You are on page 1of 6

Multivariable Predictive Control of a Pilot Flotation Column

Danny Calisaya
*
, ric Poulin
*
, Andr Desbiens
*
, Ren del Villar

, Alberto Riquelme
*
*
Department of Electrical and Computer Engineering

Department of Mining, Materials and Metallurgical Engineering


LOOP (Laboratoire dobservation et doptimisation des procds), Universit Laval, Qubec, QC, Canada

AbstractThe aim of this work is the control of hydrodynamic
variables of a pilot flotation column working with a three-phase
system (air-water-ore) in an industrial environment. Since
hydrodynamic variables are closely related to metallurgical and
economical performances of the unit, the implementation of
such a control strategy is crucial to optimize its operation. The
hydrodynamic variables here considered are the gas hold-up in
the collection zone and the fraction of wash-water underneath
the interface. They are controlled by manipulating gas flow rate
and wash-water flow rate respectively. Hydrodynamic variables
are controlled through a constrained model predictive control
(MPC) strategy. This choice is dictated by the interdependency
of these controlled variables and the capability of predictive
controllers to handle process constraints. Identification tests
lead to a representative model of the system under nominal
operating conditions but indicate variations of process behavior
with changing operating regimes. Results show good control
performances, confirming the potential use of MPC to control
hydrodynamic variables for real-time optimization of full-scale
flotation columns.
I. INTRODUCTION
N the effort to enrich the ore from the mine, mineral
separation by flotation columns is an important area of
research for the metallurgical industry. Although the first
commercial use of this process goes back almost 30 years,
the research on the control of hydrodynamic variables of the
process as well as sensors to measure them are not yet fully
developed. This step is necessary for the development and
the implementation of an indirect (or hierarchical) real-time
optimization strategy [1]. For the first time, simultaneous
constrained control of gas hold-up in the collection zone and
the fraction of wash-water underneath the interface in an
industrial environment are presented in this paper.
Different advanced control techniques, such as adaptive
self-tuning control [2], fuzzy predictive control [3],
decentralized control [4], multivariable nonlinear predictive
control [5], or predictive control based on a neural network
model [6] have been tested for optimizing and controlling the
column flotation process. From all these techniques, the
model predictive control (MPC) is the interest of several
researchers given its capability of handling interactions and
constraints which is necessary for real-time optimization. To
mention a few, constrained nonlinear predictive control
based on IMC-optimization [7], multivariable predictive
control of laboratory flotation columns, [8]-[10] could be
cited as examples. Extensive and critical reviews regarding
instrumentation, control, and optimization of flotation
columns can be found in [11] and [1].
This article presents the development, the implementation,
and the evaluation of MPC to control strategic hydrodynamic
variables of a three phase (air-water-ore) pilot flotation
column in an industrial environment (Agnico-Eagle, Laronde
Division). The hydrodynamic variables are the volumetric
wash-water fraction underneath the interface
w
and the gas
hold-up
g
in the collection zone. This work extends a
previous study achieved on a two-phase (water-air) system
[12], where bias rate, i.e. net downward flow rate of water
through the interface, is the variable used to assess the
cleanliness of the froth [13]. In the present context, the
estimation of bias is done using the volumetric wash-water
fraction underneath the interface instead of bias directly
since these variables are linearly related [14] and
measurement of
w
is simpler. The method could be easily
implemented on a full scale column using a commercial
g

sensor and the proposed sensor for
w
. Such an approach
would not require modification of existing local control
loops.
Section II gives a brief description of the flotation column
process and introduces manipulated and controlled variables
as well as the related instrumentation. Section III discusses
the identification of the system and the validation of
resulting dynamic linear models. The MPC algorithm and
tuning is presented in section IV. Section V presents
industrial results and analyses controller performances,
whereas conclusions are given in Section VI.
II. PROCESS DESCRIPTION
Fig. 1 is a schematic of a flotation column showing its
three input streams: pulp feed, air injection, and wash-water
addition, as well as its two output streams: concentrate and
tailings. In normal operating conditions, the column content
shows two distinct regions in terms of the amount of air
content: the collection zone in the lower part, with less than
30% air, and the cleaning zone (upper part) with more than
70% air.
To perform the separation between valuable mineral and
gangue, some chemical reagents (collectors, frothers,
activators or depressants, and pH modifiers) are usually
added, either at an earlier stage (conditioning) or at the
column itself. The produced froth carries the valuable
mineral particles, due to its hydrophobic properties (natural
or induced) and overflows into the column concentrate
I
2012 American Control Conference
Fairmont Queen Elizabeth, Montral, Canada
June 27-June 29, 2012
978-1-4577-1096-4/12/$26.00 2012 AACC 4022


launder, whereas the hydrophilic particles (gangue) are
withdrawn through the bottom port as tailings. The added
wash-water helps in improving the concentrate quality,
releasing the entrained undesirable particles from the froth [1].

Fig. 1. Diagram of a flotation column
The pilot flotation column used in this work consists in a
7.32 m height and 151 mm diameter Plexiglas tube, provided
with the necessary feed and tailing ports, froth overflow
launder, a perforated ring for wash-water addition over the
froth, and a porous sparger at the bottom for air injection.
The column is fully automated, with local control loops to
regulate all flow rates (tailings, wash-water, and air) and
froth-depth, as shown in Fig. 2. A controller for
hydrodynamic variables (
w
and
g
) and, obviously, the
necessary sensors for measuring all relevant variables, are
detailed here after. The pulp processed during tests
corresponds to the feed of the 3
rd
cleaner column of the
copper circuit, containing 15% solids and having 99% of the
ore particle-size smaller than 100 m.
The froth depth (interface) is measured using the
conductivity profile based sensor and determined by an
algorithm relying on the largest slope method developed by
[15], a weighted average interpolation of the two possible
froth depth values [16]. The conductivities are measured by a
Field-Programmable Gate Array (FPGA) [17].
The measure of gas hold-up (volumetric fraction of gas) in
the collection zone is made using a sensor installed near the
interface and the method described by [18]-[19] and
evaluated by Maxwell's equation for electrolyte mixtures
[20], proposed by [19], and used by [12]. The final relation
for measuring the gas hold-up is:

|
|
.
|

\
|

=
sgl sl
sgl sl
g
k k
k k
5 . 0
100 c

(1)

where k
sl
is the conductivity of pulp only (solid and liquid)
and k
sgl
is the conductivity of pulp-gas mixture. These are
measured by the siphon and open cells respectively (Fig. 2).
The conventional bias rate variable is replaced in this work
by the concept of fraction of wash-water underneath the
interface proposed by [14] and adapted by [21]. Considering
that the gas hold-up is measured sufficiently close to the
interface and assuming no solid conductivity, then the
additivity rule in a three-phase system leads to the following
expression:
|
|
|
|
|
|
.
|

\
|
u
|
|
.
|

\
|

=
) 1 (
1
1 5 . 0
'
100
s w f
g
g
sgl f
w
k k
k k
c
c
c

(2)
where k
f
is the conductivity of feed pulp, k
sgl

is the
conductivity of pulp-gas mixture underneath the interface, k
w

is the conductivity of wash-water and
s
u is the percentage
of solid into pulp. More details about this relation are found
in [21].

Fig. 2. Process and instrumentation diagram of the pilot flotation
column
The first manipulated variable is the gas superficial
velocity. It is calculated from the gas mass flow meter
reading
ref
g
e at reference conditions (21.1
o
C and 101.3
kPa) and the gas specific gravity
ref
g
:
c
ref
g
ref
g
c
ref
g
ref
g
A
A
Q
J

e
= =

(3)
Tailings
Air (J
g
)
Concentrate
Wash water
(J
w
)
Ew
g
Collection
zone
Cleaning
zone
Feed
Interface
w
PT
FPGA
Wash
Water
Feed
Concentrate
Tailings
Input
Opto 22
TT2
Speed
Reference
Level
Set-Point
Air
Set-Point
FT
FIC
Mass Flow
Controller
Wash Water
Set-Point
LC

Air

Syphon
Cell
Open
Cell
TT1
FC FT
Conductivities
Level
Estimation
4023


where
ref
g
Q is the gas volumetric flow rate and A
c
is the cross
sectional area of column. The superficial gas velocity J
g
must
be compensated for temperature and pressure at test
conditions as shown below:
|
.
|

\
| +
|
.
|

\
|
+
=
16 . 294
15 . 273
23 . 1033
23 . 1033 T
P
J J
ref
g g
(4)
where the absolute pressure P is measured in cm H
2
O and
the temperature T

is in Celsius degrees. The pressure PT and
temperature sensors TT1 are shown in Fig. 2
The second manipulated variable is the wash-water
superficial velocity J
w
, obtained by dividing the volumetric
flow rate by the column cross sectional area:
c
w
w
A
Q
J = (5)
The froth depth is controlled using a PI controller
manipulating the speed of tailings pump. The wash-water
flow rate is controlled by a PI controller implemented with a
Moore Mycro 353 controller. Graphical interfaces and data
acquisition are performed by a HMI/SCADA software
iFIX working under a Windows XP operating system.
An Opto 22 I/O system is used to centralize sensor and
actuator signals as shown in Figs. 2 and 3. Algorithms for the
froth depth control and multivariable predictive control of
g

and
w
are implemented in MatLab. All signals are sampled
at a two seconds interval.
Fig. 3. HMI/SACAD system of the pilot flotation column
III. IDENTIFICATION OF THE PROCESS TRANSFER FUNCTIONS
The identification of the process transfer functions has
been made using the iterative prediction error minimization
method provided by MatLab System Identification
Toolbox. Based on prior knowledge about the system, low-
order continuous-time models were selected [12].
The identified transfer functions, and the standard-
deviation of each parameter, are as follows:
1 ) 2 . 62 ( 2 . 358
) 3 . 38 ( 2 . 272
) (
) (
) (
) 2 . 2 ( 74
,
+

= =

s
e
s J
s
s G
s
w
w
J
w w
c
c

(6)
1 ) 2 . 10 ( 5 . 2
) 5 . 1 ( 04 . 0
) (
) (
) (
) 7 . 2 ( 2 . 56
,
+

= =

s
e
s J
s
s G
s
w
g
J
w g
c
c

(7)
1 ) 2 . 77 ( 195
) 3 . 12 ( 2 . 45
) (
) (
) (
) 2 . 4 ( 8 . 30
,
+

= =

s
e
s J
s
s G
s
g
w
J
g w
c
c

(8)
1 ) 4 . 27 ( 8 . 89
) 7 . 4 ( 0 . 25
) (
) (
) (
) 1 . 0 ( 8 . 74
,
+

= =

s
e
s J
s
s G
s
g
g
J
g g
c
c

(9)
The standard deviation of the gain, time constant, and
delay of
w w
J
G
, c
and
g g
J
G
, c is less than 25%, which
allows to consider the model found as a correct
approximation of the real process. Equation 7 shows a
standard deviation larger than nominal values for the gain
and time constant suggesting that the relation between J
w
and

g
is not significant, which is reasonable from a physical
point of view. The addition of wash-water has few impacts
on the gas hold-up in the collection zone. The transfer
function
w g
J
G
, c will thus be considered as zero for control
purposes. The standard deviation of the parameters of
g w
J
G
, c are in the range of 35%. This slightly high standard
deviation can be explained by disturbances within the
process (variations in the feed composition) during tests.

Fig. 4. Identification and validation data for
w w
J
G
, c
and
w g
J
G
, c
: input/output data (-), model (- -)
Speed
reference
(Feed)
RS232/
Modbus
FPGA
Moore
(Mycro
353)
Windows XP


MatLab


IFix

RS 485
I/O Opto22
Speed
reference
(Wash Water)
Air Set-Point
Electrodes
Temperatures Pressure Air
flow rate
Wash Water
flow rate
Speed
reference
(Tailing)

g
(
%
)

w
(
%
)
J
w
(
c
m
/
s
)

g
(
%
)

w
(
%
)
J
w
(
c
m
/
s
)
Time (s)
Time (s)
Identification
Validation
4024


Fig. 4 shows the fit of models for identification and
validation data for
w w
J
G
, c
and
w g
J
G
, c where the operating
point has been removed. In case of
w w
J
G
, c
, both
identification and validation tests shows a good fit. As
mentioned above for
w g
J
G
, c , no strong relation between J
w

and
g
can be observed in Fig. 4. The identification and
validation tests for
g w
J
G
, c and
g g
J
G
, c are presented in
Fig. 5. For both transfer functions, the model fitting is
reasonable.
Finally, the percentage of variance explained by the
different models (best fit) are gathered in Table I for
identification and validation data. In case of
w w
J
G
, c
and
g g
J
G
, c transfer functions, the percentage for validation
data is higher than the one obtained for identification data.
This is explained by the relatively higher disturbance level
during identification test. Regarding
w g
J
G
, c , the best fit is
nearly zero as expected.

Fig. 5. Identification and validation data for
g w
J
G
, c
and
g g
J
G
, c
:
input/output data (-), model (- -)
TABLE I
EXPLAINED VARIANCE FOR IDENTIFICATION AND VALIDATION DATA
Transfer Function Identification Validation
w w
J
G
, c

50 % 63 %
w g
J
G
, c

1 % 0.5 %
g w
J
G
, c

41 % 34 %
g g
J
G
, c

68 % 75 %
In summary, the transfer functions of the system used for
designing the controller in the next section are:
(
(
(
(
(

(
(
(
(
(
(

+
+

+
=
(
(
(
(
(


) (
) (
1 90
25
0
1 195
45
1 358
272
) (
) (
75
30 74
s J
s J
s
e
s
e
s
e
s
s
g
w
s
s s
g
w
c
c

(10)

IV. DESIGN OF MPC

The control strategy used for the pilot flotation column was
developed using the MatLab MPC toolbox. The MPC
controller output, | |
T
n
k u k u k u
u
) ( ) ( ) ( = , where n
u
is the
number of manipulated variables, is defined by:
) | ( ) 1 ( ) ( k k u k u k u A + = (11)
It is the result of the minimization of the following
optimization problem with respect to the sequence of input
increments and to the slack variables:
|
| | |

+
|
.
|
+ A + + +

\
|
+ +


=
A

= =
+ A A
u
p y
c
n
j
j
u
j j
H
i
n
j
j
y
j
e k k H u k k u
e k i k u w i k r
k i k y w
1
2
2 2
1
0 1
)), | 1 ( ),... | ( (
) | ( ) 1 (
) | 1 ( min

(12)
The optimization problem is subject to following
constraints:
) ( ) / ( ) ( i u k i k u i u
jmax j jmin
s + s (13)
e i y k i k y e i y
jmax j jmin
+ s + s ) ( ) / ( ) ( (14)
0 ) / ( = + A k h k u , 1 , , =
p c
H H h (15)
0 > e

(16)
where ) | ( k k u A is the first element of the optimal sequence,
H
c

and H
p
are the control and prediction horizon
respectively, n
y
is the number of system outputs,
y
j
w is the
weight for the output variable j,
u
j
w
A
is the weight for change
in each manipulated variable and
j
u A is the change in
manipulated variable j. The predicted output is
j
y and
j
r is
the reference value over the whole prediction horizon.
The slack variable e is used to relax the constraints
imposed to controlled variables. The weight on the slack
variable penalizes the violation of constraints and is given
by:
{ }
u
j
y
j
w w
A
= , max 10
5
(17)
The larger is with respect to input and output weights, the
more the constraint violation is penalized as explained in [22].

g
(
%
)

w
(
%
)
J
g
(
c
m
/
s
)

g
(
%
)

w
(
%
)
J
g
(
c
m
/
s
)
Time (s)
Time (s)
Identification
Validation
4025


The global specification for MPC tuning was to obtain a
closed-loop response with dynamics similar to the open-loop
system with a limited coupling between
g
and
w
. The
selected control period is 10 seconds. This value takes into
account the overall dynamics of the system whereas allowing
the controller to properly reject disturbances. The prediction
horizon H
p
was set to 150 control periods and H
c
to one.
Constraints imposed to the manipulated and controlled
variables as well as operating points and are listed in
Table II.
TABLE II
CONSTRAINTS ON MANIPULATED AND CONTROLLED VARIABLES
Variable Unit Minimum Maximum
w
J

cm/s 0.05 0.18
g
J

cm/s 0.1 0.5
w
c

% 10 40
g
c

% 5 20
The most critical constraints are the maximum of J
w
, the
minimum of
w
, and the minimum of J
g
. The first prevents
breaking the froth, the second ensures its cleanliness, and the
last is necessary to sustain froth. Other constraints ensure
working within the preselected and feasible operating range.
Weight values for manipulated and controlled variables are
given in Table III.
Fig. 6 presents a simulation of the closed-loop system. It
shows a set-point change on
w
at 100 seconds and two set-
point changes on
g
at 2 000 and 4 000 seconds. For each set-
point change, the controller smoothly adjusts the manipulated
variables to reach the reference within an acceptable time
response whereas the coupling effect on
w
is limited which
is coherent with the previously stated specification.

Fig. 6. Simulation of control performance: manipulated/controlled
variable (-), set-point (- -), constraint (- .)
TABLE III
WEIGHTS ON MANIPULATED AND CONTROLLED VARIABLES
Variable Weight (w
u
) Variable Weight (w
y
)
w
J

210
w
c

0.26
g
J

210
g
c

0.26
V. INDUSTRIAL RESULTS
The controller designed in Section IV was implemented on
the pilot flotation column with the same parameter settings.
The froth depth set-point was kept constant at 36 cm during
tests.
Fig. 7 presents control performances obtained for the first
test (Test A). The set-point of the fraction of wash-water
underneath the interface was changed at 100, 2 000, 3 600,
5 400 and 7 600 seconds, whereas the set-point of gas hold-
up was kept constant. For the first three steps on
w
, it is
worth noticing that the manipulated variable J
w
is
momentarily saturated. The time response for
w
is
acceptable but the control action J
w
is slightly more
aggressive than the one observed in simulation. As expected,
set-point changes on
w
has few impacts on
g
.

Fig. 7. Test A: manipulated/controlled variable (-), set-point (- -),
constraint (- .)
Fig. 8 presents the results of the second test (Test B) where
a disturbance on
w
was introduced at 450 seconds and the
gas hold-up set-point changed at 1 600, 3 600 and 6 400
seconds. In this case, control actions are rather smooth
indicating that the time response could be reduced by a
proper readjustment of MPC parameters. However, both test
results generally comply with performance obtained in
simulation.
J
g
(
c
m
/
s
)

g
(
%
)
J
w
(
c
m
/
s
)


w
(
%
)
Time (s)
J
g
(
c
m
/
s
)

g
(
%
)
J
w
(
c
m
/
s
)


w
(
%
)
Time (s)
4026



Fig. 8. Test B: manipulated/controlled variable (-), set-point (- -)
VI. CONCLUSIONS
The paper described the development, implementation and
evaluation of a controller for hydrodynamic variables of a
pilot flotation column operating in an industrial environment.
The considered variables are the gas hold-up and the wash-
water fraction underneath the interface. The selected control
algorithm is constrained multivariable MPC. Results have
shown that MPC is able to control the process while
respecting constraints and has delivered interesting
performances in the context of an industrial campaign for
both set-point tracking and disturbance rejection, under
various operating conditions. It is thus reasonable to consider
that such a controller could be successfully integrated in a
real-time optimization strategy. Further works will notably
focus on finding relationships between froth depth, gas hold-
up and wash water fraction underneath the interface and the
metallurgical and economic performance of the unit. The
final objective is the real-time optimization of a full-scale
column.
ACKNOWLEDGEMENT
The authors would like to acknowledge Agnico-Eagle
Laronde Division, XStrata-Nickel, COREM and NSERC for
providing the necessary funding for this project, and to
Agnico-Eagle personnel for on-site support.
REFERENCES
[1] D. Sbrbaro and R. del Villar (Eds)., Advanced control and supervision
of mineral Proc. Plants, Springer, 2010.
[2] S.-L. Jms-Jounela, "Simulation study of self-tuning adaptive control
for rougher flotation," Powder Technology, vol. 69, no 1, pp. 33-46,
1992.
[3] S. M. Vieira, J. M. C. Sousa, and F. O. Durao, "Real-time fuzzy
predictive control of a column flotation Process," Fuzzy Systems
Conf., 2007. FUZZ-IEEE, London, pp. 1-6, 2007.
[4] M. Maldonado, A. Desbiens, and R. del Villar, "Decentralized control
of a pilot flotation column: A 3X3 system," Canadian Metallurgical
Quarterly, vol. 47, no. 4, pp. 377-386, 2008.
[5] M. Milot, A. Desbiens, R. del Villar, and D. Hodouin, "Identification
and multivariable nonlinear predictive control of a pilot flotation
column," XXI Int. Mineral Proc. Congress. IFAC, pp. 137142, 2000.
[6] M. Swati, "Artificial neural network based system identification and
model predictive control of a flotation column," Journal of Process
Control, vol. 19, pp. 991-999, 2009.
[7] A. Desbiens and J. Bouchard, "Constrained nonlinear predictive
control based on IMC-optimization," 11th IFAC Symposium on
Automation in MMM Proc., 2004.
[8] O. D. Chuk, V. Mut, E. Nez, and L. Gutierrez, "Multivariable
predictive control of froth depth and gas holdup in column flotation,"
Tokyo, Japan, pp. 87-91, 2001.
[9] E. Nez, A. Desbiens, R. del Villar, and C. Duchesne, "Multivariable
predictive control of a pilot flotation column. Part 2: identification and
control," Int. Conf. on Mineral Process Modeling, Simulation and
Control, pp. 291301, 2006.
[10] M. Maldonado, A. Desbiens, R. del Villar, E. Poulin, and A. Riquelme,
"Nonlinear control of bubble size in a laboratory flotation column," in
Symposium on Automation in MMM Proc., Cape Town, South Africa,
pp. 19-24, 2010.
[11] J. Bouchard, A. Desbiens, R. del Villar, and E. Nunez, "Column
flotation simulation and control," 2009.
[12] M. Maldonado, A. Desbiens, and R. del Villar, "Potential use of model
predictive control for optimizing the column flotation process.," Int.
Journal of Mineral Proc., 2009.
[13] J. A. Finch and G.S. Dobby, Column Flotation, Pergamon Press,
Oxford, UK., 1990.
[14] M. Maldonado, A. Desbiens, R. del Villar, and J. Chirinos, "On-line
bias estimation using conductivity measurements," Minerals
Engineering, vol. 21, pp. 851855, 2008.
[15] C. Gomez, A. Uribe-Salas, J. Finch, and B. Huls, "A level detection
probe for industrial flotation columns," in Proceedings of an Int.
Symposium of Complex Ores, Ores Halifax, Canada , 1989, pp. 325-
334, vol. II.
[16] M. Maldonado, A. Desbiens, and R. del Villar, "An update on the
estimation of the froth depth using conductivity measurements,"
Minerals Engineering, vol. 21, no. 12-14, pp. 856860, 2008.
[17] A. Riquelme, A. Desbiens, Poulin, and Ren del Villar, "A novel
method for measuring conductivity of dispersions," submitted to
Measurement, 2012.
[18] A. Uribe-Salas, C. O. Gomez, and J.A. Finch, "Bias detection in
flotation columns," Agar, G., Huls, B., Hyma, D. (Eds.), Column91
Proceedings of an Int. Conf. on Column Flotation, vol. 2. Canadian
Institute of Mining, Metallurgy and Petroleum, pp. 391407, 1991.
[19] F. Tavera and R. Escudero, "Gas hold-up and solids hold-up in
flotation columns: on-line measurement based on electrical
conductivity," Transactions of the Institutions in MMM Proc. and
Extractive Metallurgy, vol. 11, 2002.
[20] J. C. Maxwell, A Treatise of electricity and magnetism Oxford
University press, Oxford University Press, London UK, 1892.
[21] R.-M. Estban, "Validation industrielle de la mesure du diffrentiel
deau dans une colonne de flottation." MS thesis, Department of
Mining, Materials and Metallurgical Engineering, Laval University,
Qubec, Canada, 2011.
[22] H. Mukai and E. Polak, "A second-order method for the general
nonlinear programming problem," Journal of Optimization Theory and
Applications, vol. 26, no. 4, pp. 515-532, 1978.


J
g
(
c
m
/
s
)

g
(
%
)
J
w
(
c
m
/
s
)


w
(
%
)
Time (s)
4027

You might also like