Minerals Engineering 69 (2014) 1–12

Contents lists available at ScienceDirect

Minerals Engineering
journal homepage: www.elsevier.com/locate/mineng

Estimation of the bubble size and bubble loading in a flotation froth

using electrical resistance tomography
Antti Nissinen a,⇑, Anssi Lehikoinen b, Mika Mononen c, Seppo Lähteenmäki d, Marko Vauhkonen a
a
Department of Applied physics, University of Eastern Finland, P.O. Box 1627, FIN-70211 Kuopio, Finland
b
Rocsole Ltd, Microkatu 1, FIN-70211 Kuopio, Finland
c
Outotec Ltd, Viestikatu 7, FIN-70601 Kuopio, Finland
d
Inmet Pyhäsalmi Mine Ltd, FIN-86801 Pyhäsalmi, Finland

a r t i c l e i n f o a b s t r a c t

Article history: Flotation process is widely used in mineral industry for the separation of valuable minerals from low-
Received 16 April 2014 grade ore slurry. There are several parameters such as the bubble size and bubble loading that predict
Accepted 2 July 2014 the efficiency of the flotation process. These parameters can be used for the control of the flotation pro-
Available online 21 July 2014
cess. There are already some techniques that can be used for online monitoring of these parameters, for
example, the high-speed video imaging and a probe sensor based on electrical resistance tomography
Keywords: (ERT). These methods, however, suffer for some limitations. The high speed video imaging gives informa-
Electrical resistance tomography
tion only on the surface of the froth and in the previously proposed ERT based techniques the conductiv-
Flotation
Bubble size measurement
ity of the froth is typically modeled to be smoothly varying. However, in reality the froth is composed of
Bubble loading different size of bubbles having highly conductive surface and very low conductive interior which config-
uration cannot be modeled with smoothly varying conductivity distribution. In this paper, we propose a
computational approach in which the structure of the froth is modeled and both the bubble size and the
conductivity of the boundary of the bubbles are estimated. The proposed approach utilizes data measured
with the standard ERT probe. The estimated bubble size and conductivity of the boundary of the bubbles
are compared to online measured camera based estimates of the bubble size and bubble loading. The pro-
posed approach is evaluated with simulated measurements and real data from Pyhäsalmi Mine. The
results show that there is a high correlation between the camera based and the ERT based estimates of
the bubble size. Furthermore, some of the parameters obtained from the ERT based method correlate well
with the camera based estimate of the bubble loading.
Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction stability was evaluated using electrical impedance spectroscopy

In mineral industry the froth flotation process is widely used for
the separation of minerals from low-grade ore slurry. The grade
and recovery in the flotation process depends on the structure
and stability of the froth, see Neethling and Cilliers (2003),
Farrokhpay (2011). The bubble size distribution, bubble loading,
height of froth, liquid content of the froth, the velocity of the froth
and bursting rate of the bubbles are important parameters that
affect the structure and stability of the froth. There are several
methods to estimate these parameters. The bubble size, the burst-
ing rate of the bubbles and bubble loading were estimated using
camera based methods in Grau and Heiskanen (2002), Kaartinen
et al. (2006). The froth stability column was applied to evaluate
the stability of the froth in Barbian et al. (2005, 2006). The froth
⇑ Corresponding author.
E-mail address: antti.nissinen@uef.fi (A. Nissinen).
in Hu et al. (2009). Simple conductivity probes, floats and pressure
transducers have been used for the estimation of the height of
froth, see Maldonado et al. (2008).
One possible technique for the froth flotation process monitor-
ing is electrical resistance tomography (ERT). In electrical resis-
tance tomography, electrodes are attached on the boundary of an
object or on the measurement probe and currents are injected into
the object through these electrodes. The voltages on all electrodes
are measured and the conductivity of the object is reconstructed
based on the measured voltages and known currents; for reviews
on ERT, see Cheney et al. (1999), Kaipio et al. (2000), Borcea
(2002). Note that in the literature the term electrical impedance
tomography (EIT) is often used when more accuare term would
be ERT.
Electrical resistance tomography has already been applied for
imaging of flotation process in Cilliers et al. (1999), Kourunen
et al. (2008), Normi et al. (2009), Kourunen et al. (2011),

http://dx.doi.org/10.1016/j.mineng.2014.07.001
0892-6875/Ó 2014 Elsevier Ltd. All rights reserved.
2 A. Nissinen et al. / Minerals Engineering 69 (2014) 1–12
Lehikoinen et al. (2011), Reunanen et al. (2011). In these papers, the
conductivity of the froth is typically modeled to be smoothly vary-
ing. However, in reality the froth is composed of different size of
bubbles having highly conductive surface and very low conductive
interior which configuration cannot be modeled with smoothly
varying conductivity distribution.
The aim of this study is to estimate the bubble size and also to
obtain information on the bubble loading using ERT. For this pur-
pose, an approximative froth model is constructed and the conduc-
tivity of the boundary of the bubbles and conductivity of the slurry
is estimated. The vapor inside the bubble is very low or non-con-
ductive and the surface of the bubble can be highly conductive
due to the water and mineral particles on the surface of the bub-
bles. Therefore it is believed that the conductivity of the boundary
of the bubbles can give information on the bubble loading. In this
approximative froth model, an approximative and fixed bubble
size is used. Since this is not exactly correct, the so called approx-
imation error approach is utilized. With this approach it is possible
to compensate for errors caused by the unknown parameters, such
as the bubble size in this study. For more details on approximation
error approach applied for unknown parameters, see Kolehmainen
et al. (2011), Nissinen et al. (2011b). The benefit of the approach is
that it enables the estimation of the bubble size without changing
the froth model during the computations.
The approximation error approach was originally proposed in
Kaipio and Somersalo (2005, 2007). Since then the approach have
been applied for different approximation errors and applications.
The approximation error approach was applied to ERT in Kaipio
and Somersalo (2005), Lehikoinen et al. (2007), Nissinen et al.
(2008, 2011a), for example. In optical tomography the approach
was applied in Heino and Somersalo (2004), Arridge et al. (2006),
Kolehmainen et al. (2011), Tarvainen et al. (2010), for example. In
electrical capacitance tomography the approximation error
approach was applied in Banasiak et al. (2012). Recently, the approx-
imation error approach was applied to seismic imaging of the aquifer
dimensions in Lähivaara et al. (2014). The approach was recently
applied to atmospheric modeling in Lipponen et al. (2013a).
In this study, the proposed method is evaluated with simulated
measurements and also with real measurements conducted at
Inmet Pyhäsalmi Mine at Pyhäsalmi Finland. The real measure-
ments were obtained using the Outotec LevelSense™ sensor. It
has been found earlier that this sensor can be used to control and
optimize the flotation process, see Lehikoinen et al. (2011),
Vauhkonen et al. (2013). The ERT based estimates of the bubble size
and conductivity of the boundary of the bubbles are compared to
camera based estimates of the bubble size and bubble loading. Fur-
thermore, the average conductivity of the froth is obtained from
Outotec LevelSense™ sensor and it is used as a reference parameter.
The rest of this paper is organized as follows. In Section 2, a brief
preview of the Outotec LevelSense™ sensor and measurement sys-
tem is given. Furthermore, the approximative model for the conduc-
tivity is explained and the measurement model is represented. In
Section 3, the Bayesian formulation of the ERT problem is reviewed
and the estimation of the conductivity and bubble size is explained.
The details of the computations (parameters and used methods) and
the reference estimates are explained in Section 4. The proposed
approach is evaluated using simulated and real data in Section 5
and the conclusions are given in Section 6.
2. Materials and methods
2.1. Measurement probe
In ERT, electrodes are typically attached on the surface of the
object and the measurements are conducted using these
electrodes. This measurement arrangement is used when relative
small objects such as flow pipes and small measurement tanks
are investigated. The size of the flotation cell is typically so large
that the measurement devise that measures on the boundary of
the cell would be impractical and expensive to construct. There-
fore, a probe sensor where the electrodes are on the surface of
the probe is used for the monitoring on flotation process.
In this study, the real measurements were carried out using
Outotec LevelSense™ sensor. In this sensor N el ¼ 22 ring shaped
electrodes are attached on the boundary of the measurement
probe. The height of the electrodes is 1 cm and the space between
the electrodes is 2 cm. The radius of the measurement probe is
3 cm. The arrangement of the sensor in the flotation cell and a pic-
ture of the sensor in the flotation cell at Pyhäsalmi mine are shown
in Fig. 1. Using the electrodes, electric currents are injected into the
domain and corresponding voltages V are measured using the same
electrodes. The currents are injected between all adjacent pairs of
electrodes and the voltages are measured between all adjacent
pairs for all current injections. The measured voltages are collected
to one measurement vector V 2 R441 . The amplitude of the injected
current is 13 mA. The conductivity distribution r around the probe
can be estimated based on the known currents and measured
voltages.
In Outotec LevelSense™ sensor, several parameters are com-
puted using the measurements. These parameters include the con-
ductivity of the froth (bulk conductivity), the height of the froth,
the place of the froth/slurry interface and conductivity of the
slurry, for example. These parameters are used in the control of
the flotation process. For more detailed information on the Outotec
LevelSense™ sensor and controlling of the flotation process, see
Lehikoinen et al. (2011), Vauhkonen et al. (2013).
2.2. Modeling of the conductivity
The conventional approach in ERT is to model the conductivity
to be smoothly varying. However, in reality the froth is composed
of different size of bubbles having highly conductive surface and
very low conductive interior which configuration cannot be mod-
eled with smoothly varying conductivity distribution. In this work,
the structure of the froth is modeled in the finite element model in
order to estimate both bubble size and conductivity of the bound-
ary of the bubbles.
A schematic image of the computation domain is shown in
Fig. 2. The ring shaped electrodes are shown on the boundary of
the probe. When ring shaped electrodes are used, the measure-
ments are averaged in the angular direction. The angular position
of the inhomogeneity cannot be reconstructed with these mea-
surements. A reasonable approximation with this type of applica-
tion and measurement configuration is that the conductivity is
invariant in the angular direction. Recently, a 2D cylindrically sym-
metric forward model for probe geometry in ERT was proposed in
Nissinen et al. (2013). This model is accurate model as far as the
conductivity is rotationally invariant. In Nissinen et al. (2013) it
was shown that with asymmetric targets the estimated conductiv-
ity using cylindrically symmetric model is close to the angular
average of the asymmetric 3D conductivity. The cylindrically sym-
metric forward model is used also in this work. The solution of the
forward problem is computed using finite element method (FEM).
One major benefit of the cylindrically symmetric forward model is
that the construction of the 2D model for the froth is much easier
than full 3D model. Furthermore, the cylindrically symmetric for-
ward model is computationally faster than full 3D model.
The computation domain X  R2 is divided to three separate
regions. These regions are the boundaries Xb of the bubbles in
the froth, air Xa inside the bubbles and slurry Xs . The bubbles
are modeled to be squares and the length of the edge of the square
 A. Nissinen et al. / Minerals Engineering 69 (2014) 1–12
Fig. 1. (a) The arrangement of the LevelSense™ sensor in the flotation cell. (b) The LevelSense™ sensor inserted in a flotation cell in Inmet Pyhäsalmi Mine at Pyhäsalmi.
Fig. 2. A schematic picture of the measurement probe and the 2D computation
domain X used in the cylindrically symmetric model. The ring shaped electrodes
(black rings) are shown on the boundary of the probe. The radius of the
measurement probe is 3 cm, radius of the 2D computation domain is 25 cm and
height of the domain is 94 cm.
is the bubble size c. The model of the froth and closer view of the
finite element mesh is shown in Fig. 3. In this work, the conductiv-
ity is modeled to be constant in regions Xb ; Xa and Xs . The conduc-
tivity inside these regions is denoted as rb ; ra and rs , respectively.
The conductivity of air inside bubbles is modeled to be constant
ra ¼ 107 mS cm1. The small value for ra is used because the
finite element solution of the forward problem would not be pos-
sible with zero conductivity regions inside the domain. Note that
this model for the froth is very approximative. The actual shape
of the bubbles is complex and the size of the bubbles varies in
the froth. Furthermore, the conductivity of the actual froth is not
3
Fig. 3. The finite element model of the conductivity. (a) Computation domain X, the
boundaries Xb of the bubbles in the froth are shown with black color and slurry is
denoted as Xs . The electrodes are shown with red patches on the boundary of the
computation domain. (b) Closer view of the mesh. The elements that model the
boundaries Xb of the bubbles are shown with gray color and air inside the bubbles
is denoted with Xa .
cylindrically symmetric. Due to these approximations the esti-
mated values of the conductivity and the bubble size might contain
errors. The main objective in this work is to obtain information on
the temporal variation, not the absolutely correct static value, of
the conductivity and the bubble size. One advantage of the simpli-
fied cubical froth model is that the different bubble sizes can be
simulated using one FE mesh. Furthermore, the thickness of the
boundary of the bubbles is equal in all simulated conductivity dis-
tributions. It is clear that all the simplified assumptions are not
true in real measurement cases, causing some uncertainty in the
estimation. However, it is assumed that the simplified and compu-
tationally fast bubble model works well enough (for example for
control purposes) also in practical cases.
4 A. Nissinen et al. / Minerals Engineering 69 (2014) 1–12
2.3. Measurement model
In order to solve the inverse problem in ERT we need to have an
observation model that connects the injected currents, conductiv-
ity and voltages on the electrodes. The observation model is of the
form
V ¼ Uðr; cÞ þ e; e  N ðe ; Ce Þ ð1Þ
m
where V 2 R ; m ¼ 441, is the vector of the measured voltages,
r 2 RN ; c is the bubble size and e 2 Rm is a Gaussian distributed
measurement noise with mean e 2 Rm and covariance matrix Ce .
The dependence of the forward solution on the bubble size is
expressed by the parameterization c. The forward problem Uðr; cÞ
is solved using the cylindrically symmetric forward model and
finite element method. Furthermore, reduced-order finite element
model is used in order to reduce the computation time, see
Lipponen et al. (2013b). The conductivity r 2 RN is represented
using piecewise constant basis functions in the finite element
model. The number of elements in the FE mesh is N. The individual
elements in the FE mesh M are denoted by Xj ; j ¼ 1; . . . ; N. The mass
center of the element Xj is denoted by xj . The conductivity of each
element in different regions of the computation domain is repre-
sented as
8 rately known bubble size.
< rb ; if x 2 Xb
j
>
rj ¼ ra ; if x 2 Xa
j ð2Þ
>
: 3.2. Approximation error approach
rs ; if xj 2 Xs
3. Estimation of conductivity and bubble size
The conductivity rb of the boundary of the bubbles and the con-
ductivity rs of the slurry are estimated using the Bayesian approach
for ERT inverse problem. The estimation procedure is done in two
stages. Firstly, the conductivity is estimated using approximative
bubble size c ~ and the approximation error approach is applied to
compensate for the errors due to incorrect bubble size. In the second
stage, the bubble size c is estimated. The modified approximation
error approach that enables the estimation of the unknown param-
eters in the forward model was proposed in Nissinen et al. (2011b).
For more information on approximation error approach applied for
unknown parameters see Kolehmainen et al. (2011) and for Bayes-
ian approach for inverse problems in general, see Kaipio et al.
(2000), Kaipio and Somersalo (2005), Calvetti and Somersalo
(2007), Tarantola (2004), Kolehmainen et al. (2011).
3.1. Bayesian approach for ERT inverse problem
3.1.1. Posterior model
In the Bayesian framework, all unknowns and measurements
are considered as random variables and the uncertainty related
to their values is encoded in their probability distribution models.
The joint probability density of conductivity r, bubble size c and
the measurements V can be written as
pðr; c; VÞ ¼ pðr; cÞpðV j r; cÞ ¼ pðVÞpðr; c j VÞ; ð3Þ
where pðV j r; cÞ is the likelihood model and the probability density
pðr; cÞ is the prior model of r and c. The posterior density, which is
given by the Bayes formula
pðV j r; cÞpðr; cÞ
pðr; c j VÞ ¼ ; ð4Þ
pðVÞ
is the complete probabilistic model of the ERT problem and repre-
sents the uncertainty in the unknowns given the measurements.
Typically some point and spread estimates are computed from the
posterior. However, in this case this posterior model is practically
useless, since the construction of the posterior model is difficult
due to unknown bubble size.
3.1.2. Conventional posterior model
In the conventional set-up for the inverse problem, some fixed
value for the unknown parameters are used. In flotation process
application, the bubble size c is unknown. Let c ~ be an approxima-
tive bubble size that is used when the conductivity is estimated (in
the inverse problem). This approximative bubble size can be the
average bubble size of the process, for example. In the sequel,
the tilde ~ refers to the models that are to be used in the inversion.
In the Bayesian formulation, all variables that are known, such as
measurements, or are treated as fixed parameters, appear as condi-
tioning variables. Thus, if we fix c ¼ c ~, instead of pðr; c j VÞ in (4),
we actually consider
pðV j r; c ¼ c~ÞpðrÞ
pðr j V; c ¼ c~Þ ¼ : ð5Þ
pðVÞ
Any point estimates computed from (5), such as the maximum a
posteriori estimate, are bound to be highly misleading due to
approximative bubble size. In this paper, the approximation error
approach is applied for the compensation of errors due to inaccu-
3.2.1. Measurement models
Let us start with an assumption that the bubble size is known
accurately. The measurement model
V ¼ Uðr
 ; cÞ þ e ð6Þ
denotes a (sufficiently) accurate model between the unknowns and
measurements. Here the bubble size is assumed to be so accurate
that the error in the forward model is smaller than the measurement
error. The conductivity r  is a parameterization of the conductivity
using an accurate bubble size. In this paper, the approximation errors
due to inaccurate modeling of the shape of the bubbles (square
shape) is not taken into account. Furthermore, the approximation
errors due to assumption of invariance of the conductivity in the
angular direction (in the cylindrically symmetric model) are neither
considered.
As explained above, in practice the bubble size is unknown and
therefore the reconstruction of conductivity r is carried out using
some approximative bubble size. In such a case, the accurate model
(6) is replaced by the approximate measurement model:
V  Uðr; c
~Þ þ e; ð7Þ
where c~ is an approximative bubble size. The model Uðr; c~Þ in (7) is
the forward model that is to be used in the inversion and we refer to
the model Uðr; c ~Þ as the target model.
3.2.2. Approximation error model
In the approximation error approach, instead of using the
approximation (7), the accurate measurement model (6) is written
in the form
V ¼ Uðr; c
~Þ þ ðUðr
 ; cÞ  Uðr; c
~ÞÞ þ e ¼ Uðr; c
~Þ þ eðr
 ; cÞ þ e
¼ Uðr; c
~ Þ þ g; ð8Þ
where we denote g ¼ e þ e and eðr  ; cÞ represents the modeling
error due to the incorrect bubble size. Notice that the only term
in (8) that depends on the bubble size c is the approximation error
e.
 A. Nissinen et al. / Minerals Engineering 69 (2014) 1–12
3.2.3. Simultaneous estimation of the conductivity and approximation
error
In the approximation error approach, the approximation error e
is treated as random variable. In the conventional approximation
error approach, the statistics of the approximation error e is esti-
mated and approximative posterior model is derived based on this
statistics and observation model (8), see details in Kaipio and
Somersalo (2005), Kolehmainen et al. (2011). In the modified
approximation error approach the approximation error is esti-
mated simultaneously with conductivity. The approximation error
depends on the bubble size and the bubble size can be estimated
based on the estimated approximation error.
The estimation of the approximation error is conducted as fol-
lows. The approximation error is modeled with Gaussian distribu-
tion e  N ðe ; Ce Þ with mean e and covariance matrix Ce . Let
k1 P k2 P . . . P km denote the eigenvalues of the covariance
matrix Ce of the approximation error and let fw1 ; . . . ; wm g be the
corresponding eigenvectors. Note that for the Gaussian approxima-
tion error
e  e 2 spfw1 ; . . . ; wm g:
Using this relation, the realization of the modeling error is decom-
posed to the mean added with two orthogonal components ðe0 ; e00 Þ,
that is
Xp Xm
e ¼ e þ ak wk þ j¼pþ1 bj wj :
|fflfflfflfflfflfflfflk¼1
ffl{zfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflffl
ffl{zfflfflfflfflfflfflfflfflffl}
e0 e00
Using (10), the measurement model (8) is written in the modified
form
X
p
V ¼ Uðr; c
~Þ þ ak wk þ e þ e00 þ e
k¼1
~Þ þ W a þ e þ e00 þ e;
¼ Uðr; c
where W ¼ ½w1 ; w2 ; . . . ; wp  (m p matrix), a ¼ ða1 ; a2 ; . . . ; ap ÞT 2 Rp ,
e0 represents a low-rank projection of e in the basis of principal eigen-
vectors of Ce , i.e. dimension p is chosen (significantly) smaller than
m, and e00 represents the non-estimated part of the approximation
error. The dependence of the measurement model (11) on the bubble
size is embedded in the approximation error e ¼ e þ W a þ e00 . Note
that in practice the Gaussian model for the approximation error is
not always exactly correct. However, the computed examples have
shown that feasible conductivity reconstructions can be obtained
by using Gaussian model for the approximation error, see Nissinen
et al. (2011b).
Our objective is now to construct an approximation for the pos-
terior density pðr; a j VÞ using the measurement model (11) and
estimate both the conductivity and the projection coefficients
a 2 Rp . To obtain a computationally efficient approximation
p~ ðr; a j VÞ, we make a technical approximation that ðr; a; e; e00 Þ
are mutually Gaussian and uncorrelated.
Following the approach in Kolehmainen et al. (2011), we obtain
the approximate likelihood
p~ ðV j r; aÞ ¼ N ðV  Uðr; c~Þ  W a  e  e ; Ce00 þ Ce Þ
and the approximate posterior density becomes
p~ ðr; a j VÞ / pþ ðrÞp~ ðV j r; aÞpðrÞpðaÞ;
where pþ ðrÞ is the positivity prior for the conductivity. By the prop-
erties of the eigenvalue decomposition, the prior distribution for the
projection coefficients is pðaÞ ¼ N ð0; Ca Þ, where
Ca ¼ diagðk1 ; k2 ; . . . ; kp Þ.
Comparing this modified approximation error approach to the
conventional approximation error approach (see e.g. Kolehmainen
et al. (2011), Nissinen et al. (2011a)), the main difference in the
5
modified approach is that the Gaussian approximation is pre-mar-
ginalized only partially over the approximation error and part of
the approximation error is treated as unknown parameters.
The posterior model (12) represents all information about the
unknown variables based on the measurements and prior informa-
tion. Point and spread estimates are typically computed from the
posterior density. In this work, the maximum a posteriori estimate
(MAP estimate) is computed. The computation of the MAP estimate
from the posterior model (12) amounts to solving the minimiza-
tion problem
n
~Þ  W a  e  e Þk2
ðr; aÞMAP ¼ arg min kLe00 þe ðV  Uðr; c
rP0;a
 o

þ Lr ðr  r Þk2 þ kLa ak2 ; ð13Þ
where the Cholesky factor LTe00 þe Le00 þe ¼ ðCe00 þ Ce Þ1 ; LTr Lr ¼ C1
r ;
Pm
LTa La ¼ C1
a and C e00 ¼
pþ1 kj wT
j wj .
3.2.4. Reconstruction of the bubble size
Once the MAP estimation problem (13) has been solved, an
ð9Þ approximate estimate for the bubble size c is computed. For this,
we use the Gaussian approximation of the joint density of
e0 ¼ W a and c and find the MAP estimate
cMAP ¼ arg max p~ ðc j ^e0 Þ; ð14Þ
ð10Þ given by
c^ ¼ Cce0 C1 ^0
e0 e þ c ; ð15Þ
where ^e0 ¼ W aMAP is the estimate of the approximation error and c
is the prior mean of the bubble size. The confidence intervals for the
estimated bubble size can be computed from the covariance of
p~ ðc j ^e0 Þ, given by
T
ð11Þ Cc^ j ^e0 ¼ Cc  Cce0 C1
e0 Cce0 : ð16Þ
Note that the estimation of the conductivity r and the projection
coefficients a in (13) is carried out using an approximative bubble
size c
~ and therefore time consuming modifications of the froth
model during reconstruction is avoided. We refer to the MAP esti-
mate (13)–(16) as MAP with the approximation error model (MAP-
AEM). The computation of the covariances are explained in detail
in the next section.
3.2.5. Estimation of modeling error statistics
In cases in which the measurement model is linear and the prior
model and measurement error model are Gaussian, the approxi-
mation error statistics can be computed analytically, see Kaipio
and Somersalo (2005). In other cases, like in this study, the statis-
tics is, however, typically estimated by Monte Carlo simulation.
For the Monte Carlo simulation, we generate a set of N s ¼ 2000
samples from the prior models pðcÞ and pðr  Þ. The samples of the
conductivity r and bubble size c are denoted as:
fr ð‘Þ ; cð‘Þ ; ‘ ¼ 1; 2; . . . ; N s g. These samples are then used for the
computation of the accurate forward solution Uðr  ð‘Þ ; cð‘Þ Þ and for
the target model solution Uðrð‘Þ ; c ~Þ for each of the N s samples. For
the computation of the target model solution, approximative bub-
ble size c ~ ¼ 10 mm was used.
The conductivity samples for the boundary of the bubble and
ð12Þ
slurry were drawn independently from Gaussian densities
rb  N ðr0 ; 1Þ and rs  N ðr0 ; 0:25Þ. The mean of these densities
was r0 ¼ 4:2 mS cm1. The conductivity of the air inside the bub-
bles was kept fixed to value ra ¼ 107 mS cm1. The samples of
the bubble sizes were drawn from the uniform density
c  Uð3; 15Þ, the unit of the bubble size is (mm). Note that the bub-
ble size is approximated to be Gaussian distributed in the estima-
tion of the bubble size, but the approximation error approach does
6 A. Nissinen et al. / Minerals Engineering 69 (2014) 1–12

not require to use same distribution in the estimation of the 4. Methods

approximation error statistics.
The conductivity samples r  ð‘Þ and rð‘Þ were all represented in 4.1. Solution of the minimization problem and computation of bubble
the same mesh Minv , see details of the mesh in Table 1. The bub- size
bles in the mesh were modeled by finding the elements that belong
to boundaries of the bubbles and setting the conductivity of these The estimates for the conductivity and approximation error was
elements to rb , see Eq. (2). Similarly the elements that correspond solved minimizing the functional (13) using Gauss–Newton itera-
to air and slurry were found and the conductivity of these elements tion. The Gauss–Newton iteration is of the form
was changed to simulated values. 1

The estimation of the approximation error statistics was hiþ1 ¼ hi þ ðJ T C1 1
eþe00 J þ Cpr Þ J T C1
eþe00 ðV  Uðri ; cÞ  W ai  e  e Þ
~
repeated for different height of froth, the simulated heights of 
the froth were 23; 24; . . . ; 35 cm. The range of the height was cho-
 C1
pr ðhi  hpr Þ ; ð17Þ
sen based on the variation of the height of froth during the mea-
where the conductivity r and projection coefficients a are stacked
surements. The Outotec LevelSense™ estimates the height of
to variable h and
froth and therefore the correct statistics can be selected based on      
the estimated height. ri r Cr 0
hi ¼ ; hpr ¼ ; Cpr ¼ :
Given the accurate and target forward solutions, the samples eð‘Þ ai 0 0 Ca
of the approximation error were obtained as
The Jacobian matrix J is of the form
eð‘Þ ¼ Uðr ð‘Þ ; cð‘Þ Þ  Uðrð‘Þ ; c~Þ  
@U @U
J¼ @ rb @ rs
W :
The means and covariances e ; c ; Cc ; Ce , Ce00 ; Ce0 and Cce0 for the
Gaussian approximations were computed as sample averages based The number of estimated projection coefficients is chosen to be
on the samples feð‘Þ ; cð‘Þ ; ‘ ¼ 1; . . . ; 2000g. For this task, the samples p ¼ 5. This number of basis functions for approximation error was
e0ð‘Þ ; e00ð‘Þ of the projection components were obtained from the sam- found to represent the approximation error with sufficient accu-
ples feð‘Þ g by racy, for more details on selection of p, see Nissinen et al. (2011b).
e0ð‘Þ ¼ WW T ðeð‘Þ  e Þ; e00ð‘Þ ¼ QQ T ðeð‘Þ  e Þ; In the simulated examples, the correct values of the contact
impedances are used, that is the contact impedances used in the
where e is the sample average and Q ¼ ½wpþ1 ; wpþ2 ; . . . ; wm  simulation of the measurements are used also in the inverse prob-
(m m  p matrix). lem. The contact impedances were z‘ ¼ 0:1 X cm2 . When the mea-
Let n denote the stacked variables sured data is used, the contact impedances are estimated, for more
  details on the estimation of contact impedances, see Heikkinen
e0
n¼ : et al. (2002).
c The measurement error e was zero mean and the covariance
The second order joint statistics (the mean n and covariance matrix matrix Ce was diagonal matrix and the standard deviation of the
Cn ) of the projected approximation error e0 and the bubble size c are noise was 0.1% of the simulated or measured voltages. The stan-
estimated as dard deviation corresponds to the accuracy of the LevelSense™
measurement system.
1X Ns
1 X Ns
T All estimates for the conductivity and the projection coefficients
n ¼ nð‘Þ ; Cn ¼ fð‘Þ fð‘Þ ;
Ns ‘¼1 Ns  1 ‘¼1 were computed using approximative bubble size c ~ ¼ 10 mm. Fur-
thermore, the square shaped model for the bubbles was used as
where described in Section 2.2. The bubble size was modeled to be uni-
! !   form in the froth and the height of the froth was assumed to be
e0ð‘Þ
ð‘Þ e0ð‘Þ e0
n ¼ ; fð‘Þ ¼  known. In the simulated cases, correct height of the froth was used
c ð‘Þ
c ð‘Þ c  in the inverse problem and when real measurement data was used
and the estimated height of froth was used. This was obtained from
  Outotec LevelSense™ sensor. Once the Gauss–Newton iteration
Ce0 C e0 c (17) converged, the estimate for the bubble size c ^ was computed
Cn ¼ :
Cce0 Cc using (15).

The Gaussian approximation for the joint density is written as

pðe0 ; cÞ  N ðn ; Cn Þ. Other means and covariances (e ; Ce , Ce00 ) are 4.2. Prior parameters and computation of Jacobian
computed similarly as sample averages.
Note that the estimation of the approximation error statistics by The estimated conductivity is of the form r ¼ ½rb ; rs , where rb
Monte Carlo simulation is a computationally extensive task. How- is the conductivity of the boundary of the bubbles and rs is the
ever, this task can be done offline before the measurements and conductivity of the slurry. Only these two parameters for the con-
needs to be done only once for a given measurement setup, and ductivity are estimated. The conductivity inside the bubbles ra is
for the expected range of uncertainties, in this paper, the range assumed to be known and fixed value ra ¼ 107 mS cm1 is used.
of bubble sizes and the specified prior model for the conductivity. The derivatives @@U @U
rb and @ rs were obtained by computing the Jaco-
bian using piecewise constant basis functions for conductivity
Table 1 and then summing the columns that correspond to the elements
Finite element meshes, where N n is the number of nodes, N the number of elements in that belong to Xb and Xs , correspondingly.
the meshes and size is the element size (mm).
The prior covariance matrix for the conductivity r ¼ ½rb ; rs 
Mesh Nn N Size Usage was chosen to be
Msim1 3,313,441 6,617,600 0.25 Simulation of measurements  
Msim2 829,521 1,654,400 0.5 Simulation of measurements 1 0
Minv 207,961 413,600 1 Inversion Cr ¼ :
0 0:2
 A. Nissinen et al. / Minerals Engineering 69 (2014) 1–12 7

The variances were found by running the simulations using differ-
ent variances and selecting best variances by visual inspection of
estimates. The starting point of the iteration h0 is the mean of the
prior density r0 ¼ rpr ¼ 4:2 mS cm1 and a0 ¼ 0.
4.3. Reference estimates used with real measurements
Froth thickness h (Outotec LevelSense™): In Outotec Level-
The results with the proposed approach are compared to
parameters that are obtained from camera based measurements
and parameters from Outotec LevelSense™ sensor. The reference
parameters are.
Bubble loading bload (camera based): This estimate is computed
using camera measurements obtained from the same flotation
cell in which the ERT measurement probe is, see details of the
The bulk conductivity rbulk of the froth (Outotec LevelSense™):
method in Grau and Heiskanen (2002), Kaartinen et al. (2006).
The researchers of the camera method have observed by visual
inspection of froth images that bubbles with high mineral load
do not have a total reflectance point for a RGB image. This
method gives the combined area of bubbles that do not have
a total reflectance points in percentages of the whole image
area.
Bubble size bsize (camera based): This estimate is computed
using the same images as for the bubble loading, see details in
Grau and Heiskanen (2002), Kaartinen et al. (2006). The bubble
size is calculated using a watershed segmentation algorithm,
where the boundary of each bubble is determined. The bubbles
are found using total reflectance points and the number of the
pixels in each bubble is calculated. The result of the method is
the mean bubble size (mm) on the surface of the froth.
Sense™ sensor the froth thickness is obtained by first comput-
ing the vertical conductivity profile based on the ERT
measurements and then finding the jump in the conductivity
between slurry and froth. This estimate is used in our compu-
tations as a known variable and it is not estimated in the pro-
posed method. For more detailed information about the
computation of the froth thickness, see Lehikoinen et al.
(2011).
This estimate gives the average conductivity of the froth in the
measurement location based on the ERT measurements. For
more detailed information about the computation of the bulk
conductivity, see Lehikoinen et al. (2011).
5. Results
In this section, the results based on the simulated and real mea-
surements are evaluated.

Fig. 4. The results in Case 1 when measurements are simulated using mesh Msim2 . (a) The simulated froth thickness h (cm). (b) The simulated conductivity rb (mS cm1) of
the boundary of the bubbles (thick red line) and estimated conductivity rb (mS cm1) (thin blue line). (c) Simulated bubble size c (thick red line) and estimated bubble size c
^
(thin blue line). (d) The simulated conductivity rs (mS cm1) of slurry (thick red line) and estimated conductivity rs (mS cm1) (thin blue line). (For interpretation of the
references to color in this figure legend, the reader is referred to the web version of this article.)
8 A. Nissinen et al. / Minerals Engineering 69 (2014) 1–12

5.1. Results with simulated measurements
The measurements were simulated using two meshes Msim1
and Msim2 . The density of these meshes is deliberately different
in order to evaluate the effect of the discretization errors on the
estimation of the conductivity and bubble size. The inverse prob-
lem is solved using a mesh Minv , this mesh is coarser than the
meshes that are used for the simulation of the measurements in
order to avoid so-called inverse crime. The element size is constant
in all meshes and the structure of the elements in the meshes is
shown in Fig. 3. The details of the meshes are shown in Table 1.
The elements are isosceles triangles and the element size (mm)
in Table 1 refers to the length of legs of the triangle.
The measurements were simulated using the cylindrically sym-
metric forward model. The same adjacent current and measure-
ment patterns were used as in the real measurements, leading to
441 simulated voltage measurements, that is, V 2 R441 . Gaussian
mutually independent noise was added to the simulated measure-
ments. The added noise e was zero mean and the standard devia-
tion of the noise was 0.1% of the simulated voltages.
5.2. Simulated case 1
The first simulated test was conducted using two different
mesh densities for the simulation of the measurements. The
estimates and the correct parameter values are shown in Figs. 4
and 5. The results in Figs. 4 and 5 corresponds to simulated volt-
ages computed using meshes Msim2 and Msim1 , respectively. The
correct parameter values are shown with red thick lines and esti-
mated values are shown with blue thin lines. In these simulated
cases, the froth thickness h, conductivity rb of the boundary of
the bubbles and conductivity rs of slurry were kept constant. It
was found that the estimated rb is scaled by the fraction of the ele-
ment sizes in the mesh used for the simulation of the measure-
ments and for the inversion. These scaling factors are 0.5 mm/
1 mm = 0.5 and 0.25 mm/1 mm = 0.25. This scaling of the esti-
mated conductivity is due to the discretization errors The temporal
variation of the bubble size is estimated quite correctly. The bubble
size is estimated more correctly when the mesh density in the sim-
ulation of measurements is close to the one used in the inversion
(see Figs. 4 and 5). There is clear cross talk with the estimated bub-
ble size c^ and the conductivity rs in Fig. 5.
5.3. Simulated case 2
In the second simulated test, the froth thickness h and conduc-
tivity rb were also varied during the test. The results correspond-
ing to second test are shown in Fig. 6. The measurements were
simulated using mesh Msim2 in this case. The temporal variation
of rb is reliably estimated but the actual values are shifted due

Fig. 5. The results in Case 1 when measurements are simulated using dense mesh Msim1 . (a) The simulated thickness of froth (cm). (b) The simulated conductivity rb
(mS cm1) (thick red line) and estimated conductivity rb (thin blue line). (c) Simulated bubble size c (thick red line) and estimated bubble size c
^ (thin blue line). (d) The
simulated conductivity rs (mS cm1) (thick red line) and estimated conductivity rs (mS cm1) (thin blue line). (For interpretation of the references to color in this figure
legend, the reader is referred to the web version of this article.)
 A. Nissinen et al. / Minerals Engineering 69 (2014) 1–12 9

Fig. 6. The results in Case 2 (measurements are simulated using mesh Msim2 ). (a) The simulated thickness of froth (cm). (b) The simulated conductivity rb (mS cm1) (thick
red line) and estimated conductivity rb (mS cm1) (thin blue line). (c) Simulated bubble size c (thick red line) and estimated bubble size c
^ (thin blue line). (d) The simulated
conductivity rs (mS cm1) (thick red line) and estimated conductivity rs (mS cm1) (thin blue line). (For interpretation of the references to color in this figure legend, the
reader is referred to the web version of this article.)

to the difference between the simulation and inverse mesh densi-
ties. There seems to be nearly constant bias in the estimated con-
ductivity rb due to the discretization error. This bias is easy to
correct in simulated cases by using calibration. In case of real mea-
surements, the complication is that the actual thickness of the
boundary of the bubbles is unknown. Furthermore, the thin struc-
tures in the froth are difficult to model with standard FEM meshes.
Therefore, it is expected that the estimated rb will be biased in
case of real measurements.
The estimated bubble size follows the temporal variation of the
actual bubble size but it is in sometimes over/under estimated. The
bubble size is over or under estimated severely just after time
instants 100, 150 and before 200. At these time instants, the bubble
size c and conductivity rb simultaneously change and the actual
values have low probability with respect to the prior. The errors
in the estimated bubble size are largest when the actual conductiv-
ity rb is low. The reason for this is still unknown. It should be noted
that the estimates of the bubble sizes that have small probability
with respect to the prior model pðcÞ are expected to be erroneous.
In this work, the statistics of the approximation error approach was
computed using bubble size samples that were drawn from uni-
form density Uð3; 15Þ, therefore, the maximum simulated bubble
size 15 mm has a low probability. The same effect was observed
in the chest imaging application in Nissinen et al. (2011b). In the
cited paper, the estimated shape of chest was inaccurate when
the shape of actual chest was very improbable with respect the
prior model. The negative bubble sizes are due to the approxima-
tive Gaussian joint density of the parameters c and e. The
estimated bubble size c^ is unreliable when the actual bubble size
too far from the mean of the prior density pðcÞ. One possible
approach to prevent negative estimates would be to apply positiv-
ity constraint to the estimation of the bubble size. The average
computation time of the proposed method was 44 s for the estima-
tion of the conductivities and bubble size.
5.4. Results with real measurements
The proposed approach was tested in real process environment
at Inmet Pyhäsalmi Mine in Pyhäsalmi, Finland. The ERT measure-
ments were conducted using Outotec LevelSense™ sensor, details
of the sensor and measurement parameters are given in Section
2.1. The probe was immersed in a zinc flotation cell and measure-
ment data was collected. Simultaneously, the bubble size bsize and
bubble loading bload were measured using a camera based tech-
nique. The froth height h and bulk conductivity rbulk of the froth
were obtained from Outotec measurement system. The camera
based parameters and parameters from Outotec LevelSense™ were
used as reference parameters for the proposed method, see details
about the reference parameters in Section 4.3.
The estimates from the real measurements are shown in Fig. 7.
The correlation coefficient q between the estimates and the refer-
ence parameters are shown below the figures. The froth thickness
h is shown in Fig. 7(a). The estimated conductivity rb and the
camera based bubble loading bload are shown in Fig. 7(b). There
is a small shift in the estimates rb and bload after time 21:00. This
shift might be due to the different location of the ERT and camera
10 A. Nissinen et al. / Minerals Engineering 69 (2014) 1–12

Fig. 7. (a) The froth thickness h obtained from the Outotec LevelSense™. (b) Camera based estimate for the bubble loading bload (red thick line) and estimated conductivity rb
(mS cm1) of the boundary of the bubbles. (c) Camera based estimate for the bubble size bsize (red thick line) and ERT estimate for the bubble size c ^ (blue thin line).
(d) Estimated conductivity rs (mS cm1) of the slurry (red thick line). (For interpretation of the references to color in this figure legend, the reader is referred to the web
version of this article.)

measurement systems. Furthermore, the camera system measures
the surface of the froth and the ERT sensor gives information
from the whole froth volume near the measurement probe. The
estimated bubble size c ^ and camera based estimate bsize are
shown in Fig. 7(c). The actual values of the bubble sizes are differ-
ent with these two methods but the temporal variation of the
bubble size is very similar. It must be noted, that the froth model
used in this work is very approximative (square shaped bubbles)
and therefore the estimated bubble size very likely contain errors.
However, based on these results it seems that the temporal vari-
ation of the bubble size can be estimated using the proposed
method. The estimated conductivity rs of the slurry is shown in
Fig. 7(d). In the simulated cases it was observed that the changing
bubble size was reflected also to the estimated conductivity of
the slurry, see Fig. 5. The same type of reflection is not seen (at
least not so clearly) with the measured data.
The bulk conductivity rbulk and estimated bubble size c ^ are
shown in Fig. 8(a). There is a very clear negative correlation
(correlation coefficient q ¼ 0:94) between these two parameters.
It seems that the bulk conductivity estimate is mostly affected by
the bubble size at least during these measurements. The high cor-
relation of the bubble size and bulk conductivity was also reported
in Lehikoinen et al. (2011). The estimated conductivities rs and rb
are shown in Fig. 8(b). There is clear positive correlation between
these variables, q ¼ 0:89.
The correlation coefficients between the estimated variables
and reference parameters are shown in Table 2. As can be seen
the, camera based bubble size bsize is highly correlated with the
proposed ERT based bubble size c ^. Furthermore, bubble size c ^
and the bulk conductivity rbulk are also highly correlated. This
gives evidence that the proposed method can estimate the tempo-
ral variation of the bubble size. Furthermore, the froth height h,
 A. Nissinen et al. / Minerals Engineering 69 (2014) 1–12 11

Fig. 8. (a) Estimated bulk conductivity rbulk (mS cm1) of the froth (red thick line) and estimated bubble size c
^ using ERT method (blue thin line). (b) Estimated conductivity
rs (mS cm1) (red thick line) and estimated conductivity rb (mS cm1) (blue thin line). (For interpretation of the references to color in this figure legend, the reader is referred
to the web version of this article.)

Table 2 It was found that the ERT and camera based estimates for the
The correlation between variables, where h is the height of froth, bload is the bubble bubble size has very high correlation (correlation coefficient was
load estimated with camera based method, bsize is the bubble size estimated with
camera based method, c ^ is ERT based estimate for bubble size, rs is the estimated 0.92). The absolute values of the bubble size estimates with these
conductivity of slurry, rb is the estimated conductivity of boundary of bubbles and two modalities were different. There are several reasons for this.
rbulk is the estimated bulk conductivity. The camera based method images only the surface of the froth
h bload bsize c^ rs rb rbulk and the ERT measurement probe collects information from the
whole volume of froth. The model of the froth structure in this
h 1.00 0.75 0.81 0.90 0.74 0.52 0.93
study was very approximative (square shaped bubbles) and this
bload 0.75 1.00 0.73 0.67 0.76 0.66 0.73
bsize 0.81 0.73 1.00 0.92 0.67 0.55 0.86 approximative model induces errors to the estimates. Furthermore,
c^ 0.90 0.67 0.92 1.00 0.67 0.51 0.94 cylindrically symmetric forward model was used. The real froth is
rs 0.74 0.76 0.67 0.67 1.00 0.89 0.62 not cylindrically symmetric and this approximation causes also
rb 0.52 0.66 0.55 0.51 0.89 1.00 0.42 errors to the estimated bubble sizes. The proposed method could
rbulk 0.93 0.73 0.86 0.94 0.62 0.42 1.00
be improved significantly by modeling the approximation errors
due to approximative froth model and cylindrically symmetric for-
ward model. The modeling of these errors would require the full
^ and bulk conductivity rbulk are strongly correlated. 3D model for froth and it is therefore computationally very
bubble size c
demanding task. In spite of these approximations, it seems that
The bubble load bload has strongest correlation with the conductiv-
the temporal variation of the bubble size can be estimated using
ity rs of the slurry.
the proposed approach.
The conductivity of the boundary of the bubbles and the camera
6. Conclusions based bubble loading was also found to correlate (correlation coef-
ficient was 0.66). Although the correlation coefficient is smaller
In this paper, an approach was proposed to estimate the bubble than in the case of bubble size, there seems to be a clear similarity
size and conductivity of the boundary of the bubbles using ERT in the estimated bubble loading and the conductivity of the bound-
probe measurements. The bubble size is important parameter that ary of the bubbles. In some time intervals there was a clear time shift
is used in the control of the froth flotation process and it is believed between these two estimates. The time shift can be due to difference
that the conductivity of the boundary of the bubbles gives informa- in the bubble loading at different depths in the froth. Furthermore,
tion on the amount of solids (bubble loading) in the froth. The pro- the location of the camera measurement and ERT measurement sys-
posed approach was evaluated with simulated and real tem within the flotation cell is different which can cause the differ-
measurements. The real measurements were conducted using ence in the estimates. The bubble size and the bulk conductivity was
Outotec LevelSense™ sensor in Inmet Pyhäsalmi mine. The refer- also found to correlate very clearly (correlation coefficient was
ence estimates in this work were the bubble size and bubble load- 0.94). This suggests that the bulk conductivity estimate gives
ing obtained from camera based measurements. In the earlier mainly information on the bubble size in the froth.
approaches to monitor the froth flotation using ERT, the conductiv- In the future, the proposed method should be tested with longer
ity has been modeled to be smoothly varying. The complication test measurements from flotation process. The correlation between
with the earlier approaches is that both the bubble size and bubble the estimated variables can be different when longer period of
loading affect to the conductivity and therefore the estimate of the measurements are used. Furthermore, the correlation between
conductivity of the froth can give misleading information. One of the estimated variables can be different when different ore is used.
the aim of this work was to obtain more precise information by Based on the results shown here, it seems that the estimation of
separating the bubble size and conductivity of the boundary of the temporal variation of the bubble size is possible using the pro-
the bubbles in the computation model. posed approach.
12 A. Nissinen et al. / Minerals Engineering 69 (2014) 1–12
Acknowledgments
The work was supported by the TEKES projects 2172/31/2010
and 2180/31/2011. Furthermore, this work was supported by the
Academy of Finland (project 250215 Finnish Centre of Excellence
in Inverse Problems Research).
References
Arridge, S.R., Kaipio, J.P., Kolehmainen, V., Schweiger, M., Somersalo, E., Tarvainen,
T., Vauhkonen, M., 2006. Approximation errors and model reduction with an
application in optical diffusion tomography. Inv. Probl. 22, 175–195.
Banasiak, R., Ye, Z., Soleimani, M., 2012. Improving three-dimensional electrical
capacitance tomography imaging using approximation error model theory. J.
Electromagnetic Waves Appl. 26, 411–421.
Barbian, N., Hadler, K., Cilliers, J.J., 2006. The froth stability column: measuring froth
stability at an industrial scale. Miner. Eng. 19, 713–718.
Barbian, N., Hadler, K., Ventura-Medina, E., Cilliers, J.J., 2005. The froth stability
column: linking froth stability and flotation performance. Miner. Eng. 18, 317–
324.
Borcea, L., 2002. Electrical impedance tomography. Inv. Probl. 18, R99–R136.
Calvetti, D., Somersalo, E., 2007. An Introduction to Bayesian Scientific Computing –
Ten Lectures on Subjective Computing. Springer, iSBN 978-0-387-73393-7.
Cheney, M., Isaacson, D., Newell, J.C., 1999. Electrical impedance tomography. SIAM
Rev. 41, 85–101.
Cilliers, J., Wang, M., Neethling, S., 1999. Measuring flowing foam density
distributions using ERT. In: Proc 1st World Congress on Industrial Process
Tomography, pp. 108–112.
Farrokhpay, S., 2011. The significance of froth stability in mineral flotation–a
review. Adv. Colloid Interface Sci. 166, 1–7.
Grau, R.A., Heiskanen, K., 2002. Visual technique for measuring bubble size in
flotation machines. Miner. Eng. 15, 507–513.
Heikkinen, L.M., Vilhunen, T., West, R.M., Vauhkonen, M., 2002. Simultaneous
reconstruction of electrode contact impedances and internal electrical
properties: II. Laboratory experiments. Meas. Sci. Technol. 13, 1855–1861.
Heino, J., Somersalo, E., 2004. A modeling error approach for the estimation of
optical absorption in the presence of anisotropies. Phys. Med. Biol. 49, 4785–
4798.
Hu, S., Ofori, P., Firth, B., 2009. Monitoring of froth stability using electrical
impedance spectroscopy. Int. J. Miner. Process. 92, 15–21.
Kaartinen, J., Hätönen, J., Hyötyniemi, H., Miettunen, J., 2006. Machine-vision-based
control of zinc flotation – a case study. Control Eng. Pract. 14, 1455–1466.
Kaipio, J.P., Kolehmainen, V., Somersalo, E., Vauhkonen, M., 2000. Statistical
inversion and Monte Carlo sampling methods in electrical impedance
tomography. Inv. Probl. 16, 1487–1522.
Kaipio, J.P., Somersalo, E., 2005. Statistical and computational inverse problems.
Applied Mathematical Sciences, vol. 160. Springer-Verlag.
Kaipio, J.P., Somersalo, E., 2007. Statistical inverse problems: discretization, model
reduction and inverse crimes. J. Comput. Appl. Math. 198, 493–504.
Kolehmainen, V., Tarvainen, T., Arridge, S.R., Kaipio, J.P., 2011. Marginalization of
uninteresting distributed parameters in inverse problems – application to
diffuse optical tomography. Int. J. Uncertainty Quantification 1 (1), 1–17.
Kourunen, J., Niitti, T., Heikkinen, L.M., 2011. Application of three-dimensional
electrical resistance tomography to characterize gas holdup distribution in
laboratory flotation cell. Miner. Eng. 24, 1677–1686.
Kourunen, J., Rinne, A., Saloheimo, K., Heikkinen, L.M., 2008. Electrical tomography
imaging of flotation process in a mechanical flotation cell. In: Proc 5th
International Symposium on Process Tomography.
Lähivaara, T., Ward, N.F.D., Huttunen, T., Koponen, J., Kaipio, J.P., 2014. Estimation of
aquifer dimensions from passive seismic signals with approximate wave
propagation models. Inv. Probl. 30, 015003.
Lehikoinen, A., Finsterle, S., Voutilainen, A., Heikkinen, L.M., Vauhkonen, M., Kaipio,
J.P., 2007. Approximation errors and truncation of computational domains with
application to geophysical tomography. Inverse Probl. Imaging 1, 371–389.
Lehikoinen, A., Laakkonen, P., Vauhkonen, M., Rinne, A., Saloheimo, K., Lähteenmäki,
S., 2011. Measuring flotation process using probe sensor based on 3d electrical
resistance tomography. In: Proc of Flotation11.
Lipponen, A., Kolehmainen, V., Romakkaniemi, S., Kokkola, H., 2013a. Correction of
approximation errors with random forests applied to modelling of cloud droplet
formation. Geosci. Model Dev. 6, 2087–2098.
Lipponen, A., Seppänen, A., Kaipio, J., 2013b. Electrical impedance tomography
imaging with reduced-order model based on proper orthogonal decomposition.
J. Electron. Imaging 22, 023008.
Maldonado, M., Desbiens, A., der Villar, R., 2008. An update on the estimation of the
froth depth using conductivity measurements. Miner. Eng. 21, 935–939.
Neethling, S.J., Cilliers, J.J., 2003. Modelling flotation froths. Int. J. Miner. Process. 72,
267–287.
Nissinen, A., Heikkinen, L.M., Kaipio, J.P., 2008. The bayesian approximation error
approach for electrical impedance tomography—experimental results. Meas.
Sci. Technol. 19, 015501.
Nissinen, A., Kolehmainen, V., Kaipio, J.P., 2011a. Compensation of modelling errors
due to unknown domain boundary in electrical impedance tomography. IEEE
Trans. Med. Im. 30 (2), 231–242.
Nissinen, A., Kolehmainen, V., Kaipio, J.P., 2011b. Reconstruction of domain
boundary and conductivity in electrical impedance tomography using the
approximation error approach. Int. J. Uncertainty Quantification 1, 203–222.
Nissinen, A., Sbarbaro, D., Heikkinen, L., Vauhkonen, M., 2013. Reduced forward
models in electrical impedance tomography with probe geometry. Inverse
Probl. Sci. Eng., doi: 10.1080/17415977.2013.872098..
Normi, V., Lehikoinen, A., Mononen, M., Rintamäki, J., Maksimainen, T., Luukkanen,
S., Vauhkonen, M., 2009. Predicting collapse of the solid content in a column. In:
Proc of Flotation 09.
Reunanen, J., Mononen, M., Vauhkonen, M., Lehikoinen, A., Kaipio, J.P., 2011.
Machine learning approach for locating phase interfaces using conductivity
probes. Inverse Probl. Sci. Eng. 19, 879–902.
Tarantola, A., 2004. Inverse Problem Theory and Methods for Model Parameter
Estimation. SIAM, Philadelphia.
Tarvainen, T., Kolehmainen, V., Pulkkinen, A., Vauhkonen, M., Schweiger, M.,
Arridge, S.R., Kaipio, J.P., 2010. Approximation error approach for compensating
modelling errors between the radiative transfer equation and the diffusion
approximation in diffuse optical tomography. In: Inv. Probl. 26, 015005 (18pp).
Vauhkonen, M., Lehikoinen, A., Kourunen, J., Mononen, M., Huuskonen, J., 2013.
Flotation process optimization using Outotec FrothControl solution. In:
Canadian Mineral Processors 2013.