method in this work takes the full
for inhomogeneous materials while XRF has higher resolution cov- spectrum of LIBS measurements into account. This will result in
ering larger area on the surface. On the other hand, LIBS has a very more reliable and better accuracy of the interpretations. Results
simple and easy-to-implement spectroscopic setup, wider range of enable quantitative and continuous mineralogical analysis of
elements can be detected with online LIBS and measuring elements online monitoring of the slurry samples.
with atomic number less than 20 is rather simpler by LIBS com-
pared with XRF (Gaft et al., 2007). According to (Khater, 2013),
measuring elements lighter than titanium using XRF method and 2. Experimental
without helium gas purge or performing within a vacuum chamber
is very difficult. 2.1. Laser-induced breakdown spectroscopy
The interest in LIBS application for geological analysis has been
recently increased due to multiple advantages over other analytical Laser-induced breakdown spectroscopy (LIBS) is an analytical
methods, e.g. simplicity, economic, no sample preparation, simul- laser-based technique operating based on detection and analysis
taneous multi-elemental analysis, and broad range elemental of atomic emissions. It was first introduced by Brech and Cross
detection. As there is plenty of chemical information in LIBS spec- (1962) and after the 1990s by improvement in the laser technolo-
tra, numerous analytical techniques have been proposed. Lanza gies, LIBS became appealing for engineering applications. It is a
et al. (2010) applied Principal component analysis (PCA) to classify simple spectroscopic technique mainly built on a pulsed laser
carbonates. Zhu et al. (2014) conducted a comparison of partial source, a focusing lens, a spectrometer and a data analysis unit.
least squares discriminant analysis (PLS-DA) and support vector LIBS operates by employing a low-energy pulsed laser (usually tens
machines (SVM) to classify samples based on their LIBS spectra. to hundreds of mJ per pulse) and using a focusing lens it generates
A multivariate analysis was applied on LIBS data to extract quanti- a plasma where a small volume of the sample is vaporized. LIBS is
tative elemental compositions and identify the rock types by Clegg also sometimes called laser-induced plasma spectroscopy (LIPS).
et al. (2009). Sirven et al. (2006) applied Partial Least Squares The plasma consists of a mixture of lights (wavelengths) emitted
Regression (PLSR) to quantitatively analyze the chromium concen- by present atomic and ionic species. A portion of plasma light is
tration in soil samples. However, all the proposed methods have collected and a spectrometer disperse the light into the intensity
been applied either in laboratory conditions or on the solid and wavelength components and after digitizing the signal the
samples like rocks. data are displayed. The displayed spectrum is used for elemental
Due to the affluence of complex information in each LIBS analysis of the samples as each element emits at specific wave-
spectrum, processing all the data online and in a reliable manner lengths upon cooling of plasma (Cremers and Radziemski, 2006).
is challenging. Majority of the proposed analytical techniques are Today LIBS is in use extensively in many different applications
performed offline on a database of the collected spectra and efforts and fields such as environmental monitoring (Ayyalasomayajula
have been made to reduce the time of computation. In et al., 2012; Awan et al., 2013), geology (Rakovskỳ et al., 2012;
N. Khajehzadeh et al. / Minerals Engineering 98 (2016) 101–109 103
Díaz Pace et al., 2011) or space exploration (McCanta et al., 2013; in the underflow of the cells, and the gangue mineral (quartz) is
Wiens et al., 2013). As LIBS technique illuminates samples with a lifted to the froth (Filippov et al., 2014). As the flotation is an
high-power pulsed laser, in aqueous samples the illuminated expensive process, the primary analysis is about whether the grade
liquid is vaporized and ionized locally and thus the high- from the magnetic separation is high enough to bypass the flota-
temperature plasma contains all presented elements in the samples. tion circuit. The secondary analysis would be the silica in the flota-
As a result, LIBS technique is a suitable technique for slurry analy- tion circuit, for chemical control. It is also advantageous to
sis. A review of the LIBS analysis for liquid samples has been differentiate between different Si-bearing minerals as quartz has
reported by Yu et al. (2014). different flotation properties than mixed silica minerals.
Fluids have intrinsic motions of molecules and ions. Molecular
2.2. Measurement setup motions, shaking and splashing of the liquid affects the LIBS
plasma and causes inconsistency of the collected spectra. More-
The measurement setup used in this work is a LIBS system over, part of the light energy is absorbed by water and thus the
(measurement head of Courier 8 SL) from LIBS laboratory of Outo- detected signal is weaker than the solid samples (Whiten, 2007).
tec, Finland. Theoretically, the technology is able to detect all In order to compensate such instabilities and mitigate the noise,
elements but the actual elemental range depends on the character- three LIBS spectra per sample were produced. Each spectrum cov-
istics of the laser and spectrograph, ambient conditions and sample ered the wavelength range from 190 nm to 507 nm and was the
composition (e.g. particle size, chemical composition and solid per- product of 200 accumulated laser pulses. Such procedure was
centages). The product consists of a pulsed laser, a focusing lens, a repeated three times and then the three accumulated spectra were
spectrometer and a desktop PC to perform the data collection and averaged to produce one spectrum (total of 600 pulses per spec-
control the system. The laser-induced plasma was created using a trum). This process was repeated two more times per slurry sam-
120 mJ laser pulse of 8 ns duration generated by a Q-switched Nd: ple and in total 3 accumulated and averaged spectra per sample
YAG laser operating at 1064 nm with up to 10 pulses/s repetition were recorded.
rates. A convex lens was used to focus the laser beam and heat Given the fact that the noise produced by most of the natural
small portion of the sample into hot plasma. The light spectrum processes and measurement setups have the characteristics of
emitted by the atoms and ions in the cooling incandescent plasma the white noise, accumulation of the spectra will majorly reveal
is measured after a short time to avoid continuum radiation. A the elemental peaks and diminishes the noisy magnitudes. As the
delay time of 1 ls and gating time of 233 ls was applied on the slurry is circulating and the laser beam hits different particles each
detector. The detector applied in the system is a Peltier cooled time, it is likely that the accumulated spectrum consists of the
CCD camera attached to an echelle spectrograph with resolution combination of all the elements and minerals. The slurry samples
of 32 pm (FWHM) for Si line at 288.16 nm. were mixed by dry sample with about 6 L of water. The average
Fig. 1 demonstrates a schematic diagram of the LIBS measure- solids content of the tailing samples was 21% (w/w).
ment cell and slurry flow. This laboratory setup was designed to
mimic the in situ LIBS analyzer (Courier 8 SL). The sampled slurry 2.4. Mineralogical assays of the slurry samples
is circulated into the LIBS measurement cell using a pump. Slurry
flows into the measurement cell consisting of a stabilized surface The quantitative mineralogical information of the samples was
where the LIBS apparatus is scanning a fixed point of the surface. provided by Mineral Liberation Analyzer (MLA) method. The con-
The circulation system is also equipped with a mixer, and the centration of Magnetite was measured by SATuration MAGnetiza-
speed of the slurry coming out of the measurement cell is about tion ANalayzer (SATMAGAN). SATMAGAN has been developed
25 L/min. particularly for the accurate measure of magnetite especially in
iron ore concentrators (Rapiscan Systems, 2005). It is important
to note that MLA is a 2D or surface scanning technique, hence
2.3. Slurry samples and data collection
the accuracy is less than 3D analysis.
Table 1 demonstrates the mineralogical assay data of 14 tailing
The samples set under study were 14 slurry samples from tail-
samples analyzed by MLA and SATMAGAN analyzers in the labora-
ings stream of a flotation plant. The plant belongs to an iron ore
tory. According to the results, quartz and Fe-oxides (magnetite/
concentrator, where the process is magnetic and gravity separation
hematite) are the abundant minerals. LIBS technique is incapable
followed by a reverse cationic flotation of quartz. In the reverse
of distinguishing between magnetite (Fe3O4) and hematite
flotation process the valuable mineral (magnetite) is concentrated
(Fe2O3) due to their identical elemental contents. Therefore, in this
work the collection of hematite and magnetite is measured as Fe-
Slurry flow
oxides.
Plasma
LIBS probe 2.5. Pre-processing of the LIBS spectra
PC
Detector Data analyzing of LIBS spectra is preceded by data preprocess-
Measurement
ing. First a baseline correction is performed on each spectrum. In
cell
Laser this work, baseline correction has been performed using an asym-
Power metric least squares smoothing algorithm (Eilers and Boelens,
unit
Return 2005) to highlight the peaks. In order to make the PLS algorithm
vessel automated and prevent any error which may be caused by the
Slurry collinearity or singularity in the matrix of independent variables
(LIBS spectra), the zero columns (wavelengths) were eliminated.
Since the recorded spectra contain noisy peaks, many spurious
Tubing peaks might be taken into account by PLS when mapping the
Pump inputs to the outputs. This will increase the computational time
and adds to the variability of the results and usually further user
Fig. 1. Laboratory measurement setup for the mineralogical LIBS slurry analysis. interaction is needed to interpret the results properly. Therefore,
Table 1
Mineralogical assays (%) of the tailing samples measured by MLA and SATMAGAN.
Mineral Sample (%)

T1 T2 T3 T4 T5 T6 T7 T8 T9 T10 T11 T12 T13 T14
Quartz 34.8 40.0 30.4 45.1 60.8 52 44.8 37.1 48.2 59.3 61.5 54.4 38 46.6
K-feldspar 2.2 3.5 3.1 2.4 2.1 2.3 2.4 2.2 1.5 2.6 2.4 2.8 2.4 1.8
Aegirine-augite 3.0 2.7 4.3 2.9 1.5 2.1 1.9 3.2 2.1 1.4 1.4 1.5 2.5 1.1
Biotite 1.9 1.9 2.1 4.0 4.2 3.5 3.5 2.9 2.0 1.9 1.7 1.7 1.6 1.3
Galuconite 0.1 0.2 1.1 1.8 2.5 1.3 1.3 0.6 0.1 0.2 0.2 0.4 0.3 0.5
Stilpnomelane 4.3 4.2 3.8 4.0 2.1 2.8 2.6 3.7 5.3 3.0 1.8 1.9 3.2 3.6
Actinolite 0.1 0.2 0.0 0.0 0.0 0.1 0.1 0.1 0.1 0.2 0.2 0.3 0.5 0.3
Ferrorichterite 8.6 8.1 2.5 3.7 1.9 3.4 3.5 7.7 9.1 3.2 2.1 2.3 3.6 1.9
Magnetite 42.0 37.1 47.2 31.1 17.6 26 32.3 39.1 29.1 19.6 21.0 27.0 39 35.6
Hematite 1.7 1.0 4.3 3.8 5.8 5.0 6.2 2.4 1.3 7.0 6.6 6.3 7.0 5.0
Unknown 0.9 0.9 0.9 0.8 1.2 1.0 1.0 0.8 0.8 1.3 1 1.3 1.1 1.9
a magnitude threshold was defined to robustly detect those local where W is the weight matrix. Now:
maxima or valleys which are significantly greater than the values
B ¼ W UT ð6Þ
around. The peak finding is performed using the basic principle
of first-derivative while a minimum peak distance threshold is which gives the original mapping as in Eq. (1):
specified to ignore those peaks being in the close vicinity of the
local maxima. Such close-distance peaks may occur as a result of Y ¼ XW UT þ F ¼ XB þ F ð7Þ
self-absorption in the line-shape of the LIBS emission lines. In practice, PLS is implemented using NIPALS algorithm. It
improves the time of computation since it computes only the first
2.6. Partial least squares and NIPALS algorithm few required latent variable dimensions rather than computing the
whole components of a high-dimensional structure. Furthermore,
Considering the LIBS spectra as a matrix of independent vari- the dimension of the latent basis found by eigenvector-based PLS
ables (X), the input matrix is high dimensional and very prone to approach is restricted to the number of either input or output vari-
contain collinearity. Since the target is to generate a model esti- ables while iterative PLS approach dispels these constraints
mating the concentration of the mineral compositions in each (Hyötyniemi, 2001). Given U and Z as the score matrices of Y and
slurry sample using the input matrix of pixel intensities, we define X respectively and W as the projection of X so that Z has the max-
matrix Y to contain the mineralogical concentrations data gener- imum correlation with U, the following steps are are taken for
ated by MLA technique in the laboratory. The model is trained NIPALS algorithm to obtain one latent dimension:
using Partial Least Squares (PLS) regression. PLS is a well-known
statistical method introduced by Wold (1985) and can be consid- 1. Initialize U i .
ered as the extension of Principal Component Regression (PCR).
2. wi ¼ X T U i =ðU Ti U i Þ.
PCR takes only the variance of the input variables into account.
3. Z i ¼ Xwi .
However, PLS is capable of both overcoming the collinearity of
4. /i ¼ Y T Z i =ðZ Ti Z i Þ.
the independent variables and it takes the correlation between
both inputs and outputs into account. It is a useful technique when 5. U i ¼ Y T /i =ð/Ti /i Þ.
the target is to predict a set of dependent variables from a large set 6. Repeat steps 2.–5. until convergence.
of independent variables. 7. Deflate X and Y matrices: X X Z i hTi , Y Y Z i /Ti .
PLS technique is explained using the eigenvector-based proce-
dure (Wold et al., 2001). However, in practice PLS is implemented As a result, NIPALS-based PLS algorithm generates the model B
using an iterative algorithm explained afterward. employing linear combination of the independent variables X so
The PLS model makes a linear mapping, B, between the that the spectral changes are correlated with the mineral concen-
independent variables, X, and dependent variables, Y: trations in the output matrix Y.
Y ¼ XB þ F ð1Þ
3. Results and discussion
where F is the residual term. When computing matrix B, the struc-
ture of both data matrices X and Y is taken into account; both matri- In order to train the model using PLS, first the three accumu-
ces are interpreted as a product of scores and loadings, which are lated and averaged spectra produced per slurry samples were aver-
further utilized in the model estimation: aged to have one so-called training spectrum for each sample. For
Y ¼ U UT þ G ð2Þ 14 slurry samples and 10 minerals to be predicted, the output
matrix is Y ð1410Þ and the input matrix is X ð14111Þ . In this work,
X ¼ Z HT þ E ð3Þ
the complexity of the model is defined by the number of latent
where U and U are the scores and loadings of Y, and Z and H are the variables in the PLS model. Assigning large values for the latent
scores and loadings of X, respectively. As a result of the PLS estima- variables causes overfitting of the model. Therefore, the regulariza-
tion, the loadings and scores are selected so that the residual terms tion of the model is performed using the idea of early stopping. The
G and E are small. Furthermore, the scores of X are used as predic- model is trained using leave-one-out cross validation. At each iter-
tors of Y: ation of the algorithm, one spectrum is left out for validation and
13 spectra are employed for training. The algorithm runs 14 times
Y ¼ Z UT þ F ð4Þ and the mean squared error (MSE) of both training and validation
the X score can be computed from X as: data are calculated for each mineral separately as a criterion for
model selection. By increasing the number of latent variables, an
Z ¼ XW ð5Þ iterative reduction of the training error is expected for the training
data. However, the error computed with respect to the validation spectra generated by LIBS technique are accumulation and average
data, called validation error, is expected to show first a decrease of 600 laser shots. Therefore, during the validation process it is
and then followed by an increase when the model starts to very plausible that the averaged spectra of different samples are
over-fit. Consequently, each mineral is evaluated separately since relatively identical and PLS algorithm might find the left-out spec-
different minerals may be predicted more precisely with different trum similar to some of those spectra in the training set. Thus,
number of latent variables. The complexity of the model is selected while increasing the number of latent variables, the validation
by two criteria simultaneously: (1) when the training error slumps error may decrease accordingly. This justification is better per-
and then starts to have a steady decrease and (2) when the valida- ceived by evaluating the mineral concentration values in Table 1.
tion error is around a local (or global) minimum at the same time Some of the samples have quite similar mineral grades and hence
that training error starts having steady decrease. their collected LIBS spectra resemble.
Fig. 3 illustrates the same procedure for model selection of
3.1. Model selection quartz and Fe-oxides. It is clear that PLS with two latent variables
is sufficient to predict the concentration of quartz and Fe-oxides
Fig. 2 demonstrates the idea of model selection using MSE of the without overfitting.
training and validation data for biotite, glauconite, ferrorichterite Fig. 4 shows the process of model selection for the rest of the
and stilpnomelane. For biotite, galuconite and ferrorichterite min- minerals (K-feldspar, aegirine-augite, actinolite and unknown min-
erals the training error declines rapidly around the latent variables eral groups). The trends of the validation errors indicate that the
number 3 and 4 and then it continues decreasing gradually. At the model cannot properly predict those minerals without overfitting.
same time, the validation error seems to have a significant decay K-feldspar can be justified knowing the fact that two major
around latent variables number three and four (vertical dashed elemental LIBS lines of potassium occur at 766 nm and 769 nm
lines). Eventually, four latent variables were designated to predict wavelengths and those two wavelengths are not in the detection
the concentration of biotite, glauconite and ferrorichterite miner- range of the applied spectrometer in this work. Given Table 1,
als. For stilpnomelane, it can be seen that seven latent variables the concentration of actinolite and unknown mineral groups are
seem to be required in order to do the prediction without overfit- significantly low. Although actinolite seems to be predictable with
ting. Although for many of the optimization algorithms the graph respect to its validation trend, the low concentration causes the
of the validation error is expected to decrease until a minimum PLS method to fail. The justification for failure in predicting
and then increase as the model goes towards overfitting, this issue aegirine-augite concentration is not known yet and it will be
can be justified knowing the fact that in this work the collected further explored.
Biotite Glauconite
1.5 1
Validation
Training Validation
1 Training
MSE
MSE
0.5
0.5
0 0
0 5 10 15 0 5 10 15
No. of latent variables No. of latent variables
Ferrorichterite Stilpnomelane
6 1
Validation
Validation Training
4 Training
MSE
MSE
0.5
2
0 0
0 5 10 15 0 5 10 15
Fig. 2. Selecting the number of latent variables using training and validation errors for biotite, glauconite, ferrorichterite and stilpnomelane.
Quartz Fe-oxides
60 40
Validation Validation
Training 30 Training
40
MSE
MSE
20
20
10
0 0
0 5 10 15 0 5 10 15
Fig. 3. Selecting the number of latent variables using training and validation errors for quartz and Fe-oxides. 2 latent variables are selected for quartz and Fe-oxides.
K-feldspar Aegirine-Augite
1 1.5
Validation Validation
Training Training
1
MSE
MSE
0.5
0.5
0 0
0 5 10 15 0 5 10 15
Actinolite Unknown
0.06 0.15 Validation
Validation Training
Training
0.04 0.1
MSE
MSE
0.02 0.05
0 0
0 5 10 15 0 5 10 15
Fig. 4. Validation and training error for unpredictable minerals.
3.2. Estimation of mineral concentrations from LIBS data through every single spectrum while the model has been produced
through the averaged spectra. There are three spectra per slurry
After performing the model selection, it turned out that three sample and overall 14 samples. Fig. 5 holds parity plots represent-
models are required to predict six out of ten minerals. Biotite, glau- ing the estimation of the grades of biotite, glauconite and ferrorich-
conite, ferrorichterite require four latent variables. The concentra- terite versus the laboratory assay data. R-squared values indicate
tion of quartz and Fe-oxides are estimated using two latent the agreement between modeled data and observed values.
variables and stilpnomelane needs seven. As explained in the pre- The concentration of biotite and glauconite minerals were esti-
vious section, the model is trained using the average of three spec- mated with the R-squared values of 73.9% and 80.6% respectively
tra per sample and validated by leave-one-out cross validation and these values are moderately good for mineral quantification.
technique. Now the concentrations of the minerals are estimated Ferrorichterite mineral was estimated with R-squared value of
Biotite R2 : 73.9% Glauconite R 2 : 80.6% Ferrorichterite R 2 : 69.5%

5 2.5 10
4 2 8
Estimated (%)
Estimated (%)
Estimated (%)
3 1.5 6
2 1 4
1 0.5 2
0 0 0
0 2 4 6 0 1 2 3 0 5 10
Measured (%) Measured (%) Measured (%)
Fig. 5. Comparison of the concentrations estimated using PLS method versus the laboratory assays. Three spectra per each slurry sample and in total 42 spectra were
evaluated for biotite, glauconite and ferrorichterite. Minerals grades were estimated using PLS and 4 latent variables.
Quartz R 2 : 70.0% Fe-oxides R 2 : 82.6% Stilpnomelane R 2 : 72.0%
60 50 5
Estimated (%)
Estimated (%)
Estimated (%)
40 4
50
3
30
40
2
20
30
30 40 50 60 20 30 40 50 2 3 4 5
Measured (%) Measured (%) Measured (%)
Fig. 6. Comparison of the concentrations estimated using PLS method versus the laboratory assays. Three spectra per each slurry sample and in total 42 spectra were
evaluated for quartz, Fe-oxides (hematite/magnetite) and stilpnomelane minerals. Quartz and Fe-oxides were estimated with two latent variables and stilpnomelane with
seven latent variables.
Table 2 also large relative errors due to the same reason. However, given
Summary of the results including: average and standard deviation of the mineral the standard deviations (STD) of the minerals, it can be seen that
concentrations measured by MLA technique in laboratory, R-squared, RMSE and
relative error of the predictions and selected number of latent variables for each
biotite, glauconite and ferrorichterite have quite large variations
mineral. given their average values and this justifies their large relative
errors. It should be also taken into account that mineral contents
Mineral Laboratory Prediction No. of
assays latent
were predicted quantitatively using the smallest possible number
variables latent variables to avoid overfitting. In this work, quartz and
Avg. STD R-squared RMSE Relative Fe-oxides are the most important minerals to be estimated and
(wt.%) (%) error (%) the results are acceptable for slurry samples.
Biotite 2.44 0.97 73.90 0.47 19.26 4
Glauconite 0.75 0.73 80.60 0.26 34.67 4
3.3. Verification of PLS performance and results
Ferrorichterite 4.40 2.70 69.50 0.94 21.36 4
Quartz 46.67 9.94 70.00 4.40 9.43 2
Fe-oxides 36.23 8.06 82.60 2.86 7.90 2 In Section 3.2, PLS method together with NIPALS algorithm
Stilpnomelane 3.30 1.01 72.00 0.42 12.73 7 showed encouraging performance for the rapid, on-stream and
quantitative estimation of mineral contents from high-dimensional
spectra of LIBS technique. However, in this section the reliability
of the results is further investigated. Table 3 demonstrates the
69.5% (Fig. 5). In practice, for slurry samples R-squared values lar- correlation coefficients of 10 minerals in 14 slurry samples
ger than 60% are relatively acceptable and can be used for some represented in Table 1.
process control purposes. The correlation between quartz and Fe-oxides is expected to be
Fig. 6 demonstrates R-squared values of 70.0%, 82.6% and 72.0% strong and negative as more quartz implies less being of Fe-oxides.
for quartz, Fe-oxides and stilpnomelane respectively. It can be This conjecture is validated in Table 3, the correlation coefficient
observed that estimation of the minerals having high abundances between quartz and Fe-oxides is 0.94. The performance of PLS
is more reliable and accurate. The quantities of quartz and algorithm can be justified by evaluating the coefficients in matrix
Fe-oxides were estimated using only two latent variables of PLS B that PLS technique assigns to the variables associated with quartz
model and the coefficient of determination is quite acceptable for and Fe-oxides. Fig. 7 represents the coefficients of PLS model
both. Since stilpnomelane was estimated using 7 latent variables generated using two latent variables.
and the dimension of output Y is 10, there is still the risk of over- For clarity, only variables from 235 nm to 290 nm are displayed
fitting. However, this issue will be further evaluated at the exten- and some popular LIBS lines of Si and Fe are specified in Fig. 7. It
sion of this work with larger number of training samples. can be seen that the weights of the model variables for the two
Table 2 shows a summary of the results and calculations. minerals are in opposite signs as expected. It should be noted that
Relative error can be interpreted as the percentage of the modeling PLS trains the model for each mineral using linear combinations of
error with respect to the average value of the mineral concentra- all the variables. Therefore, even if quartz mineral consists of only
tions. Glauconite has large relative error (34.67%) due to its very silicon and oxygen, the model might find correlation between the
low average concentration (0.75%). Biotite and ferrorichterite have other elements and the target mineral. This explains the positive
Table 3
Correlation coefficients of 10 assayed minerals from 14 tailing samples.
Mineral Biotite Galuconite Ferrorichterite Stilpnomelane K-feldspar Augite Actinolite Unknown Quartz Fe-oxides
Biotite 1.00 0.89 0.12 0.17 0.12 0.09 0.68 0.35 0.16 0.32
Glauconite 0.89 1.00 0.45 0.30 0.06 0.05 0.59 0.06 0.18 0.22
Ferrorichterite 0.12 0.45 1.00 0.76 0.10 0.40 0.14 0.57 0.47 0.19
Stilpnomelane 0.17 0.30 0.76 1.00 0.17 0.52 0.19 0.35 0.63 0.41
K-feldspar 0.12 0.06 0.10 0.17 1.00 0.38 0.05 0.19 0.23 0.27
Aegirine-augite 0.09 0.05 0.40 0.52 0.38 1.00 0.37 0.66 0.84 0.72
Actinolite 0.68 0.59 0.14 0.19 0.05 0.37 1.00 0.48 0.04 0.17
Unknown 0.35 0.06 0.57 0.35 0.19 0.66 0.48 1.00 0.35 0.12
Quartz 0.16 0.18 0.47 0.63 0.23 0.84 0.04 0.35 1.00 0.94
Fe-oxides 0.32 0.22 0.19 0.41 0.27 0.72 0.17 0.12 0.94 1.00
Fe Si Mg
Quartz
1Fe 239.6 251.6 Fe-Oxides 279.5
238.2 Mg Si
Fe
Si 280.26 288.2
240.5
0.5 252.9
Fe
241
a.u
-0.5
-1
240 245 250 255 260 265 270 275 280 285
nm
Fig. 7. Variable coefficients (from 235 nm to 290 nm) generated by PLS using two latent variables for quartz and Fe-oxides.
1 Mg 1 Ca Ca Glauconite
Glauconite
Mg 396.8 Ferrorichterite
279.5 Ferrorichterite 393.4
285.1 Biotite
Biotite
Mg Mg 0.5
Si Ca
0.5 280.2 286 288.2 422.7
a.u
a.u
0 -0.5
-1
-0.5
280 285 290 295 300 395 400 405 410 415 420
nm nm
Fig. 8. Variable coefficients (from 275 nm to 305 nm (left) and 390 nm to 425 nm (right)) generated by PLS using four latent variables for glauconite, ferrorichterite and
biotite.
and large coefficients of magnesium at 279.5 nm and 280.26 nm concentrate on online and quantitative analysis of iron-oxide min-
for quartz mineral. erals such as hematite and magnetite.
Fig. 8 represents the coefficients of the model variables for glau-
conite ((K,Na)(Fe3+,Al,Mg)2(SiAl)4O10(OH)2), ferrorichterite (Na
(CaNa)Fe2þ5 [Si8O22](OH)2) and biotite (K(Mg,Fe)3(AlSi3O10)(F,OH)2) References
minerals in the wavelength ranges from 275 nm to 305 nm
(Fig. 8-left) and from 390 nm to 425 nm (Fig. 8-right). Those ranges Awan, M.A., Ahmed, S.H., Aslam, M.R., Qazi, I.A., Baig, M.A., 2013. Determination of
heavy metals in ambient air particulate matter using laser-induced breakdown
were selected since the best lines of Ca and Mg elements will occur
spectroscopy. Arabian J. Sci. Eng. 38 (7), 1655–1661.
within them. The detector used in this study had the detection Ayyalasomayajula, K.K., Yu-Yueh, F., Singh, J.P., McIntyre, D.L., Jain, J., 2012.
range from 190 nm to 507 nm. Potassium (K) and Na did not have Application of laser-induced breakdown spectroscopy for total carbon
any significant line within that detection range. Regardless of K quantification in soil samples. Appl. Opt. 51 (7), B149–B154.
Beckhoff, B., Kanngießer, B., Langhoff, N., Wedell, R., Wolff, H., 2007. Handbook of
and Na, it is expected that ferrorichterite can be distinguished from Practical X-ray Fluorescence Analysis. Springer Science & Business Media.
glauconite and biotite mainly using Ca and Mg lines according the Brech, F., Cross, L., 1962. Optical microemission stimulated by a ruby laser. Appl.
their chemical formula. Fig. 8-left illustrates the positive coeffi- Spectrosc. 16 (2), 59.
Chang, C.-I., 2003. Hyperspectral Imaging: Techniques for Spectral Detection and
cients assigned to the variables associated with Mg for glauconite Classification, vol. 1. Springer Science & Business Media.
and biotite while negative coefficients for ferrorichterite. Fig. 8- Clegg, S.M., Sklute, E., Dyar, M.D., Barefield, J.E., Wiens, R.C., 2009. Multivariate
right similarly demonstrates the positive coefficients assigned to analysis of remote laser-induced breakdown spectroscopy spectra using partial
least squares, principal component analysis, and related techniques.
Ca for ferrorichterite and negative coefficients for glauconite and Spectrochim. Acta Part B: Atom. Spectrosc. 64 (1), 79–88.
biotite as expected. Such evaluation adds to the validity of the Cremers, D.A., Radziemski, L.J., 2006. Handbook of Laser-induced Breakdown
model trained using PLS technique for the purpose of quantitative Spectroscopy. Cambridge University, p. 25.
Das, R.S., Agrawal, Y., 2011. Raman spectroscopy: recent advancements, techniques
mineral analysis. and applications. Vibrat. Spectrosc. 57 (2), 163–176.
Díaz Pace, D.M., Gabriele, N.A., Garcimuño, M., D’Angelo, C.A., Bertuccelli, G.,
Bertuccelli, D., 2011. Analysis of minerals and rocks by laser-induced
breakdown spectroscopy. Spectrosc. Lett. 44 (6), 399–411.
4. Conclusion Eilers, P.H., Boelens, H.F., 2005. Baseline Correction with Asymmetric Least Squares
Smoothing. Leiden University Medical Centre Report.
Filippov, L., Severov, V., Filippova, I., 2014. An overview of the beneficiation of iron
This study focused on the quantitative mineral identification ores via reverse cationic flotation. Int. J. Mineral Process. 127, 62–69.
from the data produced by a LIBS measurement technique Gaft, M., Sapir-Sofer, I., Modiano, H., Stana, R., 2007. Laser induced breakdown
spectroscopy for bulk minerals online analyses. Spectrochim. Acta Part B: Atom.
designed for on-stream slurry analysis in flotation. The goal was
Spectrosc. 62 (12), 1496–1503.
to provide knowledge about mineral contents of the slurries in Geladi, P., Kowalski, B.R., 1986. Partial least-squares regression: a tutorial. Anal.
addition to the elemental contents to improve the process control Chim. Acta 185, 1–17.
and have better recovery. PLS technique as a multivariate statisti- Goncharov, A.F., Struzhkin, V.V., 2003. Raman spectroscopy of metals, high-
temperature superconductors and related materials under high pressure. J.
cal technique along with NIPALS algorithm was presented to be Raman Spectrosc. 34 (7–8), 532–548.
suitable for predicting the concentration of the minerals in the iron Hyötyniemi, H., 2001. Multivariate Regression: Techniques and Tools. Helsinki
ore flotation slurry. Three models were trained to estimate the University of Technology.
Khajehzadeh, N., Kauppinen, T.K., 2015. Fast mineral identification using elemental
grade of six minerals with acceptable R-squared values. Quartz libs technique. IFAC-PapersOnLine 48 (17), 119–124.
and Fe-oxides were the most essential minerals to be predicted Khater, M.A., 2013. Laser-induced breakdown spectroscopy for light elements
from the slurry and the proposed method turned out to be quite detection in steel: state of the art. Spectrochim. Acta Part B: Atom. Spectrosc.
81, 1–10.
consistent for them. However, the relative errors of biotite, glau- Lamberg, P., Vianna, S., 2007. A technique for tracking multiphase mineral particles
conite and ferrorichterite were large because of their low concen- in flotation circuits. In: Proceedings of the 7th Meeting of the Southern
trations and high variations within the sampled slurries. Increasing Hemisphere on Mineral Technology, Ouro Preto, Brazil, pp. 20–24.
Lanza, N.L., Wiens, R.C., Clegg, S.M., Ollila, A.M., Humphries, S.D., Newsom, H.E.,
the number of samples will enhance the trained models and accu-
Barefield, J.E., 2010. Calibrating the chemcam laser-induced breakdown
racy of the predictions. spectroscopy instrument for carbonate minerals on mars. Appl. Opt. 49 (13),
The main achievement of this work was extracting mineralogi- C211–C217.
Lund, C., Lamberg, P., Lindberg, T., 2013. Practical way to quantify minerals from
cal information by automatically utilizing the information on all
chemical assays at malmberget iron ore operations–an important tool for the
the important spectral lines in the LIBS measurements. This tech- geometallurgical program. Miner. Eng. 49, 7–16.
nique is rapid and gives the possibility of online mineralogical McCanta, M., Dobosh, P., Dyar, M., Newsom, H., 2013. Testing the veracity of libs
analysis without any modification of the conventional LIBS setup. analyses on mars using the libssim program. Planet. Space Sci. 81, 48–54.
McMillan, N.J., Harmon, R.S., De Lucia, F.C., Miziolek, A.M., 2007. Laser-induced
Moreover, the presented method could overcome the restrictions breakdown spectroscopy analysis of minerals: carbonates and silicates.
come from the slurry and aqueous samples. The future work will Spectrochim. Acta Part B: Atom. Spectrosc. 62 (12), 1528–1536.
Outotec, 2015. Outotec courier SL on-line analyzers, brochure. <http://www. Wiens, R., Maurice, S., Lasue, J., Forni, O., Anderson, R., Clegg, S., Bender, S., Blaney,
outotec.com/ImageVaultFiles/id_1607/d_1/cf_2/OTE_outotec_courier_sl_online_ D., Barraclough, B., Cousin, A., 2013. Pre-flight calibration and initial data
analyzers_brochure-l.PDF>. processing for the chemcam laser-induced breakdown spectroscopy instrument
Rakovskỳ, J., Musset, O., Buoncristiani, J., Bichet, V., Monna, F., Neige, P., Veis, P., on the mars science laboratory rover. Spectrochim. Acta Part B: Atom.
2012. Testing a portable laser-induced breakdown spectroscopy system on Spectrosc. 82, 1–27.
geological samples. Spectrochim. Acta Part B: Atom. Spectrosc. 74, 57–65. Wold, H., 1985. Partial least squares. In: Encyclopedia of Statistical Sciences. Sknl
Rapiscan Systems, 2005. SATuration MAGnetization ANalyser (SATMAGAN S135). Johnson.
<http://www.indehk.com/PDF/SatmagandescriptionandIIinfoOct2005.pdf>. Wold, S., Sjöström, M., Eriksson, L., 2001. Pls-regression: a basic tool of
Saarhelo, K., Paakkinen, U., Pennanen, P., 1988. On-line analysis in industrial chemometrics. Chemometr. Intell. Lab. Syst. 58 (2), 109–130.
mineral applications. In: 8th Industrial Minerals Inst. Congress, Boston. Yu, X., Li, Y., Gu, X., Bao, J., Yang, H., Sun, L., 2014. Laser-induced breakdown
Sirven, J.-B., Bousquet, B., Canioni, L., Sarger, L., 2006. Laser-induced breakdown spectroscopy application in environmental monitoring of water quality: a
spectroscopy of composite samples: comparison of advanced chemometrics review. Environ. Monit. Assess. 186 (12), 8969–8980.
methods. Anal. Chem. 78 (5), 1462–1469. Zhu, X., Xu, T., Lin, Q., Liang, L., Niu, G., Lai, H., Xu, M., Wang, X., Li, H., Duan, Y., 2014.
Warren, B.E., 1969. X-ray Diffraction. Courier Corporation. Advanced statistical analysis of laser-induced breakdown spectroscopy data to
Whiten, B., 2007. Calculation of mineral composition from chemical assays. Miner. discriminate sedimentary rocks based on czerny–turner and echelle
Process. Extract. Metall. Rev. 29 (2), 83–97. spectrometers. Spectrochim. Acta Part B: Atom. Spectrosc. 93, 8–13.
Publication V
Khajehzadeh, N., Haavisto, O. and Koresaar, L. On-stream mineral iden-

tification of tailing slurries of an iron ore concentrator using data fusion of
LIBS, reflectance spectroscopy and XRF measurement techniques. Miner-
als Engineering, 113, 83-94, https://doi.org/10.1016/j.mineng.2017.08.007,
August 2017.

c 2017 Elsevier Ltd.
115
0LQHUDOV(QJLQHHULQJ ²
Minerals Engineering
journal homepage: www.elsevier.com/locate/mineng
On-stream mineral identification of tailing slurries of an iron ore

concentrator using data fusion of LIBS, reflectance spectroscopy and XRF
measurement techniques
⁎
Navid Khajehzadeha, , Olli Haavistob, Lauri Koresaarb
a
Aalto University, School of Electrical Engineering, Department of Electrical Engineering and Automation, Finland
b
Outotec (Finland) Oy, Rauhalanpuisto 9, 02230 Espoo, Finland
A R T I C L E I N F O A B S T R A C T
Keywords: This article is extension of the earlier work (Khajehzadeh et al., 2016), where quantitative mineralogical in-
Laser-induced breakdown spectroscopy formation of slurry samples was achieved using an on-stream LIBS analyzer. Despite the great advances in the
X-ray fluorescence analytical methods and laser-based measurement techniques, the industrial developers are still demanding novel
Reflectance spectroscopy ideas enabling differentiation between minerals having similar elemental contents such as hematite and mag-
Data fusion
netite or silicon-bearing minerals such as quartz and other mixed silica minerals since they have different flo-
Mineral interpretation
tation properties. The available analytical techniques for LIBS spectral analysis (including the earlier work of this
Partial least squares
research) could not distinguish between such minerals with identical elemental contents. This work at first
presents data fusion of LIBS and reflectance spectroscopy and then discusses the data fusion of reflectance
spectroscopy and X-ray fluorescence (XRF) measurement techniques operating on the same slurry samples. The
results will show that such data integrations enable on-stream and quantitative identification of slurry mineral
contents specially for hematite, magnetite, quartz and ferrorichterite which are important minerals in iron ore
beneficiation.
1. Introduction difficulties of quantitative and online mineral identification of slurries

(Khajehzadeh et al., 2016). Here we shortly address the issue again to
Climate and environmental changes have increased the concerns introduce the structure of this study as the extension of the previous
about natural and specifically water resources. Mineral concentration work.
processes typically consume large amounts of water and therefore every In order to optimize the process control, mineralogical information
action resulting in saving water and energy is an advantage. For ex- is necessary. However, on-stream and close-to-real time mineral iden-
ample, one of these water consuming processes is flotation. It is ex- tification is challenging. Developing new measurement setups only for
pensive, complex and usually slimes influence the process. In con- the purpose of mineral identification is costly, requires subtle calibra-
centrator plants where the quality of ore is poor the magnetic and tion and needs proper space for installation in the plant. Therefore,
gravity separations are usually followed by a flotation process. It is very quantitative identification of mineral contents through the elemental
advantageous to bypass the flotation when the grade from magnetic assays collected from the streams of the process is efficient and eco-
separation is high enough. Additionally, the flotation process of quartz nomic. The concept of mineral estimation from the chemical assays is
has different properties from other silicon-bearing minerals. Therefore, usually expressed by the term elemental-to-mineral conversion.
accurate and online measuring of mineral grades of slurry samples in Lamberg and Vianna (2007) estimate mineral quantities from elemental
flotation processes is quite beneficial. contents using minimization of a set of mass balancing equations.
As lasers represent rapid and high-tech measurement techniques, However, their techniques are still slow for on-stream measurement of
the laser-based methods have recently become of increasing importance the minerals and are not considered as automated techniques. Direct
in the mining industry. However, the majority of the online and on- mineralogical measurement techniques like Raman spectroscopy (Das
stream analytical techniques is based on elemental analysis or per- and Agrawal, 2011), X-ray diffraction (XRD) (Warren, 1969) and hy-
formed offline when applied to mineralogical identification. The pre- perspectral imaging (Chang, 2003) are mainly applied in laboratories or
vious work of this research thoroughly discussed the importance and there are substantial limitations to their online usage (Goncharov and
⁎
Corresponding author.
E-mail addresses: navid.khajehzadeh@aalto.fi (N. Khajehzadeh), olli.haavisto@outotec.com (O. Haavisto), lauri.koresaar@outotec.com (L. Koresaar).
http://dx.doi.org/10.1016/j.mineng.2017.08.007
Received 12 December 2016; Received in revised form 22 July 2017; Accepted 15 August 2017
(OVHYLHU/WG$OOULJKWVUHVHUYHG
N. Khajehzadeh et al. 0LQHUDOV(QJLQHHULQJ ²
Struzhkin, 2003; Saarhelo et al., 1988). Several researchers have re- This work is structured as follows. First, the results of the earlier
ported feasibility of quantitative mineralogical analysis using re- work (LIBS mineralogical analysis) are presented briefly. Then, the
flectance data obtained from visible and near infrared to shortwave possibility of quantitative mineral identification using reflectance
infrared (VNIR and SWIR) wavelengths, (Haest et al., 2012a and Haest spectroscopy is evaluated. Next, data fusion is performed on the re-
et al., 2012b). flectance and LIBS spectra and the improvement of the mineral esti-
Laser-induced breakdown spectroscopy (LIBS) (Cremers and mation will be demonstrated. Then, the data related to two channels of
Radziemski, 2006) and X-ray fluorescence (XRF) (Beckhoff et al., 2007) iron peaks coming from the K-alpha and K-beta lines of XRF spectra are
measurement techniques are two elemental techniques widely applied recorded. After that, those two XRF intensities of iron peaks are ap-
in the flotation processes (Outotec, 2015). Cooper (1976) reported a pended to the reflectance spectra (data fusion of XRF and reflectance
centralized XRF analyzer for direct elemental analysis of several slurry spectroscopy) and finally significant improvement in the mineral esti-
lines. The term centralized stands for a separate sampler which samples mations of the slurry samples will be presented.
the slurry lines and brings them sequentially to one XRF analyzer for
analysis. The idea of centralized XRF reduced the cost of instrumenta- 2. Experimental
tion, but increased the sampling time of successive measurements from
the same slurry streams when there are several slurry lines. A con- 2.1. Slurry samples
tinuous monitoring of the slurry elemental contents was reported by
Haavisto and Hyötyniemi (2011). They demonstrated a fusion of XRF A set of 14 slurry samples was available from the tailings stream of a
and visible and near-Infrared (VNIR) reflectance spectroscopy func- flotation plant where the plant belonged to an iron ore concentrator.
tioning on the same slurry samples. The iron ore concentrator was operating by magnetic and gravity se-
LIBS has a very simple spectroscopic setup for implementation and paration processes. Additionally, the processes were followed by a re-
provides more emission lines per elements than XRF. It has better verse cationic flotation of quartz (gangue mineral) (Filippov et al.,
performance than XRF when measuring elements lighter than titanium 2014). From the process control point of view, it would be interesting to
(Khater, 2013). LIBS technique operates using high energy of laser know whether the grades of iron bearing minerals (mainly magnetite
pulses to excite small areas on the samples. When LIBS is applied to and hematite) after magnetic separation are high enough to bypass the
aqueous samples like slurry, majority of the liquid is vaporized by the flotation. Also the online analysis of silicon-bearing minerals would be
high energy of laser pulse and the elements present in plasma emit light useful since quartz behaves differently in flotation than other mixed
with unique wavelengths (Yu et al., 2014). The most common excita- silica minerals. Even though only tailing samples are analyzed in this
tion wavelength is 1064 nm because of higher temperature of plasma. study, the results should indicate also the performance of the proposed
Several researchers have reported advantages and feasibility of shorter methods in the other slurry streams at the plant. The slurry samples for
wavelengths of LIBS analysis, i.e., 532 nm and 266 nm (Russo et al., laboratory tests were generated by mixing the dry samples with 6 L of
2000; Barnett et al., 2008). water. The solid content of the tailing samples was on average 21% (w/
In (Khajehzadeh et al., 2016) we applied a statistical method called w).
Partial Least Squares (PLS) along with Non-Linear Iterative Partial Least Prior to the mineral estimation of the measured spectra, training
Squares (NIPALS) algorithm (Geladi and Kowalski, 1986) to map the data is required for modeling. Therefore, Mineral Liberation Analyzer
elemental lines of LIBS spectra to the mineral contents of the slurry (MLA) was used to produce the mineralogical assays of the slurry
samples. Three models were trained to estimate the grade of six mi- samples. SATuration MAGnetization ANalayzer (SATMAGAN) tech-
nerals with acceptable R-squared values. Quartz and Fe-oxides were the nique was applied to measure the concentration of magnetite (Rapiscan
most essential minerals to be predicted from the slurries. However, LIBS Systems, 2005). It is worth noting that the samples under study are the
technique is incapable of distinguishing between minerals with similar same samples presented in the previous work of this research (See:
elemental contents such as hematite and magnetite, thus we measured Khajehzadeh et al., 2016 Table 1].
their collection as Fe-oxides. Table 1 represents cross-correlation between elemental and mineral
Haavisto et al. (2006) reported the visible and near-infrared (VNIR, contents of the slurry samples. The elemental assays have been done
400–1000 nm) reflectance spectra of the froths and slurries as re- using XRF technique in laboratory. It can be seen that quartz and
presentative of elemental contents. The capability of reflectance spec- magnetite have strong correlation with total iron content. On the other
troscopy to discriminate the iron-bearing minerals will be discussed at hand, ferrorichterite correlates with CaO and MgO. Hematite has a
Section 3.2. slight correlation with CaO and MgO but does not correlate with FeO
As reported by Khajehzadeh et al. (2016), LIBS was unable to dif- and SiO2.
ferentiate between hematite and magnetite. This research discusses
capability of reflectance spectroscopy technique to identify hematite 2.2. Laser-induced breakdown spectroscopy
and magnetite contents quantitatively. Then, the estimations are im-
proved by adding XRF data to the model. Therefore, this study extends 2.2.1. LIBS measurement setup
the online mineralogical assay of slurry samples by fusion of LIBS, re- The measurement setup of LIBS system in this research was a la-
flectance spectroscopy and XRF spectra. All measurements are per- boratory prototype of Courier 8 SL (Outotec, 2015). The laser used was
formed on the same slurry samples. an ULTRA 100 Q-switched Nd:YAG laser operating at 1064 nm with
120 mJ/pulse and up to 10 pulses/s repetition rate. In order to avoid
Table 1 continuum radiation in the cooling incandescent plasma, an aquisition
Cross-correlation coefficients between minerals and elements. delay time of 1 μ s and gate time of 233 μ s was applied to the detector.
The detector is a Peltier cooled CCD camera attached to an echelle
spectrograph with resolution of 32 pm (FWHM) for Si line at
288.16 nm. Fig. 1 shows the same schematic diagram of the LIBS ana-
lyzer as in the previous work Khajehzadeh et al. (2016) while the lo-
cation of the stabilizer has been depicted here. The slurry is circulated
into the measurement cell using a pump and the LIBS system scans one
fixed point on the stabilized slurry surface.
In order to overcome the instabilities in the LIBS spectra stemming
from the intrinsic characteristics of liquids such as shaking, motion and

Slurry flow performed on the matrix of independent variables.

Plasma Fig. 2 represents the LIBS spectra prior and after preprocessing. It
LIBS probe can be seen that after preprocessing (elimination of negative and zero
Stabilizer PC
surface wavelengths and peak finding procedure) the number of LIBS lines has
Detector
been reduced to 111 and the variables are scaled so that they all have
Measurement
cell similar importance prior to modeling. Therefore, the x-axis of the pre-
Laser
Power processed spectra shown in Fig. 2 is representative of the selected
unit variable numbers from 1 to 111 (not the original wavelengths (nm)).
Return
vessel
2.3. Reflectance spectroscopy
Slurry
Infrared reflectance spectroscopy utilizes the spectral signatures

resulting from interaction of electromagnetic radiation with sample
Tubing
material, which are due to the physicochemistry and relative abun-
Pump dances of infrared-active mineral species, their grain size and the
Fig. 1. Laboratory measurement setup for the mineralogical LIBS slurry analysis. viewing geometry (Clark, 1999). In mineralogy the main concept of
optical reflectance spectral analysis is absorption so that different mi-
nerals have distinguished absorption properties. The wavelength range
splashes during measurement and to compensate the light energy ab-
of infrared radiation is from red edge of visible (700 nm) to 1 mm.
sorbed by water, 200 laser pulses for each sample were accumulated.
Infrared radiation is further divided into near infrared (NIR), short-,
Such accumulation mitigates the noise and exposes the elemental
mid-, and long-wavelength infrared (SWIR, MWIR, LWIR) and far-in-
peaks. This procedure was repeated three times and the three accu-
frared radiation. The reflectance spectroscopy in the visible and near-
mulated spectra were averaged to generate the final LIBS spectrum of
infrared ranges is a common practice since the ordinary silicon-based
the sample. Then, the data pre-processing is performed.
sensors can be used for light detection (D’Amico et al., 2009).
2.2.2. LIBS data preprocessing 2.3.1. Reflectance spectroscopy laboratory setup

First, the LIBS spectra are baseline corrected by an asymmetric least In order to enhance the performance of the on-stream slurry mineral
squares smoothing algorithm (Eilers and Boelens, 2005). The detection identification, the slurry samples measured by LIBS technique were
range of the spectrometer is from 190 nm to 507 nm and each LIBS scanned by a reflectance spectroscopy analyzer. The source of illumi-
spectrum contains about 53,480 wavelengths (variables). Such a large nation is a halogen lamp (Fig. 3). The detection system was a spectro-
number of variables causes collinearity or singularity in the matrix of graph consisting of a Specim Imspector V10 spectrophotometer at-
independent variables when training the model. Moreover, majority of tached to a Basler 0102f greyscale camera. A collimation lens is used to
those variables (wavelengths) are zero or negative values and are collect the reflected light and transfer it to the spectrograph using an
considered as redundant variables. Therefore, the wavelengths with optical fiber.
negative and zero intensities are eliminated. Next, a peak finding al-
gorithm is applied on the spectra so that local maxima which are sig- 2.3.2. reflectance data preprocessing
nificantly greater than the values around are specified and the rest are Each reflectance spectrum, collected with the above described
eliminated. The peak finding is performed using the basic principle of setup, contains 960 intensity values from 400 nm to 1000 nm. In order
first-derivative while a minimum peak distance threshold is also spe- to reduce the dimension, each spectrum is down-sampled by selecting
cified to ignore those peaks being in the close neighborhood of the local every rth point from the signal. r is selected so that the dimension of the
maxima. At the end the mean centering and variance scaling is reflectance spectra are similar or close to the dimension of LIBS spectra
Fig. 2. Accumulated LIBS spectra of 14 slurry samples prior and after

Original accumulated LIBS spectra
pre-processing. “a.u” indicates arbitrary unit.
Intensities(a.u)
10000
5000
0
200 250 300 350 400 450 500
Wavelength(nm)
LIBS spectra after preprocessing

10
Intensities(a.u)
8
6
4
2
0 20 40 60 80 100
Variable number

Measurement nλ = 2dsinθ (4)

winĚŽw where n (an integer) is the order of reflection, λ is the wavelength of the
IlluminaƟŽn radiation, d is the spacing between the crystal planes and θ is the angle
&lŽw between the incident rays and the surface of the crystal. The radiation
Slurry cell of the isolated wavelength is the measured using a proportional
circulaƟŽn SpectrŽŐraph counter. In the EDXRF method the fluorescent X-rays coming from the
pump ŽllimaƟŽn sample are measured without separation. The X-rays are measured by
lens the amount of ionization it produces on a semi-conductor material used
as the detector. The WDXRF generally has higher resolution and can
measure lower elemental grades, while the EDXRF is able to measure
multiple elements simultaneously.
Fig. 3. Laboratory measurement setup for reflectance spectroscopy of slurry samples. 2.4.1. X-ray laboratory setup
An Outotec Courier 6 SL slurry analyzer (Outotec, 2015) was used
for the laboratory measurement. A schematic layout of the analyzer
(111). The length of output signal will be equal to the length of original
setup is presented in Fig. 6. The analyzer was equipped with a 200 W X-
signal divided by r, (length(x)/r). To avoid aliasing of signal (distortion
ray tube with Mo anode. Peltier cooled EDXRF channel with a resolu-
of signal when the signal is reconstructed), the resampling of spectra is
tion of 125 eV for Mn Kα line was used for the test. The samples were
preceded by lowpass filtering using a Chebyshev type 1 filter with
circulated through the measurement cell using a pump and an agitator.
normalized cutoff frequency of 0.8/r and passband ripple 0.05 dB.
120 s counting time was used for each sample, and the final result was
According to Beer-Lambert law, the concentration of absorbing
presented as counts/s.
particles is in linear relation with the absorbance coefficient:
R = IR/ I0, (1) 2.4.2. XRF data preprocessing
Fe fluorescence radiation lines, Kα1 at 6.40 keV and Kβ1 at 7.06 keV,
A = log (1/R), (2) can be seen in the spectra (Fig. 6). The spectra also contain elastic and
inelastic scattering radiation of Mo Kα and Kβ at 17.48 and 19.61 keV,
A∝c (3)
respectively. The scattered radiation is resulting from the primary ra-
where IR and I0 are reflected and incident intensities respectively. R is diation of the Mo anode in the X-ray tube. The spectra also contain Sn
reflectance, A is defined as absorbance and c indicates the concentra- Kα and Kβ lines at 25.27 and 28.49 keV, respectively. The Sn signal is
tion. Therefore, the reflectance spectra are transformed to absorbance caused by the material used for the protective window holder.
using Eq. (2). Partial least squares regression technique seeks a linear X-ray fluorescence emission intensity decreases with decreasing
model mapping the input variables to the output. Therefore, the data atomic number. Emission of elements below Ca get completely ab-
should be pre-processed so that the relation between input and output is sorbed by the water in the slurry, protective windows and air in the
linear. Beer-Lambert law and Eq. (3) is an empiric way of such linear- spectrometer. Therefore, due to the lack of elemental peaks in the XRF
ization. spectra, XRF technique was used as a supplementary method for the
Fig. 4-up shows the raw reflectance spectra of 14 slurry samples and reflectance spectroscopy technique. The Kα and Kβ iron peaks were
Fig. 4-down shows the measured reflectance spectra converted to ab- selected by averaging the data around 6.40 keV and 7.06 keV energy
sorbace after signal resampling. It is important to note that the as- lines respectively.
sumptions of the Beer-Lambert law do not usually hold in practice. For
example, in this case the thickness of the sample and the scattering
2.5. Data modeling using partial least squares regression
effects caused by the granular sample violate the assumptions (Osborne
et al., 1993). However, the transformation is still improving the mod-
2.5.1. Partial least squares regression
eling of the reflectance spectrum data as demonstrated in the results
Suppose X denotes an m × n matrix of independent variables where
section (Table 3).
m is the number of samples and n is the number of LIBS and reflectance
spectra wavelengths together. Y(m × p) is the matrix of dependent vari-
2.4. X-ray fluorescence ables where p denotes the number of minerals to be predicted. The
purpose is to train a model estimating the concentration of the minerals
X-ray fluorescence (XRF) has been a dominant elemental analysis in each slurry sample through the matrix of the measured spectra. The
technique for slurry analysis, since 1960s (Carr-Brion, 1989). XRF uti- problem is formulated by constructing input (X) and output (Y) ma-
lizes electromagnetic radiation in the X-ray band (photon energy trices. As an example, for combined LIBS and reflectance spectroscopy
>0.1 keV). Primary radiation from an X-ray source is directed to a the matrix of inputs for 14 slurry samples is formed as X(14 × (111 + 120) ) and
sample, where the primary radiation excites electrons in the inner shells matrix of outputs for 4 minerals will be Y(14 × 4 ) .
of the atoms and causes loss of electron in the shell. Electrons in higher Partial Least Squares (PLS) regression is the technique applied for
energy shells can then fill this vacancy in lower energy shells by mapping the measurements to the mineral concentrations. PLS is a
emitting a photon at a wavelength corresponding to the energy differ- statistical method introduced by Wold (1985). It is often applied for
ence of the electron shells. As every element has a unique set of energy analysis of high dimensional spectral measurements and widely applied
levels, the elements in the sample can be determined by measuring the in chemometrics (Gemperline, 2006). The most important advantage of
radiation at the characteristic energies. PLS compared with other principal component-based methods (e.g.
The X-rays can be generated using a radioactive source or an X-ray principal component regression) is that instead of finding hyperplanes
tube, which generates the x-rays electrically. The X-ray tube can be based on the maximum variance of inputs to overcome the problem of
turned off when analysis is not needed. The radiation from the sample collinearity, it considers the correlation between both inputs and out-
can be measured using wavelength dispersive (WDXRF) or energy dis- puts and projects both independent and dependent variables to a new
persive (EDXRF) methods (Fig. 5). In the WDXRF method the char- space where the correlation between them is maximized. PLS is usually
acteristic radiation is isolated using a crystal diffracting only the desired explained using the eigenvector decomposition approach, but in prac-
wavelength according to Bragg’s law: tice PLS is implemented by an iterative procedure called non-linear

Measured reflectance spectra Fig. 4. Up, raw reflectance spectra. Down, reflectance
0.045 converted to absorbance after resampling. “a.u” indicates
arbitrary unit.
0.04
Intensities(a.u)
0.035
0.03
0.025
0.02
0.015
0.01
400 500 600 700 800 900 1000 1100
Wavelength(nm)
Absorbacne after resampling

1.9
1.8
Absorbance (a.u)
1.7
1.6
1.5
1.4
1.3
0 20 40 60 80 100 120
Variable number
2.5.2. Model selection and validation

When the problem is about on-stream mineral identification, the
number of samples is often considerably less than variables. Providing
more samples and more measurements for training data is time-con-
suming and expensive. Therefore, the problem of m < n is almost al-
ways plausible for on-stream applications. Accordingly, instead of using
conventional regression techniques where all variables are taken into
account when training the models, methods such as PLS with the
property of dimensionality reduction are encouraged. Thus, we define
the complexity of the model by the number of latent variables in the
PLS model. The number of latent variables should be selected cau-
tiously because large number of latent variables will easily cause over-
fitting of the model.
Leave-one-out is the cross-validation method used for model selec-
tion. At each iteration of the algorithm, one spectrum is left out for
validation and 13 spectra are employed for training. The algorithm runs
14 times and the mean squared error (MSE) of both training (13 spectra
at each iteration) and validation data (left-out spectrum at each itera-
tion) are calculated for each mineral separately. Then, the training and
validation errors are calculated while the number of latent variables are
increasing from 1 to m. By increasing the latent variables, continuous
decrease of the training error is expected while the validation error is
expected to have an initial decrease and then followed by a rise since
the model is over-fit and is unable to estimate the unseen data.
After selecting the model with the right number of latent variables,
Fig. 5. Schematic layout of the XRF laboratory setup. for each left-out spectrum the R-squared value (coefficient of determi-
nation) of the model prediction is calculated. Finally, the average of 14
iterative PLS (NIPALS) (Hyötyniemi, 2001). The previous phase of this R-squared values coming from 14 iterations is reported as the perfor-
research ((Khajehzadeh et al., 2016), Section 2.6), provided more de- mance of the model for each mineral. It is worth to note that R-squared
tailed explanation of PLS and NIPALS algorithms. value explains the percentage of the response variable variation that is
explained by a linear model.

Fig. 6. Raw XRF spectra.
3. Results magnetite, hematite and ferrorichterite (Na(CaNa)Fe52 +[Si8O22](OH)2).

Quartz and ferrorichterite have clear patterns for model selection since
In this chapter, the results of LIBS technique for quantitative mi- the early slump in the validation error is instantly followed by an as-
neral identification is briefly discussed. Then the capability of re- cending steep slope. The model selection for magnetite needs further
flectance spectroscopy technique for mineralogical analysis is pre- evaluation because the decay in the validation error is followed by a
sented. Next, the data fusion of LIBS and reflectance spectroscopy is gradual growth. For hematite there is no determined model since the
discussed. After that, the fusion of reflectance spectra and the two-iron decline in the validation error is not deep enough and the subsequent
channels of XRF measurement technique will be proposed. increase of the error is momentary. However, in order to compare the
results of the fused LIBS and reflectance with only LIBS, two latent
variables are selected for magnetite model and hematite model was
3.1. Quantitative mineral identification using LIBS technique trained by four latent variables according to Fig. 7. Two latent variables
were selected for quartz and three latent variables for ferrorichterite.
Fig. 7 represents the procedure of selecting the number of latent Fig. 8 demonstrates the parity plots of the estimated LIBS mineral
variables for PLS model of LIBS spectra. The model selection is per- contents versus the laboratory assays. It is clear that the estimations are
formed using MSE of the LIBS training and validation data for quartz,
Fig. 7. Model selection for LIBS spectra. Selecting the number of latent
Quartz Magnetite
40 variables using training and validation errors for ferrorichterite, quartz,
60
magnetite and hematite.
Validation
30 Training
40
MSE
MSE
20
20
10
0 0
0 5 10 15 0 5 10 15
Hematite Ferrorichterite
4 6
3
4
MSE
MSE
2
2
1
0 0
0 5 10 15 0 5 10 15

Fig. 8. Comparison of the concentrations estimated using LIBS spectra

Quartz R 2 : 69% Magnetite R 2 : 78% versus the laboratory assays. Mineral grades were estimated using PLS and
80 60
two latent variables for quartz and magnetite, three latent variables for
Estimated (%)
Estimated (%)
ferrorichterite and four latent variables for hematite.
60
40
40
20
20
0 0
0 20 40 60 80 0 20 40 60
Measured (%) Measured (%)
Hematite R 2 : 23% Ferrorichterite R 2 : 52%

10 10
Estimated (%)
Estimated (%)
5 5
0 0
0 5 10 0 5 10
Fig. 9. Model selection of reflectance spectroscopy. Selecting the number

Quartz Magnetite
30 20 of latent variables using training and validation errors for ferrorichterite,
Validation quartz, magnetite and hematite.
15 Training
20
MSE
MSE
10
10
5
0 0
0 5 10 15 0 5 10 15
2 4
1.5 3
MSE
MSE
1 2
0.5 1
0 0
0 5 10 15 0 5 10 15
in accordance with the model validations for quartz, hematite and apply parameterization of reflectance spectra to identify Fe-oxides and
magnetite, e.g. the model validation failed for hematite. Although Fig. 7 also hematite. Given the capability of reflectance spectroscopy for
represents a clear model for ferrorichterite by three latent variables, measuring hematite, the performance of the technique for miner-
MSE value of validation data is large compared to the concentration of alogical analysis is evaluated in this section.
ferrorichterite. It is important to note that in this work the estimations Fig. 9 represents the PLS model selection through the spectra ob-
have been performed using one spectrum per sample and the average of tained by reflectance spectroscopy technique. Three latent variables
the R-squared values for fourteen iterations of the leave-one-out algo- were selected for magnetite and ferrorichterite and two latent variables
rithm is reported as the accuracy of the model estimations. for quartz. Perhaps two latent variables were sufficient for hematite.
Fig. 10 represents the quantitative estimations of the slurry mineral
3.2. Quantitative mineral identification using reflectance spectroscopy contents using reflectance spectroscopy technique versus the laboratory
technique assays. While LIBS technique alone could not enable mineralogical es-
timations of hematite and the PLS model for ferrorichterite was trained
There are numerous reports about the capability of reflectance with some uncertainty, Fig. 10 demonstrates the capability of re-
spectroscopy for detection of hematite. Scheinost et al. (1998) reports a flectance spectroscopy for online mineralogical assaying of the slurry
diffuse reflectance spectroscopy of soil samples and identifies Fe-oxide samples.
minerals including hematite and magnetite. Torrent and Barrón (2002) According to the model selection presented by Fig. 9, the trained

Fig. 10. Comparison of the concentrations estimated using reflectance

Quartz R 2 : 85% Magnetite R 2 : 87% spectra versus the laboratory assays. Mineral grades were estimated using
80 60
PLS and two latent variables for quartz and hematite and three latent
Estimated (%)
Estimated (%)
variables for magnetite and ferrorichterite.
60
40
40
20
20
0 0
0 20 40 60 80 0 20 40 60
Hematite R 2 : 73% Ferrorichterite R 2 : 79%

8 10
Estimated (%)
Estimated (%)
4 5
0 0
0 2 4 6 8 0 5 10
model for hematite is subject to some uncertainty. The fall in the va- reflectance spectra. Quartz still has an explicit pattern in the validation
lidation error around the second latent variable was followed by a very error. The improvement of the model selection for hematite and mag-
short rise while a model is selected reliably when the fall in the vali- netite is clear. However, adding reflectance spectra obscures the model
dation error is deep and is followed by an increasing slope. selection of ferrorichterite because the proper model seems to have five
latent variables while with LIBS spectra the model could be selected
with three variables.
3.3. Data fusion of LIBS and reflectance spectroscopy Fig. 12 represents the estimations when reflectance spectra were
combined with LIBS spectra. It is clear that deploying reflectance
The data combination is simply performed by appending pre-pro- spectra has enhanced the estimations of almost all the minerals. Most
cessed reflectance spectra to the pre-processed LIBS spectra. There will importantly, hematite has been estimated with R-squared value of 61%.
be no need for further pre-processing of the data after this combination. In practice, for process control purposes R-squared values in the range
Although such data fusion increases the dimension of the data sig- of 60% and above start to be useful.
nificantly (almost double compared with LIBS spectra alone), during A related point to consider is that the reliability of the mineral
model validation using PLSR technique the number of latent dimen- content predictions is mainly justified by the process of model selection
sions are selected so that only the most relevant information are taken (Figs. 7, 9 and 11) while the R-squared criterion is mainly used as the
into account. demonstration of the model performance. Furthermore, the results of
Fig. 11 represents the PLS model selection of the combined LIBS and
Fig. 11. Model selection for combined LIBS and reflectance spectroscopy.
Quartz Magnetite
30 20 Selecting the number of latent variables using training and validation er-
Validation rors for ferrorichterite, quartz, magnetite and hematite.
15 Training
20
MSE
MSE
10
10
5
0 0
0 5 10 15 0 5 10 15
3 6
2 4
MSE
MSE
1 2
0 0
0 5 10 15 0 5 10 15

Fig. 12. Comparison of the concentrations estimated using combined LIBS

Quartz R2: 88% Magnetite R2: 92% and reflectance spectra versus the laboratory assays. Mineral grades were
80 60
estimated using PLS and two latent variables for quartz, three latent vari-
Estimated (%)
Estimated (%)
ables for magnetite and hematite and four latent variables for ferror-
60 ichterite.
40
40
20
20
0 0
0 20 40 60 80 0 20 40 60
Hematite R2: 61% Ferrorichterite R2: 83%

10 10
Estimated (%)
Estimated (%)
5 5
0 0
0 5 10 0 5 10
the previous work (Khajehzadeh et al., 2016) can easily lead us to be- Fig. 13 represents the PLS model selection when the data of two XRF
lieve that having more measurements for each sample will increase the iron channels were appended to the measured reflectance spectra. It is
R-squared values. clear that data fusion of XRF and reflectance spectroscopy techniques
has made the procedure of model selection much smoother for quartz,
3.4. Data fusion of reflectance spectroscopy with two iron-channels of XRF hematite and magnetite. However, the combined reflectance spectro-
measurement technique scopy and XRF seems to fail when estimating concentration of ferror-
ichterite and this can be observed from the estimations in Fig. 14.
XRF measurement techniques are widely applied for direct mea- Fig. 14 represents the quantitative estimations of the slurry mineral
surement of slurry elemental contents. However, Section 2.4.2 ex- contents using the combined reflectance spectroscopy and XRF tech-
plained that emission of elements below Ca are absorbed by water of niques versus the laboratory assays. Comparing the mineral estimations
slurry, protective windows and air. Therefore, in this work XRF spectra of different measurement techniques easily induces that the combined
of the measured slurry samples contained few peaks mainly related to XRF and reflectance spectroscopy technique is outperforming other
iron and thus on-stream mineralogical identification of slurries using techniques for hematite, magnetite and quartz. Section 3.5 presents a
XRF was omitted. summary of the results including a comparison and justification of the
Fig. 13. Model selection of the combined reflectance spectroscopy and XRF
Quartz Magnetite
20 20 measurement techniques.
Validation
15 15 Training
MSE
MSE
10 10
5 5
0 0
0 5 10 15 0 5 10 15
2 6
1.5
4
MSE
MSE
1
2
0.5
0 0
0 5 10 15 0 5 10 15

Fig. 14. Comparison of the concentrations estimated using the combined

Quartz: 95% Magnetite: 94%
80 60 reflectance spectra and XRF versus the laboratory assays. Mineral grades
were estimated using PLS and two latent variables for hematite and fer-
rorichterite and three latent variables for magnetite and quartz.
60
Measured
Measured
40
40
20
20
0 0
0 20 40 60 80 0 20 40 60
Estimated Estimated
Hematite: 85% Ferrorichterite: 59%
8 10
6
Measured
Measured
4 5
0 0
0 2 4 6 8 0 5 10
Estimated Estimated
performances. 3.5.1. Performance of reflectance spectroscopy

Section 3.2 reported the capability of reflectance spectroscopy for
mineral identification. Mineral contents could be identified by training
3.5. Discussion and summary of results distinct models. However, the trained model for hematite was subject to
some uncertainty because the fall in the validation error around the
In this work, instead of the physical fusion of the measurement second latent variable was followed by a very short rise (Fig. 9). This
techniques, the combination was performed on the collected spectra uncertainty was overcome when two iron channels of XRF data were
and is known as data fusion. For corporations that produce multiple combined with the reflectance measured spectra (see Section 3.4).
analyzers and for clients who have already installed those analyzers, it In order to emphasize the advantage of transforming reflectance to
is more economic and time-saving to perform data fusion of different absorbance using Beer–Lambert law and to provide an evidence for the
installed analyzers rather than the physical combination of them. validity of the approach, Table 3 represents the accuracy of mineral
Table 2 represents a summary of the results for quantitative esti- estimations with and without data conversion.
mation of quartz, magnetite, hematite and ferrorichterite as the most The reflectance spectroscopy measurement is affected by the solids
important minerals of the slurry samples used in this study. The table content and particle size of the slurry. In a laboratory setup with a small
includes the number of latent variables selected for PLS model of each number of slurry samples these factors can be assumed to be relatively
mineral and R-squared value of each minerals estimation. As XRF constant and a calibration model can be calculated. However, in real
technique is more accurate than LIBS technique when measuring iron, on-stream analysis applications these factors are typically unknown.
the combined XRF and reflectance spectroscopy measurement tech- Additionally, both the solids content and particle size vary depending
nique is giving better results for magnetite and hematite than other on the feed ore properties and operating point of the process meaning
techniques. Additionally, quartz is negatively correlated with iron, and that a constant calibration model does not necessarily hold for longer
that is why the combined XRF and reflectance spectroscopy works well periods of time. Therefore, application of reflectance spectroscopy for
for quartz too. Since ferrorichterite contains light elements and it was online slurry analysis needs frequent calibration while it is an expensive
discussed in Section 1 that LIBS has better performance than XRF when and time consuming procedure. One approach to solve this issue is to
measuring elements lighter than titanium, therefore combined LIBS and calibrate the reflectance model using the results of other available on-
reflectance spectra enabled accurate measuring of ferrorichterite. line analyzers. However, as hematite could not be measured accurately
The following sections will discuss the performance of different with LIBS spectra and more elemental lines of XRF is required for online
techniques and data fusions in detail. calibration of reflectance spectroscopy, this issue will be studied in the
Table 2
Summary of the results including the R-squared values of minerals content estimations. The best estimation for each mineral has been highlighted in bold.

Table 3
R2 values of estimated minerals with & without conversion of reflectance to absorbance.
extended work of this research using more samples. coefficients have been scaled between zero and one to enable re-
presentation of them in one figure. Variables 1–111 are related to the
3.5.2. Performance of combined LIBS and reflectance spectroscopy selected LIBS lines after pre-processing and the rest are related to re-
Section 3.3 presented the improved estimation of quartz, magnetite flectance spectroscopy. Given the procedure of model selection ex-
and ferrorichterite after combing LIBS and reflectance spectra. How- plained by Fig. 11 at Section 3.3, the regression coefficients of quartz,
ever, adding reflectance spectra forced the model for ferrorichterite to magnetite, hematite and ferrorichterite have been calculated using two,
be trained using five latent variables while with LIBS spectra the model three, three and five latent variables respectively. It can be seen from
could be selected with three variables. This can be due to the increasing Fig. 15 that hematite has obtained large coefficients for reflectance
the number of input variables whereas the number of training samples variables while for ferrorichterite they are close to zero. Consequently,
is constant. Ferrorichterite contains hydroxyl groups having char- fusion of reflectance and LIBS spectra provides more information for
acteristic absorption features in the SWIR (around 2300 and 2390 nm). estimation of hematite content and enhances the estimation of quartz
However, those wavelengths were not in the detection range of the and magnetite contents.
spectrograph used in this work. Furthermore, in slurry applications
water absorption hinders the measurement of reflectance in these wa- 3.5.3. Performance of combined XRF and reflectance spectroscopy
velengths. Otherwise it is expected that detecting hydroxyls using re- Section 3.4 represented that the two iron-lines of XRF spectra
flectance spectroscopy would further improve the estimation of fer- greatly assisted reflectance spectroscopy technique for quantitative
rorichterite. mineral identification of the slurries. This result can be explained so
Given Figs. 10 and 12, there might be the question how hematite that XRF is able to accurately measure the total iron content, whereas
was estimated more accurately using reflectance spectra alone com- reflectance spectroscopy is sensitive to the hematite content of the
pared to the combined LIBS and reflectance spectra. Although the ad- slurry due to the characteristic spectrum of hematite (Brown et al.,
dition of LIBS variables to the reflectance spectra is interpreted as extra 2006). Assuming that the majority of iron is in magnetite and hematite
information, PLS algorithm models also the correlation structure of the (see Table 4), the linear combination of XRF and reflectance spectra can
input variables. Therefore, adding extra variables might cause the PLS then be used by the PLS model to estimate both magnetite and hematite
algorithm to find irrelevant correlations between the input variables. well. Improved quartz modeling is most likely relying on the strong
Such ambiguity can be resolved using more samples for training and (negative) correlation between the total iron and quartz contents
validation of the model. Another point to consider is that R2 value is not (Table 1). Detection of Ferrorichterite was not improved by XRF spectra
the only metric to evaluate the performance of a trained model. Al- because of small concentration and small share of total iron, however
though the accuracy of the hematite estimation has been decreased LIBS could assist slightly.
after combination of LIBS and reflectance spectra, the reliability of the
trained model has been increased since the fall in the validation error 4. Conclusion
(see Fig. 11) is followed by an ascending steep slope and the model is
selected more decisively. This study focused on the quantitative mineral identification of the
In order to clarify how PLS algorithm assigns coefficients to in- tailing slurry samples. The samples came from an iron ore concentrator
dependent variables, Fig. 15 illustrates the regression coefficients of the where the on-stream analysis of quartz, hematite and magnetite would
combined LIBS and reflectance spectra for each mineral. The provide means to improve the process performance. The approach was
1.5 Fig. 15. PLS regression coefficients of quartz, magnetite,

hematite and ferrorichterite using two, three, three and
Quartz
Magnetite five latent variables respectively. “a.u” indicates arbitrary
1 Hematite unit.
Ferrorichterite
0.5
a.u
-0.5
-1
LIBS coefficients Reflectance coefficients
-1.5
0 50 100 150 200 250
Variable number

Table 4
Average distribution of iron in the iron-bearing minerals of the slurry samples.
to obtain mineralogical analysis by combining the measurement data of Cremers, D.A., Radziemski, L.J., 2006. Handbook of Laser-Induced Breakdown
Spectroscopy. Cambridge University, p. 25.
available slurry analyzers (data fusion) rather than introducing com- D’Amico, A., Di Natale, C., Castro, F.L., Iarossi, S., Catini, A., Martinelli, E., 2009. Volatile
pletely new measurement techniques. Three different techniques (LIBS, compounds detection by ir acousto-optic detectors. In: Unexploded Ordnance
XRF and reflectance spectroscopy) were used for the measurements. It Detection and Mitigation. Springer, pp. 21–59.
Das, R.S., Agrawal, Y., 2011. Raman spectroscopy: recent advancements, techniques and
was shown that the on-stream mineralogical assaying of the slurry flows applications. Vibrat. Spectrosc. 57 (2), 163–176.
could be improved significantly by data fusion of different slurry ana- Eilers, P.H., Boelens, H.F., 2005. Baseline Correction with Asymmetric Least Squares
lyzers. The data fusion was performed using partial least squares re- Smoothing. Leiden University Medical Centre Report.
Filippov, L., Severov, V., Filippova, I., 2014. An overview of the beneficiation of iron ores
gression technique. Data combination of reflectance spectroscopy and
via reverse cationic flotation. Int. J. Min. Process. 127, 62–69.
laser-induced breakdown spectroscopy enhanced the quantitative Geladi, P., Kowalski, B.R., 1986. Partial least-squares regression: a tutorial. Anal. Chim.
identification of hematite, magnetite, quartz and ferrorichterite. Also, Acta 185, 1–17.
Gemperline, P., 2006. Practical Guide to Chemometrics. CRC Press.
the combination of accurate XRF iron data with the reflectance spectral
Goncharov, A.F., Struzhkin, V.V., 2003. Raman spectroscopy of metals, high-temperature
measurements provided valid estimates of the slurry mineral contents. superconductors and related materials under high pressure. J. Raman Spectrosc. 34
This study demonstrated the capability of data fusion from different (7–8), 532–548.
slurry analyzers for online quantification of the important minerals in a Haavisto, O., Hyötyniemi, H., 2011. Reflectance spectroscopy in the analysis of mineral
flotation slurries. J. Process Control 21 (2), 246–253.
flotation plant. In general, the correct selection of the analyzers in- Haavisto, O., Kaartinen, J., Hyotyniemi, H., 2006. Optical spectrum based estimation of
cluded in the estimation depends on the available analyzers at the grades in mineral flotation. In: IEEE International Conference on Industrial
plant, type of ore and mineralogy. Since the proposed data fusion is Technology, 2006, ICIT 2006. IEEE, pp. 2529–2534.
Haest, M., Cudahy, T., Laukamp, C., Gregory, S., 2012a. Quantitative mineralogy from
based on both mineral and element raw data, the correlation of mi- infrared spectroscopic data. i. Validation of mineral abundance and composition
nerals and elements in the ore affect the performance of the analysis. scripts at the rocklea channel iron deposit in western australia. Econ. Geol. 107 (2),
The samples used in this study were from an iron ore concentrator 209–228.
Haest, M., Cudahy, T., Laukamp, C., Gregory, S., 2012b. Quantitative mineralogy from
where quartz was the main gangue mineral. The PLS-based approach infrared spectroscopic data. ii. Three-dimensional mineralogical characterization of
performed well when estimating abundant minerals like quartz and the rocklea channel iron deposit, western australia. Econ. Geol. 107 (2), 229–249.
magnetite. The estimations of hematite was improved specially after Hyötyniemi, H., 2001. Multivariate Regression: Techniques and Tools. Helsinki
University of Technology.
data fusion of reflectance spectroscopy and XRF. Given the results of
Khajehzadeh, N., Haavisto, O., Koresaar, L., 2016. On-stream and quantitative mineral
this research and the previous work (Khajehzadeh et al., 2016), the identification of tailing slurries using LIBS technique. Min. Eng. 98, 101–109.
proposed PLS approach functions better when estimating abundant Khater, M.A., 2013. Laser-induced breakdown spectroscopy for light elements detection
in steel: state of the art. Spectrochim. Acta Part B: Atom. Spectrosc. 81, 1–10.
minerals and the reliability descends for low-concentration minerals.
Lamberg, P., Vianna, S., 2007. A technique for tracking multiphase mineral particles in
The future work of this research will concentrate on different types of flotation circuits. In: Proceedings of the 7th Meeting of the Southern Hemisphere on
slurry samples to generalize the proposed data fusion approach. How- Mineral Technology, Ouro Preto, Brazil, pp. 20–24.
ever, theoretically it is derivable that the applied PLS technique in this Osborne, B.G., Fearn, T., Hindle, P.H., et al., 1993. Practical NIR Spectroscopy with
Applications in Food and Beverage Analysis. Longman Scientific and Technical.
research performs reliably when estimating abundant minerals of dif- Outotec, 2015. Outotec Courier SL On-line Analyzers, Brochure. < http://www.outotec.
ferent ore types. com/ImageVaultFiles/id_1607/d_1/cf_2/OTE_outotec_courier_sl_online_analyzers_
brochure-l.PDF > .
Rapiscan Systems, 2005. SATuration MAGnetization ANalyser (SATMAGAN S135).
References < http://www.indehk.com/PDF/SatmagandescriptionandIIinfoOct2005.pdf > .
Russo, R., Mao, X., Borisov, O., Liu, H., 2000. Influence of wavelength on fractionation in
Barnett, C., Cahoon, E., Almirall, J.R., 2008. Wavelength dependence on the elemental laser ablation icp-ms. J. Anal. Atom. Spectrom. 15 (9), 1115–1120.
analysis of glass by laser induced breakdown spectroscopy. Spectrochim. Acta Part B: Saarhelo, K., Paakkinen, U., Pennanen, P., 1988. On-line analysis in industrial mineral
Atom. Spectrosc. 63 (10), 1016–1023. applications. In: 8th Industrial Minerals Inst. Congress of Boston.
Beckhoff, B., Kanngießer, B., Langhoff, N., Wedell, R., Wolff, H., 2007. Handbook of Scheinost, A., Chavernas, A., Barrón, V., Torrent, J., 1998. Use and limitations of second-
Practical X-ray Fluorescence Analysis. Springer Science & Business Media. derivative diffuse reflectance spectroscopy in the visible to near-infrared range to
Brown, D.J., Shepherd, K.D., Walsh, M.G., Mays, M.D., Reinsch, T.G., 2006. Global soil identify and quantity fe oxide minerals in soils. Clays Clay Min. 46 (5), 528–536.
characterization with VNIR diffuse reflectance spectroscopy. Geoderma 132 (3), Torrent, J., Barrón, V., 2002. Diffuse reflectance spectroscopy of iron oxides.
273–290. Encyclopedia Surface Colloid Sci. 1, 1438–1446.
Carr-Brion, K., 1989. X-Ray Analysers in Process Control. Kluwer Academic Pub. Warren, B.E., 1969. X-ray Diffraction. Courier Corporation.
Chang, C.-I., 2003. Hyperspectral Imaging: Techniques for Spectral Detection and Wold, H., 1985. Partial least squares. In: Encyclopedia of Statistical Sciences.
Classification, vol. 1 Springer Science & Business Media. Yu, X., Li, Y., Gu, X., Bao, J., Yang, H., Sun, L., 2014. Laser-induced breakdown spec-
Clark, R.N., et al., 1999. Spectroscopy of rocks and minerals, and principles of spectro- troscopy application in environmental monitoring of water quality: a review.
scopy. Manual Rem. Sens. 3, 3–58. Environ. Monit. Assess. 186 (12), 8969–8980.
Cooper, H.R., 1976. On-Stream X-ray Analysis. Flotation–A.M. Gaudin Memorial.

Aalto-DD 177/2018
ISBN 978-952-60-8182-3 (printed) BUSINESS +

9HSTFMG*aibicd+
ISBN 978-952-60-8183-0 (pdf) ECON OMY

ISSN 1799-4934 (printed)
ISSN 1799-4942 (pdf) AR T +
DESIGN +
Aalto Universit y ARC HITECTURE
School of Electrical Engineering
Depar tment of Electrical Engineering and Automation SCIEN CE +
www.aalto.fi TECHNOLOGY
CROSSOVER
D O C TO R A L
DISSERTATIONS
Ve
i w pubcilato
i n stats

Analytical Techniques For Online Mineral Identification

Uploaded by

Copyright:

Available Formats

You might also like

Analytical Techniques For Online Mineral Identification

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Analytical Techniques For Online Mineral Identification

Uploaded by

Copyright:

Available Formats

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

ANALYTICAL TECHNIQUES FOR ONLINE MINERAL IDENTIFICATION

Thesis · February 2019

Doctoral Thesis View project

The user has requested enhancement of the downloaded file.

ANALYTICAL TECHNIQUES FOR

A doctoral dissertation completed for the degree of Doctor of

Aalto University publication series

© 2018 Navid Khajehzadeh

ISBN 978-952-60-8182-3 (printed)

Al t ho u g h t he s p e ct ro s c o p ic t e c hniqu e s p e rf orme d in t he l ab o rat o rie s p ro vid e ac c u rat e re s u l t s , t he

Keywords Mineral identification, X-ray fluorescence, Laser-induced fluorescence, Laser-induced

My heartfelt thanks go to all my labmates, colleagues and academic friends:

Espoo, August 30, 2018,

List of Publications vii

1.3.4 Hypothesis 4: Does data fusion of diﬀerent measurement

4. Results and discussion 47

4.3 Quantitative mineral identiﬁcation of tailing slurries using LIBS

This thesis consists of an overview of the following publications which are

I Kauppinen, T., Khajehzadeh, N. and Haavisto, O. Laser-induced ﬂuores-

II Khajehzadeh, N., Kauppinen, T., Häkkänen, H. Laser-induced ﬂuorescence

III Khajehzadeh, N., Kauppinen, T. Fast mineral identiﬁcation using elemen-

IV Khajehzadeh, N., Haavisto, O. and Koresaar, L. On-stream and quantita-

V Khajehzadeh, N., Haavisto, O. and Koresaar, L. On-stream mineral iden-

Publication I: “Laser-induced ﬂuorescence images and Raman

This article is an equal contribution between Tommi Kauppinen and Navid

Publication II: “Laser-induced ﬂuorescence imaging: detection of

Publication III: “Fast mineral identiﬁcation using elemental LIBS

Publication IV: “On-stream and quantitative mineral identiﬁcation

Publication V: “On-stream mineral identiﬁcation of tailing slurries

CIE international commission on illumination

EDXRF energy-dispersive X-ray Fluorescence

HIS hue, intensity and saturation

HLS hue, lightness and saturation

HSV hue, saturation and value

HVS human visual system

iid independent and identically distributed

keV kilo electron volt

L*a*b* average values of Lab coordinates

Lab Lightness, redness-greeness, yellowness-blueness

LHD load haul dumper

LIBS laser-induced breakdown spectroscopy

LIF laser-induced ﬂuorescence

LIPS laser-induced plasma spectroscopy

LWIR long-wave infrared

MLA mineral liberation analyzer

MSE mean square error

NIPALS non-linear iterative partial least squares

NNLS non-negative least squares

PCA principal component analysis

PCR principal component regression

PLS partial least squares

RGB red, green, blue

SVD singular value decomposition

SWIR short-wave infrared

VNIR visible and near-infrared

WDXRF wavelength-dispersive X-ray Fluorescence

XRD X-ray diﬀraction

Lab* average values of Lab coordinates