Minerals Engineering 123 (2018) 117–127

Contents lists available at ScienceDirect

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

Mineral processing plant data reconciliation including mineral mass balance T

constraints
⁎
Maryam Sadeghi , Daniel Hodouin, Claude Bazin
Department of Mining, Metallurgy and Materials Engineering, Université Laval, Quebec City, QC G1V 0A6, Canada

A R T I C LE I N FO A B S T R A C T

Keywords: The operation of mineral processing units or plants is related to the mineral composition of the ore. However,
Mineral unit performances are usually characterized in terms of metal content or recovery as this data is easier to be
Material balance obtained rather than the mineral content. This paper presents a data reconciliation method that combines
Data reconciliation material balancing calculations and mineral stoichiometric information to estimate balanced mineral compo-
Ore composition
sition, chemical assays and flow rates in various streams of a mineral processing plant. The advantage of this
method is evaluated by comparing the variance of the reconciled variables from this method to those obtained
from usual data reconciliation methods. The estimated mineral composition leads to improved process perfor-
mance evaluation.

1. Introduction
The ore that is feeding a concentrator is made of minerals that
condition the energy requirements for the ore comminution and the
performance of the subsequent separation circuit. Indeed, a change in
the mineral composition of the ore may affect grinding mill throughput
as well as the recovery or grade of the economical products. Despite this
recognized relationships, the mineral composition of the ore is seldom
measured or estimated to provide regular information for plant per-
formance evaluation. Some methods such as QEM-SCAN are available
to measure the concentration of the minerals in the ore samples
(Lamberg, 2007; Whiten, 2008), but such an approach is too time de-
manding and expensive to be applied on a daily basis for plant mon-
itoring. Therefore, plant operators usually evaluate the plant perfor-
mance using chemical assays of metals or elements that are readily
measured from ore samples.
Samples of the streams of mineral processing plants are collected on
a daily basis to be assayed for the strategic elements. The measurements
are reconciled to produce coherent data based on the material con-
servation concept (Hodouin et al., 2010). Then the reconciled data is
applied to characterize the day to day operation of the plant in terms of
grade and recovery of the valuable metals. However, such diagnosis is
only partial as elements can be transported by different minerals that
respond differently to the concentration process. A possible way to
detect a problem resulting from the change in the mineral composition
of the ore requires determining the concentrations of the various mi-
nerals in the ore and in the produced salable concentrates. Since
measuring the mineral content of the ore could be a demanding task,
one should consider estimating this data from the daily chemical assays
of the samples collected on the main streams of the circuit. Some au-
thors have already investigated the problem. Whiten (Whiten, 2008)
described a method to estimate the minerals content of an ore sample.
Lamberg and Vianna (2006) proposed a method to do a sequential
mineral reconstruction followed by a material balance. Few authors
take account for the mass conservation of the reconstructed minerals
content (Subramanian et al., 2016). These authors consider the overall
problem of material balance reconciliation and mineral content re-
construction and apply their method to batch and semi-batch processes.
They also suggest a method to simultaneously carry out the mineral
composition estimation and to balance all the species content so that
mass conservation is verified for the overall processing circuit.
The objectives of this study are: (1) to propose a more efficient data
reconciliation calculation algorithm for large size data set based on a
hierarchical structure allowing an analytical solution for the ore mi-
neral and/or chemical composition; (2) to apply the method to a si-
mulated complex sulfide Pb-Zn-Cu-Ag ore flotation plant; (3) to com-
pare the reconciled process variables and performance indices
reliability of the proposed method to the conventional ones that do not
make use of the mineral stoichiometric constraints.
The paper consists of four sections. The first section presents the
formulation of the data reconciliation and mineral reconstruction pro-
blem involving flow rates, chemical elements and mineral composition
measurements. The second section proposes a method that simulta-
neously estimates the mineral composition from the chemical assays

⁎
Corresponding author.
E-mail address: maryam.sadeghi.1@ulaval.ca (M. Sadeghi).

https://doi.org/10.1016/j.mineng.2018.04.023
Received 11 December 2017; Received in revised form 18 April 2018; Accepted 21 April 2018
Available online 13 May 2018
0892-6875/ © 2018 Elsevier Ltd. All rights reserved.
M. Sadeghi et al.
Nomenclature
Variables and functions
a Number of measured elements (1 ≤ a ≤ n )
argming (z ) Value of z which minimizes the function of g(z)
z m
A, B, C Intermediate calculation matrices
F Vector of solids flow rates Fj
f Relative flow rates vector
f Diagonal matrix of relative flow rates
F Block diagonal matrix of the relative flow rates
f∗ Vector of independent relative solids flow rates
E (n × 1) vector of e values required to express the stoi-
chiometry when using Y −
I Number of available information pieces or identity matrix
depending on the context
J() Maximum likelihood criterion
ℓ Number of plant streams
L Lagrange associated to a constrained LQ problem
m Number of selected ore minerals
M Network matrix
M Block diagonal matrix of M repeated n times or of different
Mk matrices
n Number of selected ore chemical elements
p Number of nodes in the plant
q Number of measured mineral mass fractions (0 ≤ q ≤ m )
V Variance-covariance matrix
x Generic notation for any element mass fraction xij
X Vector of x values
y, y − Generic notation for any mineral mass fraction ykj
Y Vector of y values
Y− Vector of m-1 mineral mass fractions
Z Total number of variables to be estimated
Indices
x Index for a matrix related to chemical elements
y Index for a matrix related to mineral phases
and balances this data with the estimated flow rates in the streams of a
circuit. The third section describes the simulated mineral processing
plant (a complex sulfides flotation plant) and the related data used to
test the method. The fourth section discusses the data reconciliation
results and compares the reliability of the variable estimates and metal
recoveries to the reliability of those obtained through two other sub-
optimal methods.
2. Formulation of the data reconciliation problem
A data reconciliation problem of a single process or a plant is
usually defined as a statistical procedure that allows the estimation of
measured and unmeasured process variables submitted to physico-
chemical constraints. Usually the main considered constraints are the
laws of mass and energy conservation. In the most frequent cases, one
considers the total mass conservation of one phase and of the chemical
species it is composed of, thus leading to a set of bilinear constraints
(Hodouin and Everell, 1980). However the problem might also consider
various phases, and various levels of the phase properties, i.e. tem-
perature, chemical or physical properties, size-wise chemical compo-
sition, etc…, therefore leading to non-bilinear conservation constraints
(Hodouin and Vaz Coelho, 1987; Bazin et al., 2003; Bellec et al., 2007).
It is often considered that the measurement errors follow Gaussian
distribution and the constraints are deterministic. The method applied
in all such cases, where random uncertainties are present, is based on
118
Minerals Engineering 123 (2018) 117–127
f Index for a matrix related to flow rates
i Index for chemical elements
j Index for plant streams
k Index for minerals
r Lower index indicating either the removed mineral in the
conservation constraint or the reference flow rate
Upper index indicating measurement values of any mea-
sured variable
^
Accentuation for an estimated or reconciled value of any
variable
– Upper index for any mineral content in a set of m-1 mi-
nerals
Greek symbols
α, β Matrices used for expressing the network flow rates from
the independent ones
λ Lagrange multiplier
φ Matrix containing the mass fraction coefficients corre-
sponding to the chemical stoichiometry of the minerals
ϕ (nℓ × mℓ) block matrix that is obtained through the re-
petition of φ
θ Matrix pointing at the measured variables
ψ Modified φ matrix for completeness constraint for mi-
nerals
Ψ Block matrix of ψ
Acronyms
DR1, DR2, DR3 Data reconciliation procedure using respectively
raw data, balanced data and including mineral constraints
LQ Optimization of a quadratic function under linear con-
straints
ML Maximum Likelihood estimation method
NLP Non-linear programming for optimization of a function
under non-linear constraints
RSTDV Relative standard deviation
the maximum likelihood (ML) principle. In the present study the pro-
blem is limited to the following assumptions:
• The plant is assumed to operate in permanent regime (steady-state
reconciliation).
• The plant is a p nodes and ℓ streams (index j = 1 to ℓ) network.
• The flowing material is the ore solid phase characterized by its F
flow rate vector of the F ’s.
j
• Two ore properties are considered: the chemical composition x ij
(mass fraction of element or component i on stream j, i = 1 to n) and
the mineral composition ykj (mass fraction of mineral k on stream j,
k = 1 to m).
• The mineral compositions are known through a stoichiometric ma-
trix φ (n × m) as detailed in Section 2.1.
• The measurements are unbiased and their random errors follow
Gaussian distribution without error correlation between the three
measurement levels (flow rates, chemical elements and mineral
mass fractions), but allowing error correlation within each of the
three levels.
• All the selected process variables are estimable.
• The variable estimation problem is redundant, a necessary condition
being that there are more measurements and non-redundant con-
straints than process variables to be estimated.
• The addition of the mineral stoichiometric compositions and pos-
sibly measured mass fractions add redundancy to the overall
M. Sadeghi et al.
variable estimation process.
2.1. Data reconciliation criterion and constraints
The ML estimation scheme is applied using the following three sets
of quadratic criterion for the three types of variables Fj, xij, ykj, gathered
in the vectors F, X and Y for the whole plant network:
J (F ) = (F m−θf F )T V f−1 (F m−θf F )
J (X ) = (X m −θx X )T V x−1 (X m −θx X )
J (Y ) = (Y m−θy Y )T V y−1 (Y m−θy Y )
where θ is the matrix that selects the measured variables among the
vectors F, X and Y. The weighting matrices correspond to the inverse of
the error variance-covariance matrices (V), therefore giving more
weight to accurate information, and taking account of the correlation
that might exist between some measurement errors, either because of
the use of the same measurement devices or of normalization techni-
ques brought to the raw measurements (Hodouin and Everell, 1980;
Bazin and Hodouin, 2001). The estimates (variables with a ^ symbol)
are then obtained through the following ML solution of the minimiza-
tion problem:
(F ,̂ X ,̂ Y )̂ = arg min (J (F ) + J (X ) + J (Y ))
F ,X ,Y
submitted to the following mass conservation constraints:
(a) for the ore:
Mf = 0
where f is the relative flow rate F/Fr vector. The reference stream flow
rate (Fr) is usually the plant circuit feed rate, and M is the usual (p×ℓ)
network matrix.
(b) for the chemical elements:
M f Xi = 0 for i = 1 to n
or globally:MFX = 0
where f is the (n × n) diagonal matrix of relative flow rates f, Xi is the
vector containing the chemical assays of element i in all streams, F is
the (nℓ × nℓ) block diagonal matrix of the relative flow rates vector f,
and M is the block diagonal matrix of M repeated n times.
(c) for the mineral phases:
M f Yk = 0 for k = 1 to m
or globally:MFY = 0
where the Yk vector contains the k mineral mass fractions in all bran-
ches. F with the dimension of (mℓ×mℓ) and M is the block diagonal
matrix of M repeated m times. It may happen that some mineral ele-
ments k are decomposed in some process units; in this case the corre-
sponding rows in the Mk block have to be deleted.
When either the chemical elements or the mineral phases compo-
sition is exhaustive in the plant streams, the following additional con-
straints must be considered:
n crease the number of x and y variables to be analytically calculated at
∑ x ij = 1 for j = 1 to ℓ
i=1
m tailed below.
∑ ykj = 1 for j = 1 to ℓ
k=1
When Eq. (8) and/or Eq. (9) are enforced, the constraints of Eqs. (6) and
(7) might be written only for n-1 and/or m-1 elements and minerals,
respectively, since the mass conservation equations of the other
Minerals Engineering 123 (2018) 117–127
elements and/or minerals coupled to Eq. (8) and/or Eq. (9) warrant
that the eliminated element and/or mineral are also conserved.
When a data reconciliation problem involves simultaneously che-
mical elements and mineral phase variables, the known stoichiometry
of the mineral imposes additional constraints which may create addi-
tional redundancy to the overall estimation problem. For any stream j,
the element mass fractions vector can be deduced from the mineral
mass fractions by the following relationship:
(1)
Xj = φYj for j = 1 to ℓ (10a)
(2)
or globally by:X = ϕY (10b)
(3)
where φ is the (n × m) matrix containing the mass fraction coefficients
corresponding to the chemical stoichiometry of the minerals, and ϕ the
(nℓ×mℓ) block matrix obtained through the repetition of φ for all the
plant streams. There are two possible consequences to the additional
constraints created by Eq. (10):
(1) The X variable can be estimated by data reconciliation using the
measured flow rate and chemical assays; and Eq. (10), added to
possible measurements of the ore mineral contents, allows the
calculation of a unique value or multiple values of y from the x
estimates. The mineral reconstruction problem is therefore in-
dependent from the reconciliation problem. The constraint φ does
(4) not bring any new information to the estimation of the flow rates
and chemical elements concentrations.
(2) The information in Eq. (10) and x estimates, possibly coupled to y
variable measurements if any, does not allow finding an estimate
for Y. This means that this additional information is in conflict with
(5) the element mass fractions and flow rates. This also implies that it
contains information which is redundant and could benefit to the
estimation of all the plant variables through the overall re-
conciliation procedure. This is the assumption which is made in the
following reconciliation method and which is present in the case
study used to illustrate the method. For this specific data re-
conciliation (DR) case, the following process variables status clas-
(6a) sification must be performed: (a) measured variables, either re-
dundant or not redundant; (b) unmeasured variables, either
(6b) estimable or not estimable, (c) unmeasured estimable variables,
either redundant or not redundant. The present study does not in-
tend to develop general methods and rules for unambiguously de-
fining the status of each process variable in the multiple levels of a
DR procedure; however, this problem has already been studied for
usual 2-level cases (Hodouin et al., 1991).
2.2. Solution to the data reconciliation problem
(7a)
(7b) In order to solve Eq. (4) submitted to equality constraints of Eqs.
(5)–(10), the hierarchical optimization method already used in the
BILMAT algorithm (Hodouin and Everell, 1980) is also used here for
benefiting of the constraints linear structure with respect to the vari-
ables x and y, when flow rate values are selected. As the ML criterion
part that contains x and y is quadratic, the minimization of J (X ) + J (Y )
is a linear quadratic (LQ) problem which has an analytical solution for
selected flow rate values. Therefore the minimum value of the total J of
Eq. (4) can be searched through iterations on a limited number of flow
rate variables. In addition, it is possible to use the constraints to de-
each iteration of the minimization algorithm. The principles of this
(8)
hierarchical minimization strategy are depicted in Fig. 1 and are de-
The selection of the strategy to reduce the size of X and Y vectors to
(9)
be analytically calculated depends on the structure of the problem, i.e.
the relative patterns of the selected elements and minerals for the ore
characterization. In the followings, it is assumed that the mineral de-
scription is exhaustive; therefore the X vector contains only elements
119
M. Sadeghi et al. Minerals Engineering 123 (2018) 117–127

Fig. 1. Principles of the data reconciliation procedure to estimate solids flow rates and chemical and minerals composition.

present in the minerals. of Eq. (14), and λ is the Lagrange multiplier vector.
The y variables are selected as the variables to be analytically cal- After some simplifications the Lagrange is rewritten as:
culated for a given set of flow rate. More precisely, only ℓ(m-1) y T
variables are selected, noted as the vector Y −, and the removed mineral L = (X m )T V x−1 X m + E T θxT V x−1 θx E −2(X m )T V x−1 θx E + Y − A Y−−2BY −
mass fractions (yrj) are given by: + 2λT CY − (16)

yrj = 1− ∑ ykj for j = 1 to ℓ where:

k≠r (11)
A = ΨT θxT V x−1 θx Ψ (17)
Eq. (10a) can be reformulated using y− as:
B = ((X m )T −E T θxT ) V x−1 θx Ψ (18)
Xj = ψY −j + ej for j = 1 to ℓ (12a)
C = MF (19)
or globally as:X = ΨY − + E (12b)
The data reconciliation problem is solved by zeroing the Lagrange
An example of the ϕ matrix modification to provide Ψ and E ma- equation derivatives with respect to Y − and λ:
trices is illustrated in Section 4. Replacing x by its above expression, the
∂L
y− estimation problem consists in the following minimization problem, = 2CY − = 0
∂λ (20)
assuming, for equation writing simplification, that J (Y ) = 0 , i.e. that no
y has been measured (this does not restrict the generality of the method ∂L
as shown in Appendix A1): = 2(AY −−BT + C T λ ) = 0
∂Y − (21)
Y ̂ = argmin ([X m −θx (ΨY − + E )]T V x−1 [X m −θx (ΨY − + E )])
−
leading to:
Y− (13)
λ = (CA−1C T )−1CA−1BT (22)
submitted to the m-1 mineral mass conservation constraints:
Y ̂ = A−1 (I −C T (CA−1C T )−1CA−1)ΨT θxT V x−1 (X m −θx E )
−
MFY − = 0 (14) (23)
where θx (a × n) is the observation matrix for the element mass frac- This solution is valid for the assumptions made on the measurement
tions (a is the number of measured x, 1 ≤ a ≤ n ). The upper index m and constraints structure, but for other redundancy patterns of the data
stands for the measurement values and M and F have the appropriate reconciliation problem, another structure of the search variables and
dimensions. constraints might be required for solving the minimization problem. At
The x ̂ estimates are then easily calculable from y −̂ using Eq. (12), as the upper level of the ML criterion minimization Fig. 1, the search
well as yrĵ from Y ̂ using Eq. (11) (see Fig. 1. The conservation of the y−
−
variables are a set of independent relative mass flow rates f∗ (vector of
minerals warrants the conservation of the variables yr and x because of size ℓ-p-1). Calculations are initiated with guesses for these values. The
the inclusion of the structure of yr and X in the ML criterion of Eq. (13). complete f vector is calculated, using the ore conservation constraint of
It should be noted that if some of the minerals on some of the streams Eq. (5), as:
have been measured, thus improving the information redundancy of the
f = αf ∗ + β (24)
estimation problem, the additional term J (Y −) as proposed in Eqs. ((3)
and (4)) could be easily added to Eq. (13) (see Appendix A1). where the ((ℓ-1)×(ℓ-p-1)) α matrix contains the equations returning the
The estimation problem of Eqs. (13) and (14) can be solved by the estimated value of all flow rates according to the network matrix
Lagrange method with: (Hodouin and Everell, 1980) and the (ℓ-1) β vector contains 0 and 1
values according to the relationship between the reference and other
L = [X m −θx (ΨY − + E )]T V x−1 [X m −θx (ΨY − + E )] + 2λT (CY −) (15)
flow rates.
where C is the matrix containing only the mass conservation constraints Then the optimal value Fr ̂ which minimizes J(F ) is calculated as:

120
M. Sadeghi et al. Minerals Engineering 123 (2018) 117–127

Fr ̂ = argmin (F m−θf Fr f )T V f−1 (F m−θf Fr f ) Table 1

Fr (25) Mineral composition of the BMS ore (Paktunc, 1998).

where f is related to f∗ through Eq. (24). The solution of Eq. (25) is: Minerals Composition Nomin. Min. Max.
(%) (%) (%)
F mT
V f−1 θf f
Fr ̂ = T T −1
Pyrite FeS2 45 35 55
f θf V f θf f (26) Marmatite (low iron) Zn0.97Fe0.03S 8 4 12.5
Marmatite (high iron) Zn0.91Fe0.09S 5.5 2.2 8.5
leading to the total flow rates vector F ̂ at each iteration by multiplying Galena PbS 4.2 2.2 6.3
Eq. (24) by Fr ̂ and to the J(Fr) value of Eq. (25). Chalcopyrite CuFeS2 1 0.5 1.7
Chlorite Al4.7Fe6.2Mg3.4Si6O36·8H2O 8 3 13
In parallel the J(X) part of the criterion is calculated from the
Quartz SiO2 24 15 33
iterative solution Y ̂ (see Fig. 1. The total value of the criterion J(X) + J
−

(F) is then submitted to the convergence test of the NLP (non linear
programming) supervisory algorithm which selects a new set of f∗ va-
lues for a new iteration of the calculation of y and Fr. If the minimum of
J is obtained, the overall solution F ,̂ X ,̂ Y is
̂ provided.
The proposed method does not require strong programming skills
and can be implemented in spreadsheet softwares such as Excel-
MicrosoftTM, which provides the Solver macro for the non-linear opti-
mization search of the relative independent flow rate variables.
3. Studied flotation plant
The method is illustrated with simulated data from a complex sul-
fide flotation circuit. The circuit is described with a brief presentation
of the simulator used to generate data to test the data reconciliation
method.
3.1. Studied flotation circuit
scavenger cells. Also a flash flotation is operated in the grinding circuit
to recover fine galena. The zinc cleaner consists of four cleaning flo-
tation banks, while the copper flotation circuit uses a rougher bank
followed by two banks of cleaner cells. The circuit shown in Fig. 2 is
used to identify the streams that are sampled to obtain the daily pro-
duction data.
The ore processing follows the conventional route of grinding and
flotation. The ground ore is processed in a Cu-Pb flotation circuit that
produces a bulk concentrate of copper and lead minerals. The reject of
the Cu-Pb flotation circuit is processed in a zinc circuit to recover the
zinc mineral. The galena recovered in the bulk Cu-Pb concentrate is
depressed allowing the flotation of a copper concentrate in the copper
separation circuit. The copper flotation reject is conditioned to activate
pyrite that is floated in a reverse circuit leaving zinc and lead minerals
in the reject that is processed to obtain a Pb and a Pb-Zn concentrate in
the bulk Pb-Zn flotation circuits.

The considered flotation circuit has been inspired from the

Brunswick Mining complex sulfide flotation circuit (Orford, et al., 2005;
Roberts et al., 2008) that ceased operation in 2013. The ore consists of
about 80% sulfide minerals with quartz and chlorite as the non-sulfide
gangue (Petruk and Schnarr, 1981). The main sulfide gangue mineral is
pyrite while marmatite (Zn1−x(Fex)S) carries the zinc, galena the lead
metal and chalcopyrite the copper. Silver is distributed in the sulfide
minerals.
A simplified flow sheet of the flotation circuit is shown in Fig. 2. The
circuit is indeed simplified as each unit in Fig. 2 may consist of several
flotation banks. For instance the Cu-Pb circuit consists of a bank of
sixteen rougher cells, two stages of cleaning and a bank of cleaner
3.2. Methodology to generate the data
3.2.1. Flotation simulator
The simulated ore feed is assumed to be composed of the minerals
identified in Table 1. In a real situation, marmatite has a varying pro-
portion of iron and zinc contents. For the simulation, its stoichiometry
is considered as constant, the low and high iron content marmatites
being lumped into a Zn0.95Fe0.05S marmatite.
The flotation circuit is simulated using the six minerals identified in
Table 1 assumed to be perfectly liberated, plus silver considered as a
separate phase. For each of them, separation coefficients at each

Fig. 2. Simplified scheme of Brunswick Mining flotation circuit.

121
M. Sadeghi et al. Minerals Engineering 123 (2018) 117–127

flotation stage were roughly selected. This simulator approximates the

83.0

57.3

58.0
37.0
90.0
97.6
95.0
7.9
5.7
plant operation at the mine closure, the objective being simply to il-
lustrate reconciliation methods.

Recovery

Chlorite
An example of a simulated set of data is shown in Table 2. The feed

Quartz
Pyrite
(%)
contains 0.35 %Cu, 3.64 %Pb and 8.65 %Zn. The ore silver content is

Ag
Cu
Zn

Zn
Pb

Pb
34 g/t. The Cu-Pb concentrate (stream 2) assays 29.8% Pb and 2.68
%Cu. The zinc content of the zinc concentrate (stream 6) is 51.8% Zn, a

Quartz

28.30

30.32

37.74

11.16
35.74

10.27

13.21

35.17
value that is consistent with the latest plant results (Roberts et al.,

7.62

3.37

0.10

0.53
8.68
0.63

0.66
(%)
2008). The total zinc recovery to the Zn and Pb-Zn concentrates is
88.8%, a value that is also coherent with the last reported plant results

Chlorite
of 86–89% zinc recovery (Roberts et al., 2008). The copper concentrate

10.02
8.00
4.31
8.36
2.32

0.17
7.45
9.83
0.74
4.84
0.88
5.62
0.90

7.07

9.68
(%)
assays 17.07% Cu at a recovery of 57.3% which is also consistent with
reported data (Roberts et al., 2008). The Pb concentrate assays 42.5 %
Pb at a recovery of 58%. The Pb recovery in the Pb-Zn concentrate is

Pyrite

45.00
35.33
45.95
31.91
49.81
17.22
66.94
51.10
40.67
34.53
79.47
25.64
22.98

26.46

51.56
(%)
7.9% for a total Pb recovery of 65.9% which is close to the latest
published results (Roberts et al., 2008).

Marmatite
3.2.2. Measurement strategy simulation

13.50
10.60
13.78
60.00

80.91
10.15

12.06

14.31
50.98
The simulated data of Table 2 is considered here as the ‘real’ steady

1.05

1.74
0.81

0.73

3.08

1.72
(%)
state of the plant, values which are obviously not known in practice and
can only be estimated from measurements or mass balance calculations.

Galena

34.39

38.35
16.26
42.72
21.88

49.10
Since measurements are always stained by unavoidable measurement

4.20

1.25
1.38
1.21
1.02
2.25
1.29
7.92

1.53
(%)
errors, the simulator is programmed to add errors to the simulated
values in order to produce the measurements for the plant operation.
Table 3 shows the relative standard deviations applied to the ‘real’ data

Chalcopyrite
Minerals
to generate the errors added to the measurements. The assays of some

49.30
elements such as Pb and Zn are more reproducible than other ones such

1.00
7.74
0.34
1.01
0.16
0.57
2.05
0.30

1.51
2.02
1.41
2.59

1.05

0.33
(%)
as Fe and SiO2. A set of simulated and measured data for the Cu-Pb
circuit is shown in Table 4. One can verify that the real data obeys the Others
mass conservation law while it is not the case for the measured ones.

2.93
1.58
3.06
0.85
3.67
0.06
2.73
3.60
0.27
1.77
0.32
2.06
0.33

2.59

3.55
(%)
The simulator is used to generate 50 sampling campaigns of the
same ‘real state’ of Table 2, according to the standard deviations of the 30.30

32.41

40.25

13.02
38.20

11.68

14.98

37.59
normal errors of Table 3. The standard deviations calculated a poster-
SiO2

8.69

3.95

0.14

0.71
9.89
0.85

0.88
(%)

iori are close to the selected standard deviations used for the simula-
tion. The 50 tests samples, assumed to be collected from the streams
0.46
0.25
0.48
0.13
0.57
0.01
0.43
0.56
0.04
0.28
0.05
0.32
0.05

0.41

0.55
indicated in Fig. 2, are assayed for Cu, Pb, Zn, Fe, S, Al, Mg, SiO2 and Ag
(%)
Mg

content. Solids mass flow rate measurements are also assumed to be

available as indicated in Fig. 2 in streams 1, 2, 4, 6, 9, 12 and 13. It is
0.69
0.37
0.72
0.20
0.87
0.01
0.65
0.85
0.06
0.42
0.08
0.49
0.08

0.61

0.84
(%)

assumed that there are no available measurements of the mineral phase

Al
Simulated values for mineral and element content in the streams of the flotation plant.

mass fractions.
(g/t)

211

243
206
100
227
136

255
Ag

34

17
28
14
27
30
15

16
4. Results from different data reconciliation methods applied to
the flotation circuit
29.43
29.70
29.40
37.43
27.19
36.29
40.16
28.17
40.30
28.11
45.61
24.66
32.98

22.11

28.45
(%)

The data reconciliation method estimating mineral mass fractions

S

described in Section 2 and named DR3 in the followings is applied to

the data generated by the simulator of the flotation circuit of Fig. 2. The
23.56
20.14
23.89
17.45
25.67
10.56
33.86
26.29
34.14
18.04
37.84
14.13
13.17

14.42

26.48
(%)
Fe

objective is to show how this method works and to compare the esti-
mation reliabilities (for F, X, Y and some metal and mineral recoveries)
38.43

51.83

32.65

obtained through this method with the estimates obtained through two
8.65
6.79
8.83

0.68

6.50
1.11
0.52
7.73
0.47
9.16

1.97

1.10
(%)
Zn

other simpler estimation methods.

4.1. Application of DR3 to the flotation circuit

29.78

33.21
14.08
37.00
18.95

42.52
3.64

1.08
1.20
1.05
0.88
1.95
1.11
6.86

1.32
(%)
Pb

The circuit is composed of eight nodes and fifteen streams. The DR3
Elements

data reconciliation algorithm selects six independent relative flow

17.07
0.35
2.68
0.12
0.35
0.05
0.20
0.71
0.10

0.52
0.70
0.49
0.90

0.36

0.11
(%)
Cu

rates, e.g. streams 3, 4, 6, 9, 14 and 15, as f∗ and minimizes the ML

criterion calculated using F, and X through iteration.
Flow Rate

Nine elemental assays (Cu, Pb, Zn, Fe, S, Ag, Al, Mg, SiO2) are as-
sumed to be available on each stream to reconstruct the mineral com-
396.0

360.7

282.8

305.8

310.9
(t/h)

35.3

77.9

54.9
23.0

30.7

25.6

19.6
4.6

5.1

6.0

position. The silver content in the mineral phases is considered as

negligible and therefore not changing significantly their stoichiometric
compositions. Silver can be processed either as an x variable only or
13 (Pb-Zn)

15 (Rej.)
1 (Feed)

14 (Pb)

added, as in this study, to the six minerals as a supplementary mineral

Stream
Table 2

9 (Cu)
6 (Zn)

phase. This has no impact on the reconciled values, since xAg is ne-
10
11
12
2
3
4
5

7
8

cessarily equal to yAg. The (7⨯1) Y vector is exhaustive. Therefore, Y¯ is

122
M. Sadeghi et al. Minerals Engineering 123 (2018) 117–127

Table 3 4.2. Other estimation methods

Applied relative standard deviations to simulate measurements data.
Stream Flow rate Elements 4.2.1. The basic raw data method (DR1)
DR1 uses raw values of the measured variables, and calculates di-
Cu Pb Zn Fe S Ag Al Mg SiO2 rectly the values of unmeasured variables, such as the flow rates 3, 5, 7,
(%) (%) (%) (%) (%) (%) (%) (%) (%) (%) 8, 10, 11, 14 and 15, the plant recoveries of Cu, Zn and Pb in their
1 (Feed) 10.0 5.8 5.8 5.8 7.1 8.6 6.4 7.1 7.1 5.8 respective concentrates, the recoveries of Pb and Zn in the Pb-Zn con-
2 7.0 5.0 5.0 5.0 6.4 8.1 5.7 6.4 6.4 5.0 centrate, the recoveries of pyrite, chlorite and quartz in the plant
3 6.7 6.7 6.7 7.8 9.2 7.2 7.8 7.8 6.7 tailing, and the recovery of Ag in the Pb concentrate, where Ag recovery
4 10.0 5.0 5.0 5.0 6.4 8.1 5.7 6.4 6.4 5.0 is maximum. The mineral mass fractions y can be directly extracted
5 7.6 7.6 7.6 8.6 9.9 8.1 8.6 8.6 7.6
from the raw values of x as explained in Appendix B.
6 (Zn) 7.0 4.2 4.2 4.2 5.8 7.6 5.0 5.8 5.8 4.2
7 6.7 6.7 6.7 7.8 9.2 7.2 7.8 7.8 6.7
8 7.6 7.6 7.6 8.6 9.9 8.1 8.6 8.6 7.6
9 (Cu) 7.0 4.2 4.2 4.2 5.8 7.6 5.0 5.8 5.8 4.2 4.2.2. The standard BILMAT algorithm (DR2)
10 5.0 5.0 5.0 6.4 8.1 5.7 6.4 6.4 5.0 DR2 minimizes the ML criterion J (F ) + J (X ) submitted to the ele-
11 4.2 4.2 4.2 5.8 7.6 5.0 5.8 5.8 4.2
ments conservation of Eqs. (5) and (6). It generates reconciled flow
12 7.0 5.0 5.0 5.0 6.4 8.1 5.7 6.4 6.4 5.0
13 (Pb-Zn) 7.0 4.2 4.2 4.2 5.8 7.6 5.0 5.8 5.8 4.2 rates and assays that can be used subsequently to calculate recoveries as
14 (Pb) 5.0 5.0 5.0 6.4 8.1 5.7 6.4 6.4 5.0 well as mineral mass fractions using the procedure of Appendix B. In-
15 (Rej.) 7.6 7.6 7.6 8.6 9.9 8.1 8.6 8.6 7.6 equality constraints are added to the BILMAT algorithm to warrant that
all the variables have positive values.
The DR1 algorithm does not allow mass conservation constraints to
selected as a six mineral phase vector, the seventh omitted mineral be obeyed. In the particular flow rate measurement strategy selected
being quartz, whose concentration is calculated using Eq. (11). The here, however, the calculated ore flow rates verify the ore conservation
matrices φ, ψ and E correspond to:

since the measurements are made on an independent set of streams,

thus allowing a unique estimate for all the flow rates. The DR2 algo-
rithm warrants the conservation of the ore and the chemical elements,
but the reconstructed minerals do not obey the constraints of their
conservation along the plant network.
Given a set of either measured x (DR1 algorithm) or reconciled x
(DR2 algorithm), and therefore containing random measurement or
estimation errors, there is no possible unique reconstruction of Y¯ from

Table 4
Real (Sim.) and measured (Meas.) data for the Cu-Pb circuit.
Stream number Tonnage (t/h) Cu (%) Pb (%) Zn (%) Fe (%) S (%) Ag (ppm) Al (%) Mg (%) SiO₂ (%)

Sim. Meas. Sim. Meas. Sim. Meas. Sim. Meas. Sim. Meas. Sim. Meas. Sim. Meas. Sim. Meas. Sim. Meas. Sim. Meas.

1 (feed) 396.0 423.6 0.35 0.36 3.64 3.82 8.65 8.50 23.56 29.33 29.43 27.74 34 35 0.69 0.68 0.46 0.48 30.30 29.21
2 (Cu-Pb con.) 35.3 35.0 2.68 2.65 29.78 28.73 6.79 7.25 20.14 18.94 29.70 31.16 211 190 0.37 0.36 0.25 0.24 8.69 8.19
3 (Zn circuit feed) 360.7 388.6 0.12 0.09 1.08 1.04 8.83 8.23 23.89 23.70 29.40 28.94 17 18 0.72 0.75 0.48 0.46 32.41 34.41

123
M. Sadeghi et al. Minerals Engineering 123 (2018) 117–127

Table 5 Table 6
RSTDEV of reconciled F and x values through DR1, DR2 and DR3 methods RSTDEV of estimated mineral mass fractions through DR1, DR2 and DR3
(average values for fifteen streams). methods (average values for fifteen streams).
Average RSTDEV (%) Minerals Average RSTDEV (%)

Method DR1 DR2 DR3 DR1 DR2 DR3

Flow rate 12.5 11.7 4.7 Chacopyrite 5.6 5.7 3.6

Concentrations Galena 5.2 4.1 3.4
Cu 5.6 5.7 3.6 Marmatite 5.0 6.2 3.6
Pb 5.6 4.2 3.4 Pyrite 4.5 4.1 2.9
Zn 5.5 6.5 3.6 Chlorite 5.0 4.5 3.3
Ag 9.2 5.6 4.3 Quartz 6.1 5.2 3.9
Fe 6.8 4.5 2.2
S 8.5 5.3 1.2 Total 31.5 29.9 20.7
Al 7.1 5.4 3.3
Mg 6.9 5.6 3.3
SiO₂ 5.7 4.9 3.2

Total RSTDEV of (n-1) elements 60.9 47.7 28.2

Table 7
nth element (Others)* 246.6 75 3.3 Average values of estimated recoveries of minerals and elements and their
RSTDEV.
* ∑(variance of (n-1) elements)
Minerals Recovery (%) RSTDEV of Recovery (%)

the X vector since the linear system of Eq. (12) to be solved for Y¯ is Real DR1 DR2 DR3 DR1 DR2 DR3
redundant (Whiten, 2008). As a matter of fact, in the present example value
there are nine known x values and only six unknown y−variables.
Marmatite in stream 83.0 83.1 79.3 83.3 16.3 5.6 0.8
However, a least squares method can be used to estimate y¯ or y as
6 (Zn con.)
explained in Appendix A.2. Chalcopyrite in 57.3 56.2 51.5 57.3 14.8 13.9 2.2
stream 9 (Cu
con.)
4.3. Results from the three methods Galena in stream 13 7.9 8.0 7.9 7.9 15.0 8.3 6.5
(Pb-Zn con.)
The comparison between the results of the three estimation methods Marmatite in stream 5.7 5.8 5.4 5.6 14.9 8.3 6.3
provides a way of assessing the contribution of the constraints, gener- 13 (Pb-Zn con.)
Galena in stream 14 58.0 54.9 56.7 57.6 16.1 6.5 2.4
ated by the knowledge about the minerals present in the ore, to the
(Pb con.)
performance of the data reconciliation procedure. The results from the Pyrite in stream 15 90.0 91.4 90.2 90.0 6.8 1.3 0.6
three estimation methods are compared using the relative standard (tailing)
deviation (RSTDEV) of the estimated F ,̂ x ,̂ y ̂ and recovery values. Quartz in stream 15 97.6 96.0 98.9 97.6 9.2 2.0 0.1
(tailing)
Table 5 compares the relative standard deviation of chemical spe-
Chlorite in stream 15 95.0 96.2 96.4 95.1 9.3 1.6 0.3
cies mass fractions and flow rates estimated through the three re- (tailing)
conciliation methods. The exhibited relative standard deviations are the
Elements
average values obtained from the fifteen plant streams. The line Zn in stream 6 83.0 84.0 83.8 83.3 15.9 1.9 0.8
“Others” corresponds to the complement to 1 of the sum of the assayed Cu in stream 9 57.3 56.2 58.8 57.3 14.6 4.5 2.2
element concentrations, and is physically composed of O2 and H2O. As Pb in stream 13 7.9 8.0 8.1 7.9 15.2 16.5 6.5
expected the DR3 method gives the best performance when comparing Zn in stream 13 5.7 5.8 5.6 5.6 15.6 16.4 6.3
Pb in stream 14 58.0 55.1 57.0 57.6 16.1 4.0 2.4
the sum of the average RSTDEV. For some elements the estimation
Fe in stream 15 88.2 89.5 87.3 88.3 12.5 1.2 0.6
RSTDEV through DR2 is slightly larger than the estimation RSTDEV S in stream 15 75.9 77.2 74.3 75.9 14.6 3.7 0.9
through DR1. This is related to the fact that some inequality constraints Silver in stream 14 37.0 34.5 35.8 37.0 20.2 7.0 4.2
have been activated. The superiority of DR3 is also confirmed for the SiO₂ in stream 15 97.4 95.5 97.3 97.5 10.3 0.4 0.1
estimation of “Others” with an estimation RSTDEV of 3.3% compared
to that of DR1 (247%) and DR2 (75%). The large RSTDEV of “Others”
estimated through DR1 is due to the constraint that it is the sum of values used in the simulator. The last three columns give the recovery
other measured elements error variances. standard deviations obtained through the three estimation methods.
Table 6 compares the accuracy of the estimated minerals composi- The results are presented only for the most important streams to avoid
tion by the three methods. Again DR3 gives the best estimation relia- providing unnecessary information. The low RSTDEV of the elements
bility and moreover is the only method which warrants mineral con- and minerals recovery in the presented streams confirms the advantage
servation constraints. Yet again in DR2 method some RSTDEV are of using DR3 over the two other approaches. Since there are some ex-
slightly higher than in DR1 method. This is true for marmatite and clusively related elements and minerals, e.g. marmatite and Zn, chal-
chalcopyrite. This issue comes from the fact that the DR2 algorithm copyrite and Cu and galena and Pb, the recovery values of these mi-
provides negative values for the “Other elements” in some streams that nerals should be equal to those of their exclusive element. This fact is
contain very small amounts of this species. Since the other elements are verified for DR3 but not for the other methods.
estimated using the element completeness constraint of Eq. (8), the sum It is important to underline that the recovery reliabilities are con-
of the x elements would have been larger than one, which is physically siderably improved by the reconciliation procedure. This is related to
impossible. Therefore, inequality constraints were activated in the DR2 the fact that data reconciliation not only decreases the estimate stan-
algorithm to force the program to return positive values for this species. dard deviations of F, x and y, but also creates covariance terms between
Table 7 compares the elements and minerals plant recovery values them. When using performance indices, e.g. recovery, combining sev-
estimated by the three methods. The four first columns give the average eral reconciled data for their calculation, the estimation error correla-
values of the recoveries using variable estimates compared to the real tion induced by the reconciliation procedure drastically increases their

124
M. Sadeghi et al. Minerals Engineering 123 (2018) 117–127

Table 8
Compared degrees of redundancy for DR1, DR2 and DR3 methods(Cons. = conservation, Stoich. = stoichiometric, Norm. = normalization).
Method Z Information: I Redundancy

All var. Meas. F, x Cons. F Cons. x Cons. y Stoich. x = ϕy Norm. y I R (Z – I) %

DR1 255 142 8 0 0 135 15 300 45 18

DR2 255 142 8 72 0 135 15 372 117 46
DR3 255 142 8 72 48 135 15 420 165 65

reliability (Hodouin et al., 1989). been added to the classical two level data reconciliation method. If data
Finally, Table 8 compares the redundancy degree of the three types concerning particle size distribution, either the size-wise mineral mass
of estimation methods. The degree of redundancy is defined as the fractions or size-wise minerals liberation degree, the species transfer
number of available information pieces I, i.e. the number of measure- between a liquid and a solid phase, etc is added to the data re-
ments plus the number of constraints (mass conservation, stoichio- conciliation problem, the structure complexity of the problem would
metric relationships and normalization equations), minus Z, i.e. the significantly increase.
total number of variables to be estimated in the minimization of the In addition, the coupling of data reconciliation with plant optimal
criterion J under the constraints to be verified by the Z variables conception or tuning, plant monitoring and control or real-time opti-
(Hodouin et al., 1991). The degree of redundancy for DR1 method is mization is to be evaluated to determine the advantage of using ad-
only 45 since mass conservation equations of the chemical species and vanced data filtering methods in terms of process performance im-
the mineral phases are not constrained. In the DR2 method the con- provement measured through metallurgical or economic criteria.
servation constraints of the chemical elements is added leading to a
redundancy of 117. Finally for the DR3 method the redundancy in-
5. Conclusion
creases to 165 because of the addition of the mass conservations of the
6 minerals at the eight plant nodes and the normalization to 1 of the
As mineral processing plants treat minerals rather than elements, it
mass fractions, i.e. 15 additional constraints allowing the calculation of
is of great importance to apply data reconciliation methods that si-
the seventh mineral mass fractions. The average normalized re-
multaneously estimate mass fractions of mineral phases and elements,
dundancy for the 255 process variables are 18%, 46% and 65% re-
and flow rates. Since the measurement methods of minerals content are
spectively for DR1, DR2 and DR3, therefore explaining the improved
not used for regular or daily sample analysis, process operators evaluate
estimation reliability through DR3.
the plant performance using only chemical assays of metals or elements
In general, the degree of redundancy would grow while the ratio n/
that are readily measured. Using only chemical measurements can be
m, i.e. the ratio of the number of assayed elements to the number of
confusing, because some elements are carried by different minerals in
estimated mineral mass fractions, is increasing. In addition, the mea-
the ore that respond differently to a given method of separation. The
surement of some of the mineral mass fractions would also obviously
hierarchical data reconciliation method presented in this paper pro-
improve the redundancy degree; however, this is not illustrated in the
vides mineral mass fractions using chemical assays and solids flow rates
investigated case study. For instance, if the mineral mass fractions had
of a mineral processing circuit. The method combines material con-
been measured in the four plant concentrates, the degree of redundancy
servation and completeness concepts to return a set of precise estimated
would have been augmented to 76%. Generally, when the number of
mineral content. It is clearly shown on the studied case study that plant
minerals is larger than or equal to the number of elements and the
performance can be evaluated more accurately using estimated mi-
mineral mass fractions are all unmeasured, the method does not in-
nerals content. The plant performance indicators, e.g. metal and mi-
crease the degree of redundancy.
neral recoveries and concentrate grades, are much more reliable using
The conversion stoichiometric matrix φ plays a dominant role for
this method than usual reconciliation procedures. The feasibility of the
reconstructing the minerals composition. The introduction of un-
method depends strongly on the structure of the available mineral
certainties related to the variability of the mineral stoichiometry or to
phase information. Concepts such as the mineral information un-
the dynamic variability of the pulp flow properties in the data re-
certainties and problem redundancy have to be carefully investigated
conciliation problem might complicate estimation involving minerals.
prior the implementation of the method. It is reasonable to think that
For instance, in the present study the fact that the marmatite stoi-
data reconciliation methods including mineral information would allow
chiometric composition varies with respect to zinc and iron elements in
better maintenance, optimization and control of mineral processing
the minerals between Zn0.97Fe0.03S and Zn0.91Fe0.09S is a source of es-
units.
timation uncertainties which is not addressed in this paper. These
various topics are to be explored in the future.
Applying mineral reconstruction method on the data reconciliation Acknowledgements
procedure improves the estimation of chemical element mass fractions
and flow rates, and the reliability of performance indicators such as The authors acknowledge the financial support from FRQNT,
recoveries of valuable metals. Only the ore mineral information has MITACs, COREM and ArcelorMittal Exploitation Canada.

Appendix A1

Solution to the data reconciliation problem involving measurements conducted on the mineral mass fractions, chemical contents and flow rates

If some (or all) mass fractions of minerals on some of the streams are measured, the redundancy of the estimation problem is increased. In this
case, the additional term J(Y) of Eq. (3) is to be added to the criterion of Eq. (13):

Y ̂ = argmin ([X m −θx (ΨY − + E )]T V x−1 [X m −θx (ΨY − + E )] + [Y m−θy (ΓY − + L)]T V y−1 [Y m−θy (ΓY − + L)])
−

Y− (A1.1)
where θy (q × m) is the observation matrix for mineral mass fractions, therefore returning the measured y variables, and q the number of measured y

125
M. Sadeghi et al. Minerals Engineering 123 (2018) 117–127

variables (0 ≤ q ≤ m ). The term (ΓY − + L) reconstructs the whole Y vector using the reduced vector Y − and the completeness constraint for minerals
(see Eq. (12b)). The estimation problem can be solved by the Lagrange method with:
L = [X m −θx (ΨY − + E )]T V x−1 [X m −θx (ΨY − + E )] + [Y m−θy (ΓY − + L)]T V y−1 [Y m−θy (ΓY − + L)] + λT (CY −) (A1.2)
After some simplifications the Lagrange is rewritten as:
T T T T T
L = X m V x−1 X m + E T θxT V x−1 θx E −2X m V x−1 θx E + Y m V y−1 Y m + LT θyT V y−1 θy L−2Y m V y−1 θy L + Y − AY −−2BY − + 2λT CY − (A1.3)
where:
A = ΨT θxT V x−1 θx Ψ + ΓT θyT V y−1 θy Γ (A1.4)
T T
B = (X m −E T θxT ) V x−1 θx Ψ + (Y m −LT θyT ) V y−1 θy Γ (A1.5)

C = MF (A1.6)
The data reconciliation problem is solved by zeroing the Lagrangian derivatives with respect to Y− and λ :
∂L
= 2CY − = 0
∂λ (A1.7)

∂L
= 2(AY −−BT + C T λ ) = 0
∂Y − (A1.8)
leading to:
λ = (CA−1C T )−1 [CA−1BT ] (A1.9)

Y ̂ = A−1 BT −A−1 C T (CA−1C T )−1CA−1BT

−
(A1.10)
a solution that is formally the same as the one of Eqs. ((22) and (23)), only A and B matrices differ containing the additional information about the
measured mineral compositions.

Appendix A2

Stream by stream mineral composition reconstruction from the available chemical assays

The chemical composition of the ore in any plant stream can be calculated through the stoichiometric matrix φ as shown in Eq. (10a):
Xj = φYj for any stream j = 1 to ℓ

When the mineral composition is exhaustive the chemical composition can be calculated for some or all the elements i (i = 1 to n) present in the
minerals. The inverse problem is more complex. If the selected minerals completely describe the ore, there are m-1 y values that have to be
calculated, since the missing mineral is calculable from the constrained sum of mineral mass fractions. The stoichiometric equation can therefore be
rewritten as:
Xj = ψY −j (A2.1)
The mineral composition can be reconstructed if the number of assayed elements is m-1 and ψ is invertible. In most cases, such as the flotation
example used in the present study, the number of assayed elements is larger than m-1. Here nine elements are assayed for 7–1 = 6 minerals to be
estimated. Then the redundancy degree is 9–6 = 3. As the assays inevitably contain measurement or estimation errors, the system of equations does
not have any solution and should be solved through the ML method.
In the case where the mineral reconstruction is performed on only one stream at a time, without considering the material balance constraints with
the other plant streams, the ML solution is written as:

Y ̂ = argmin [X −ψY −]T V x−1 [X −ψY −]

−

Y− (A2.2)
where the x values are either measured, as in DR1 method, or estimated chemical mass fractions through DR2 method. Even when the x values have
̂ not obey mineral conservation equations.
been previously reconciled and therefore verify the chemical elements conservation, the reconstructed y do
On the contrary, simultaneously performing the data reconciliation on all streams while assuming given values of the flow rates and the mineral
conservation constraints (DR3 method) gives an Y ̂ optimal estimate (Eq. (23)) that obeys mass conservation law. The reconstructed values of y ̂
−

appear in Tables 6 and 7 for the three different methods.

References Supervision of Mineral Processing Plants. Springer, London, pp. 15–83.

Bazin, C., Hodouin, D., 2001. Importance of covariance in mass balancing of particle size
distribution data. Minerals Eng. 14 (8), 851–860.
Bazin, C., Rochon-Tremblay, S., Gosselin, C., 2003. Estimation of gas flow rates and pellet
temperatures in an iron oxide induration furnace. Canadian Metallurgical Quarterly.
42 (3), 301–312.
Bellec, S., Hodouin, D., Bazin, C., Duchesne, C., 2007. Multi-level data reconciliation
application to a gold processing plant. IFAC Proceed. Volume 40 (11), 39–45.
Hodouin, D., 2010. Process observers and data reconciliation using mass and energy
balance equations. In: Sbárbaro, D., del Villar, R. (Eds.), Advanced Control and
Hodouin, D., Everell, M.D., 1980. A hierarchical procedure for adjustment and material
balancing of mineral processes data. Internat. J. Mineral Processing. 7 (2), 91–116.
Hodouin, D., Flament, F., Bazin, C., 1989. Reliability of material balance calculations - a
sensitivity approach. Minerals Eng. 2 (2), 157–169.
Hodouin, D., Vaz Coelho, S., 1987. Mass balance calculations around mineral processing
units using composition analyses within particle-size classes. Internat. J. Mineral
Process.. 21 (1), 65–82.
Hodouin, D., Alliot, N., Flament, F., 1991. Redundancy analysis of complex sets of mi-
neral processing data for mass balance computation. Internat. J. Mineral Process. 32
(3), 213–231.
Lamberg, P., and Vianna, S., (2007). “A technique for tracking multiphase mineral

126
M. Sadeghi et al.
particles in flotation circuits”. in Lima, RMF, Ladeira, ACQ, Da Silva, CA et. al.(eds)
Proceedings of VII Meeting of the Southern Hemisphere on Mineral Technology, Ouro
Preto, Brasil. p. 195-202.
Orford, I., et al. (2005). “The approach to process improvement at Brunswick Mining”. in
Proceedings of the 37th Annual Canadian Minerals Processors Conference, Ottawa,
Ontario, Jan. 18-20. p. 203-226.
Paktunc, A.D., 1998. MODAN: an interactive computer program for estimating mineral
quantities based on bulk composition. Comput. Geosci. 24 (5), 425–431.
Petruk, W., and Schnarr, J.R., (1981). “An evaluation of the recovery of free and un-
liberated mineral grains, metals and trace elements in the concentrator of Brunswick
127
Minerals Engineering 123 (2018) 117–127
Mining”, in CIM Bull. : p. 132-159.
Roberts, J., Deredin, C., and Paulin, J.-G., (2008). “Process improvement update at
Brunswick Mining”. in Proceedings of the 40th Annual Canadian Minerals Processors
Conference, Ottawa, Ontario, Jan. 22-24. p. 347-358.
Subramanian, S., Tammishetti, V., Rai, B., Pradip, B., 2016. Concurrent reconciliation of
chemical and mineral assays for mineral processing circuits. Internat. J. Mineral
Process. 146, 1–9.
Whiten, B., 2008. Calculation of mineral composition from chemical assays. Mineral
Process. Extract. Metall. Rev. 29 (2), 83–97.