applied

sciences
Article
A Method of Ore Blending Based on the Quality of
Beneficiation and Its Application in a Concentrator
Bingyu Liu, Dingsen Zhang * and Xianwen Gao

College of Information Science and Engineering, Northeastern University, Shenyang 110819, China;
[email protected] (B.L.); [email protected] (X.G.)
* Correspondence: [email protected]; Tel.: +86-151-4008-5725

Abstract: Ore blending is an essential part of daily work in the concentrator. Qualified ore dressing
products can make the ore dressing more smoothly. The existing ore blending modeling usually
only considers the quality of ore blending products and ignores the effect of ore blending on ore
dressing. This research proposes an ore blending modeling method based on the quality of the
beneficiation concentrate. The relationship between the properties of ore blending products and the
total concentrate recovery is fitted by the ABC-BP neural network algorithm, taken as the optimization
goal to guarantee the quality of ore dressing products at the source. The ore blending system was
developed and operated stably on the production site. The industrial test and actual production
results have proved the effectiveness and reliability of this method.

Keywords: ABC-BPNN; concentrate recovery; optimization; ore blending







Citation: Liu, B.; Zhang, D.; Gao, X. 1. Introduction

A Method of Ore Blending Based on Ore blending generally refers to the blending of existing ore according to the produc-
the Quality of Beneficiation and Its
tion index before ore dressing. The purpose is to make the mixed ore properties more
Application in a Concentrator.
convenient for production, thus improving production efficiency and reducing production
Appl. Sci. 2021, 11, 5092. https://
cost [1]. Ore blending can also ensure that the feed is uniform during ore processing,
doi.org/10.3390/app11115092
which guarantees normal production in the subsequent links. This is because the parame-
ters such as the grinding time and the beneficiation reagent are all designed. In the process
Academic Editor: María
of ore blending, the model should be established according to the actual situation. Due to
Ángeles Martín-Lara
the diverse nature of raw ore used in different concentrators, the same model cannot be
Received: 29 March 2021
applied in other places.
Accepted: 29 May 2021 Van Tonder et al. [2] proved that good mixed ore properties have a significant impact
Published: 31 May 2021 on flotation production. Some studies have applied the SIP (Stochastic Integer Program-
ming) model to the ore dressing [3–8]. In order to make rational use of resources and
Publisher’s Note: MDPI stays neutral increase the life of mines, operators often need to balance the use of high-grade ore and
with regard to jurisdictional claims in low-grade ore as much as possible, avoiding the occurrence of a decline in the mine’s
published maps and institutional affil- overall quality due to excessive mining of high-grade ore. It requires the operator to make
iations. a reasonable ore allocation based on each mining area’s material composition report before
deciding on the mining plan. Some studies use integer dynamic programming to calculate
ore blending schemes [9,10]. Some scholars use metaheuristics to solve the problem of ore
blending [11–16]. Some scholars research from the perspective of path planning. Based on
Copyright: © 2021 by the authors.
the MCNFP (Minimum Cost Network Flow Problem) model, this paper designed a set of
Licensee MDPI, Basel, Switzerland. scheduling methods that can cover all areas [17]. Dominy, Simon C. et al. [18] proposed that
This article is an open access article geometallurgy is an essential addition to any evaluation project or mining operation. This
distributed under the terms and method can achieve net present value optimization and effective ore body management
conditions of the Creative Commons while minimizing technical and operational risks. Navarra, Alessandro, etc. [19], applied
Attribution (CC BY) license (https:// the stochastic mine planning algorithm to different geometallurgical units and realized
creativecommons.org/licenses/by/ the quantification of the net present value (NPV) of the optional operation modes of the
4.0/).

Appl. Sci. 2021, 11, 5092. https://doi.org/10.3390/app11115092 https://www.mdpi.com/journal/applsci

Appl. Sci. 2021, 11, 5092 2 of 16

concentrator under the conditions of geological uncertainty, and then evaluated the oper-
ation mode of the concentrator. These methods have solved some practical problems in
the industry, but they have some limitations. They assume that the original ore’s nature is
constant at each block or that the change is weak. They believe that each block is processed
independently of others, which is often an unrealistic assumption for mining complexes
that require blending and homogenization before processing. Moreover, for the concentra-
tor, the most critical indicator is the total concentrate recovery rate. The ore blending is
also serving this indicator. Therefore, it is more in line with the manufacturers’ interests to
combine the requirements of ore blending and dressing to produce a blending scheme.
Ore blending is an essential step in the daily production of the concentrator. A good
ore blending scheme can improve grinding production workload, reasonably allocate
the amount of ore into flotation, gravity separation, etc., and improve the concentrator’s
production efficiency and economic efficiency [20]. Therefore, in the process of ore blending,
the material balance and various indicators of the ore must be considered, and attention
must be paid to the impact of the blending product on the quality of the comprehensive
concentrate. If the low-quality content in the mixed ore is too high, the hardness becomes
higher, which will increase the grinding time and consume more grinding balls. When the
content of magnetic substance is insufficient, the effect of magnetic separation will be poor,
which will affect the quality of magnetic separation concentrate. When the high-quality
content is too high, it will consume more imported raw ore and increase the production
cost. When the content is too low, the original pharmaceutical system will not be able to
meet the flotation production needs, and more reagents will be wasted. These conditions
will lead to the unqualified quality of the beneficiation concentrate. Existing methods
cannot meet production requirements. This paper presents a method of ore blending
modeling considering the total concentrate recovery and resource utilization. Section 1
introduces the research background and current situation. Section 2 first analyzes several
factors that should be paid attention to when blending ore. Then, the ABC algorithm is
used to improve the BP neural network algorithm and improve the modeling accuracy.
The new algorithm is used to fit the relationship between the nature of the raw ore and
the concentrate’s recovery rate, which is the first goal of the optimization model. Taking
the maximization of resource utilization as the second goal, a multi-objective optimized
beneficiation model was established. Section 3 uses the NSGA-II algorithm to solve the
optimization model and introduces the software system developed based on the above
method and its application in the production site. The results show the reliability and
practicability of the method. Section 4 summarizes this research and proposes ideas for
further research.

2. Materials and Methods

The raw ore used by a concentrator in this study was supplied by its own open-pit
mining site, which limits the selectivity of beneficiation to a certain extent. The nature of
this raw ore is very complicated. The composition was mainly low-grade ore, hematite,
carbonate iron ore, and silicate iron ore. With the continuous mining of surface ore, the deep
ore is continuously exposed. Due to this mine’s complex material content, the nature of
deep ore and surface ore is very different, resulting in significant changes in the original
ore’s nature during the mining process. Therefore, the existing methods could not be
applied to this concentrator. Raw ore can be roughly divided into three categories. The red
part of the ore body is rich in iron oxide (Fe2 O3 ), which is called “red ore” by the workers
and is a high-grade ore. The green part not far away is rich in ferrous ions, including ferrous
carbonate (FeCO3 ) and ferrous silicate (FeSiO3 ). This part is challenging to beneficiate and
is called “green ore” by the workers. Above them, the black part of the ore is mainly
composed of iron tetroxide (Fe3 O4 ). Workers call it magnetic iron, which is a very high-
quality raw material. These ores are concentrated in one mining site, which shows that the
nature of the ore on site is very complicated. This kind of complex raw ore leads to high
real-time requirements for ore blending.
Appl. Sci. 2021, 11, 5092 3 of 16

The field operator’s ore blending method is to put the property table of each ore
blending block into Microsoft Excel and then mix the ore by a trial method based on
considering product grade. Although this method can initially meet on-site production
needs, its real-time performance is poor and wastes many human resources. Therefore,
this research proposes an intelligent ore blending method, which can improve the real-time
and robustness of ore blending. Before establishing the optimization model, it is necessary
to analyze the constraint conditions of ore blending.

2.1. Constraints on Ore Blending

In actual production, mining and mixing of minerals are restricted by the nature
of mine products and production capacity [21]. Therefore, various indicators must be
comprehensively considered when generating an ore blending plan. There are usually two
classification methods for ore blending indicators, which are divided according to the work
order and the ore’s nature. According to the working order, the ore blending conditions can
be divided into raw ore, ore blending indicators, and ore dressing products. According to
the ore natures, it can be divided into ore’s physical properties, ore’s chemical properties,
and quantity of ore. The first method is adopted in this research.
Raw ore conditions include OG (Ore Grade), HQC (High-Quality Content), LQC (Low-
Quality Content), MPV (Minimum Production Volume), and GR (Geological Reserves).
Ore grade refers to the ratio of metal to total ore in the raw ore. In this study, it is the iron
content in the mixed mineral product, which is a percentage. High-quality content refers to
the proportion of material that is conducive to subsequent work in total ore. Low-quality
content refers to the proportion of material that is detrimental to subsequent work in total
ore. In this study, HQC is Fe2 O3 and Fe3 O4 , LQC is FeCO3 and FeSiO3 . The MPV refers to
the minimum production volume at each ore mining site. The geological reserves refer to
the maximum amount of each ore that can be produced.
Ore blending indicators include the total amount of ore, the maximum and minimum
values of mixed ore grades, the maximum and minimum values of high-quality material
content, and the maximum and minimum values of total inferior ore content. These values
are determined for the quality of the ore blending products, and their size is related to
the use of the mixed ore. Therefore, the ore blending indicators need to be transferred to
the production workshop when the production plan is determined. A useful ore blending
product index can ensure to a certain extent that the quality of grinding products and
beneficiation products meet the production standards, so it is also an essential condition
that needs to be considered when ore blending to generate a plan.
Under normal circumstances, the above are the restrictive conditions for ore blending
production. The operator determines the appropriate ore blending plan according to
the actual situation and stores the produced mixed ore in their respective warehouses,
waiting for distribution. The open-pit targeted by this research is the own property of
a concentrator. The ore mined can only be used as raw material for this concentrator.
Therefore, while ensuring the quality of the blended ore products, it is also necessary to
make the mixed ore suitable for the beneficiation plant’s production. In order to achieve
this goal, this research adds a condition to the ore blending scheme to ensure the quality of
the beneficiation products. It is the difference between this ore blending method and the
conventional ore blending method.
The last condition on ore blending is the quality of the beneficiation products. Pre-
dict the quality of the beneficiation products according to the mixed ore’s nature and guide
the ore blending plan’s adjustment to form a closed-loop control. There are many influ-
encing factors in the production process, from ore blending to beneficiation. Every time
the ore blending scheme is adjusted, it takes several hours to know the adjustment effect
through laboratory tests. It can be said that this is a nonlinear, extensive lag system, so this
research considers this to be a black box system, and a neural network is used to build a
predictive model.
Appl. Sci. 2021, 11, 5092 4 of 16

Concentrate grade and concentrate recovery rate are the key indicators that determine
the quality of beneficiation products. Not only that, they can also evaluate the metallurgical
efficiency of the beneficiation process. However, in order to maintain the stability of
production, the site often requires the concentrate grade to be kept near a certain fixed value,
so the concentrate recovery rate becomes the most concerned indicator of the production
workshop. Even the recovery rate of concentrate directly affects the performance pay
of operators. Therefore, real-time and accurate prediction of the production index of
concentrate recovery rate based on the nature of the mixed ore is the concentrator’s key
task [22]. In this study, BP neural network was used to establish the prediction model,
and the artificial bee colony algorithm was used to optimize the model, which achieved
good results.

2.2. ABC-BPNN Model

ABC (Artificial Bee Colony) algorithm is a relatively novel evolutionary algorithm
proposed by Dervis Karaboga [23]. In the ABC algorithm, the colony of artificial bees
contains three groups of bees: employed bees, onlookers, and scouts [24]. A bee waiting
on the dance area for deciding to choose a food source, is called an onlooker, and a bee
going to the food source visited by itself previously is named an employed bee. A bee
carrying out a random search is called a scout. The food source is the potential solution
to the optimization, and one-to-one corresponds to the employed bees. Suppose that in
the D-dimensional optimization problem, the population number is 2N (both the number
of employed bees and the number of observed bees are N). Food source positions are
randomly generated according to the following equation:

Xi,j = X min
j + δ( X max
j − X min
j ) (1)

where i = 1, 2, . . . , N. Xj max and Xj min are the upper bound and lower bound of the
optimization problem. δ is a random number uniformly distributed between (0, 1).
After identifying the food source, the three bees moved in sequence.
(1) Employed bees: Employed bees at their corresponding food source Xi to generate
candidate solution Vi according to the following formula. If Vi is better than Xi , replace it.

Vi,j = Xi,j + ϕ( Xi,j − Xk,j ) (2)

(2) Onlookers: When employed bee completed the operation, onlookers randomly
selected a food source for extraction based on the following probability.

f iti
Pi = (3)
N
∑ f iti
i =1

where fiti is the fitness function of food source Xi , which is proportional to food source
quality. That is, the greater fitness value of a food source, the higher probability of selection
by onlookers. When the onlooker selects the food source, it is updated according to
Equation (2).
(3) Scouts: In the ABC algorithm, if a position cannot be improved further through a
predetermined number of cycles called limit, then that food source is assumed to be aban-
doned. At this point, the scouts will generate a new food source according to Equation (1)
to replace the discarded food source.
The ABC algorithm proved to be an excellent algorithm for solving optimization
problems. Li used a wavelet neural network (WNN) and a new artificial bee colony
algorithm (ABC) to predict the gold price [25]. Bahram and Nader combined ABC with
BP and RBF neural network to predict phosphate ore grade [26]. In order to improve the
convergence speed of the algorithm, Chen et al. introduced IABC algorithm and differential
evolution (DE) algorithm into the new ABC search equation [27]. The method is applied to
Appl. Sci. 2021, 11, x FOR PEER REVIEW 5 of 16

Appl. Sci. 2021, 11, 5092 RBF neural network to predict phosphate ore grade [26]. In order to improve the conver- 5 of 16

gence speed of the algorithm, Chen et al. introduced IABC algorithm and differential evo-
lution (DE) algorithm into the new ABC search equation [27]. The method is applied to
nonlinear system control
nonlinear system controlproblems.
problems.AnuarAnuar combines
combines ABCABC with
with thethe artificial
artificial neural
neural networknet-
work to avoid the problem that the neural network is prone to fall
to avoid the problem that the neural network is prone to fall into the local optimum and into the local optimum
and applies
applies the the algorithm
algorithm to the
to the classification
classification of of crime
crime data[28].
data [28].Ghanem
Ghanem combined
combined ABC ABC
with PSO as a feed-forward neural network training method
with PSO as a feed-forward neural network training method and tested this method’s and tested this method’s
classification accuracy on
classification accuracy on multiple
multiple data
data sets
sets [29].
[29]. Cui
Cui etet al.
al. proposed
proposed ABC ABC algorithm
algorithm based
based
on distance fitness and verified it with standard data set [30]. Therefore,
on distance fitness and verified it with standard data set [30]. Therefore, ABC algorithm ABC algorithm
was
was selected
selected in in this
this paper
paper to to optimize
optimize thethe parameters
parameters of of BP
BP neural
neural network.
network.
BP
BP (Back Propagation) Neural Networks is a multi-layer feed-forward networknetwork
(Back Propagation) Neural Networks is a multi-layer feed-forward trained
trained
by errorby error inverse
inverse propagationpropagation
algorithm algorithm
proposed proposed by Rumelhart
by Rumelhart and McClelland
and McClelland in 1986.
in 1986.
It is one It
of is
theone
mostof the
widely most widely
used Neuralused Neuralmodels.
network network BPmodels.
networkBP network
can learn andcanstore
learna
and
largestore
number a large number of input-output
of input-output mode mapping mode mapping
relations relations
without without
revealing revealing the
the mathematical
mathematical
equations describingequations suchdescribing
mappingsuch mapping
relations relationsIts
in advance. in learning
advance.rule Its learning
is to userulethe
is to use the gradient descent method to continuously adjust
gradient descent method to continuously adjust the network’s weight and threshold the network’s weight and
threshold
through the through
reversethe reverse propagation
propagation to minimizetotheminimize
network’s thesquared
network’s squared
error. error. The
The topology of
topology
BP neuralofnetwork
BP neural network
model modelinput
includes includes input
layer, layer,
hidden hidden
layer, andlayer,
outputandlayer.
output layer.
Usually,
Usually,
the hidden thelayer
hidden haslayer
two or hasmore
two or more
layers oflayers
neurons.of neurons.
The number The number
of inputof input
layer layer
neuron
neuron
nodes isnodes is the number
the number of model of model
inputs,inputs,
and theandnumber
the number of output
of output layerlayer neurons
neurons is theis
number
the numberof model
of modeloutputs.
outputs.
The combination of ABC algorithm and BP neural network can establish an accurate
model
model and and avoid
avoid falling
falling into
into local optimal.
optimal. Figure
Figure 11 shows
shows thethe flow
flow chart
chart ofof the
the ABC-
ABC-
BPNN algorithm.

Start

Initialization BPNN parameters (Weights and Threshold)

Output the maximum weights and thresholds

Generate food source position (Xi)

Employed bees generate candidate solution (Vi) Calculate the error and propagate it back

Y
Vi is better than Xi
Update weights and thresholds
Replace Xi with Vi
N

Onlookers look for better food source near

Calculated values of MSE
food source Xi in limit times

Y Better food sources

N
were found Satisfy termination
condition
N
Replace Xi
Scouts generate Y
new food sources
Output the predicted value

Record the current optimal solution

End

N Y
Satisfy termination
condition

Figure 1. The flow chart of the ABC-BPNN algorithm.

Taking a concentrator as an example, this research introduces the ore blending model
based on comprehensive concentrate recovery. An ABC-BP neural network algorithm was
used to model the relationship between the recovery rate of concentrate and ore blending
Appl. Sci. 2021, 11, 5092 6 of 16

products’ properties. BP neural network algorithm is chosen because a mature algorithm

should be used in engineering implementation, and this algorithm is suitable. When fitting
by BP neural network algorithm, attention should be paid to the timeliness and rationality
of fitting data. Set a reasonable number of hidden layers and neurons, and verify the
model with the test set after training. If the error does not reach the acceptable range,
training should be performed repeatedly until the error is small enough or the training
number reaches the upper limit.

2.2.1. Data Preprocessing

After collecting and sorting the historical data of ore blending and the corresponding
recorded ore dressing data, it is necessary to screen the data. This paper adopts the
following strategies: s
n
 Pi − P ≥ 3 × 1 ∑ ( Pi − P)2
 
(4)
n i =1
n
∑ Pi
i =1
where P is a data value. P = n . When the difference between a value and an average is
greater than or equal to three standard deviations, this set of data is excluded. The first
80 percent of data was taken as training data, and the last 20 percent as test data.

2.2.2. Describing Function

The comprehensive concentrate recovery is related to ore grade h, high-quality content
k and metal ion content l. The description function of total concentrate recovery can be
expressed as:
J f ( Xi ) (5)
As the ore grade h(•), high-quality content k(•) and metal ion content l(•) are related
to the ore blending results, the above equation can be expressed as:

f [h( Xi ), k( Xi ), l ( Xi )] (6)

where Xi is the amount of ore produced at the mine outlet of No.i.

2.2.3. Normalization
Considering that there are different dimensions of data involved in the fitting, it is
necessary to normalize each group of data before training, and then use a neural network
algorithm for modeling.
In this paper, data is normalized to the range [−1, 1]. The implementation method is:

P0 = 2( P − Pmin )/( Pmax − Pmin ) − 1 (7)

where P is the input data, Pmax is the maximum value of the input data, and Pmin is the
minimum value of the input data.

2.2.4. Fitting and Saving Function

According to the scene’s actual situation, after many times optimization training,
it was finally found that the number of hidden layers was 2, which contained 6 and
8 neurons, respectively, for the model to work best. The neural network using this structure
is shown in Figure 2.
l. Sci. 2021, 11, x FOR PEER REVIEW 7 of 16

According to the scene’s actual situation, after many times optimization training, it
was
Appl. Sci. 2021, 11, 5092 finally found that the number of hidden layers was 2, which contained 6 and 8 neu- 7 of 16
rons, respectively, for the model to work best. The neural network using this structure is
shown in Figure 2.

Output Layer

Input Layer

Hidden Layer

Figure 2. Neural networkFigure 2. Neural

structure chart. network structure chart.

The neural
The neural network networksettings
parameter parameter
are settings
shown inare shown
Table 1. in Table 1.

Table
Table 1. Parameter of neural 1. Parameter of neural network.
network.

Maximum Training Learning Rate Minimum Error

Minimum Number of Input Number
Number
of of Output
Maximum Training Learning Rate Number of Input
2500 0.1 Error
0.0001 3 Output 1
2500 0.1 0.0001 3 1
In order to measure the performance of the proposed method and to evaluate the
In order to measure
prediction the performance
accuracy, of the error
the mean squared proposed
(MSE)method and
criterion to evaluateLet
is employed. the(X , O ); i = 1,
i i
prediction accuracy, the mean squared error (MSE) criterion is employed. Let (X i, Oi); i =
2, . . . , n be the training subset with size n, the MSE is calculated as follows.
1, 2, …, n be the training subset with size n, the MSE is calculated as follows.
1 n
n
1 MSE = ∑2 (yi − di )2 (8)
MSE =
n
(
y − dn i=1
i =1
i i ) (8)
where yi is the actual output of the ith sample, di is the predicted value of the ith sample.
where yi is the actual
Afteroutput of the ith of
the completion sample,
fitting,d40
i is groups
the predicted
of datavalue
were of the as
taken ithtest
sample.
data to verify the
After the completion of fitting, 40 groups of data
fitting results. The error is shown in Table 2: were taken as test data to verify the
fitting results. The error is shown in Table 2:
Table 2. Error of model.
Table 2. Error of model.
Test Error Training Error
Test Error Training Error
0.0788 0.0309
0.0788 0.0309

Figure
Figure 3 shows the 3test
shows the The
results. test results.
blue dotThe blue
is the dot is the
expected expected
output output
and the and
green the green dot
dot
is the
is the predicted predicted output.
output.
Appl. Sci. 2021, 11, x FOR PEER REVIEW 8 of 16
Appl. Sci. 2021, 11, 5092 8 of 16

Figure 3. Text
Figure of of
3. Text neural network.
neural network.

AA comparison
comparison of of
thethe mean
mean andand standard
standard deviation
deviation (SD)
(SD) between
between ABC-BP
ABC-BP and
and the
the
other methods is shown in Table
other methods is shown in Table 3. 3.

Table
Table 3. 3. Comparison
Comparison with
with other
other methods.
methods.
Algorithm
Algorithm Mean
Mean SD SD
ABC-BP
ABC-BP 76.0725
76.0725 0.0309
0.0309
PSO-BP
PSO-BP 77.1132
77.1132 0.0421
0.0421
NA-BP 75.8654 0.0674
NA-BP 75.8654 0.0674

TheThe mean
mean values
values of of
allall three
three algorithms
algorithms areare within
within the
the allowable
allowable range.
range. The
The standard
standard
deviation of ABC-BP model is superior to other algorithms. Therefore, ABC-BP
deviation of ABC-BP model is superior to other algorithms. Therefore, ABC-BP algorithm algorithm
was
was adopted
adopted forfor modeling
modeling inin this
this study.
study.

2.3.
2.3. Optimization
Optimization Model
Model
According
According toto theneural
the neuralnetwork
networkresults
resultsand
andthe
thescene’s
scene’sactual
actualsituation,
situation, the
the optimi-
optimiza-
tion model can be established. Before ore blending, each outlet’s ore properties should
zation model can be established. Before ore blending, each outlet’s ore properties should
be obtained as the ore blending model’s input data. Input the raw ore properties into the
be obtained as the ore blending model’s input data. Input the raw ore properties into the
neural network model trained above to calculate the concentrate recovery rate’s predicted
neural network model trained above to calculate the concentrate recovery rate’s predicted
value. The absolute value of the difference between the predicted value and the target value
value. The absolute value of the difference between the predicted value and the target
is used as the optimization model’s objective function. According to the field experience,
value is used as the optimization model’s objective function. According to the field expe-
determine the ore blending indicators that make the sequence work normally and take this
rience, determine the ore blending indicators that make the sequence work normally and
as the constraint conditions. In addition, the difference between the ore blending plan of
take this as the constraint conditions. In addition, the difference between the ore blending
this stope and other studies is that the unit amount of the mining volume of this stope is
plan of this stope and other studies is that the unit amount of the mining volume of this
not “1 ton”, but “1 vehicle.” It means the full tonnage of a truck. Therefore, the decision
stope is not “1 ton”, but “1 vehicle.” It means the full tonnage of a truck. Therefore, the
variable is changed from mining volume to the number of mining vehicles. Then the
decision variable is changed from mining volume to the number of mining vehicles. Then
optimization problem becomes a pure integer programming (PIP) problem. According to
thethis,
optimization problem
the first objective becomes
function ofathe
pure integer programming
optimization problem can(PIP) problem.Formula
be obtained: According
(12).
to this, In
theorder
first to
objective function of the optimization problem can be obtained:
further ensure the quality of beneficiation products, when optimizingFormula ore
(12).
blending, the improvement of resource utilization must be another goal of ore blending.
In In
theorder
miningto further ensure
production the quality
process, of beneficiation
the resource products,
utilization indexwhen
is an optimizing ore
important index
blending, the improvement of resource utilization must be another goal of ore blending.
In the mining production process, the resource utilization index is an important index
Appl. Sci. 2021, 11, 5092 9 of 16

reflecting the loss of mineral resources mining and the recovery level of valuable metals.
The resource utilization index can be expressed as:

WJ
Zl = 1 − γs
× 100% (9)
Md × θ d × 1− γ p

where Zl is the resource utilization rate; WJ is the amount of concentrate metal; Md is the
geological reserves; θd is the geological grade; γs is the dilution rate; and γ p is the loss rate.
Dilution rate: In the process of ore mining, due to the mixing of waste rock and gangue,
loss of high-grade ore, dissolution or loss of some useful components, etc., the grade of the
mined ore is lower than the geological grade of the ore in the industrial reserves calculated
before mining. This phenomenon is called ore dilution. Mining dilution rate is the ratio
of the difference between the original ore’s geological grade and the grade of the mined
ore to the original ore’s geological grade, and the percentage of industrial reserves lost to
reimbursed industrial reserves. It is one of the important factors to determine the degree of
utilization of mineral resources. The formula for calculating the dilution rate is:

θd − θc
γs = × 100% (10)
θd

where θc is the grade of the mined ore. When the ore dilution rate is high, the selected
ore grade is low, the quality of the concentrate product decreases, and the enterprise’s
economic efficiency decreases.
Loss rate: In the process of mining and production, industrial ores cannot be fully
mined, released, and transported. The loss rate is the degree of industrial ore production
loss expressed as a percentage. The calculation formula for the loss rate is:

Ml
γp = × 100% (11)
Mc

where Ml is the amount of ore lost, and Mc is the amount of ore mined. The amount of ore
mined is the sum of the minerals mined in each mining area in the ore blending plan.
In summary, a multi-objective optimization problem can be obtained.
(1) Performance Index: 

min J f ( Xi ) − ε , i = 1, 2, · · · , m (12)
 

maxZl ( Xi ), i = 1, 2, · · · , m (13)
where ε is the target value of concentrate recovery.
(2) Constraints:
(a) Raw Ore Constraints

J f ( Xi ) = f [h( Xi ), k ( Xi ), l ( Xi )], Xi ∈ Z (14)

m
Mc = ∑ Xi = Wr (15)
i =1

MPVi = ximin ≤ xi ≤ ximax = GRi (16)

(b) Mixed ore properties constraints

hmin ( Xi ) ≤ h( Xi ) ≤ hmax ( Xi ) (17)

kmin ( Xi ) ≤ k( Xi ) ≤ kmax ( Xi ) (18)

cmin ( Xi ) ≤ c( Xi ) ≤ cmax ( Xi ) (19)
Appl. Sci. 2021, 11, 5092 10 of 16

where m is the number of ore producing blocks, Wr is the total amount of

production, h( Xi ) is the ore grade, k( Xi ) is the HQC, and c( Xi ) is the LQC.
After establishing the multi-objective optimization model, this study used the NSGA-II
algorithm to solve it, and through the TOPSIS algorithm, extract the optimal solution from
the Pareto solution set and provide it to the operator as the final ore blending plan. Then, the
operator can adjust the production amount of each block according to the actual situation
and the field’s requirements to make it more suitable for mining operation.

3. Results and Discussion

Based on the above algorithm, the ore blending optimization system was developed.
The server of the system was located in the beneficiation dispatching room, and the upper
computer was located in the ore blending office of another building. The communication
between the two was realized by ethernet.
Based on this concentrator’s actual production situation, the modeling method was
used to generate the ore blending scheme and compare it with the manual ore distribution
scheme. There were four ore producing blocks in total. The ore properties of each ore
producing block are shown in Table 4:

Table 4. The ore properties of each ore producing block.

Ore Grade Iron Content High Quality Low Quality Minimum Production Geological
No
T (%) M (%) Content G (%) Content B (%) Volume (Unit) Reserves (Unit)
1 34.54 20.32 18.24 4.31 0 100
2 29.26 15.48 12.87 4.25 0 100
3 32.03 3.21 0 3.17 0 100
4 31.87 2.50 0 2.56 0 100

The property requirements of ore blending products are as Table 5:

Table 5. The property requirements of ore blending products.

Total Ore Concentrate Upper Grade Lower Grade Upper of High Lower of High Upper of Low Lower of Low
(Unit) Recovery (%) of Ore (%) of Ore (%) Quality (%) Quality (%) Quality (%) Quality (%)
100 73.8 33 29 12 9 4.5 3

A software platform was developed for this working condition and ore blending
method. In order to facilitate the operator to use this system, a human–computer interaction
interface of the system was compiled, shown in Figure 4. This interface can generate and
display the ore blending plan. After clicking the “Parameter Setting” button, it will jump
to the interface shown in Figure 5, where the operator can input relevant data.
Figure 5 shows the interface for setting parameters of the software. Data from Tables 3
and 4 need to be entered into this interface.
In the current production, the effect of ore blending results on the recovery rate
of concentrate has not been considered by operators, who only consider ore blending
products’ properties.
Appl. Sci. 2021, 11, 5092 11 of 16

Based on the above data, the modeling method is used. The optimization model can
be expressed in the following forms:
 
min J f ( Xi ) − 73.8 , i = 4
 

maxZl ( Xi ), i = 4
 m


 ∑ Xi = 100
 i =1 (20)


 29 ≤ h( X ) ≤ 33
i
s.t.

 9 ≤ k( Xi ) ≤ 12
 3 ≤ c( Xi ) ≤ 4.5



0 ≤ xi ≤ 100


where
4
∑ ( Ti × xi )
i =1 34.54 × x1 + 29.26 × x2 + 32.03 × x3 + 31.87 × x4
h ( Xi ) = = (21)
4 x1 + x2 + x3 + x4
∑ xi
i =1

4
∑ ( Gi × xi )
Appl. Sci. 2021, 11, x FOR PEER REVIEW i =1 18.24 × x1 + 12.87 × x2 + 0 × x3 + 0 × x4 11 of 16
k ( Xi ) = = (22)
4 x1 + x2 + x3 + x4
∑ xi
i =1

A software platform 4 was developed for this working condition and ore blending
∑ ( Bi × xi )
method. In order toi=facilitate 4.31 × xto
the operator 1+ use × xsystem,
this
4.25 2 + 3.17a× x3 + 2.56 × x4 interac-
human–computer
l ( Xi ) = 1 = (23)
tion interface of the system
4 was compiled, shown x1 + x2 + x3 + x4 interface can generate
in Figure 4. This
and display the ore blending ∑ xi plan. After clicking the “Parameter Setting” button, it will
i =1
jump to the interface shown in Figure 5, where the operator can input relevant data.

Figure
Figure 4.
4. Main
Main interface
interface of
of the
the ore
ore blending
blending system.
system.

Figure 5 shows the interface for setting parameters of the software. Data from Tables
3 and 4 need to be entered into this interface.
 Figure 4. Main interface of the ore blending system.
Appl. Sci. 2021, 11, 5092 12 of 16
Figure 5 shows the interface for setting parameters of the software. Data from Tables
3 and 4 need to be entered into this interface.

Figure 5. Interface forFigure

blending parameters
5. Interface setting. parameters setting.
for blending

The model was solved by optimization algorithm and the ore selectivity index was
calculated. Compared with the results obtained by the original method. This method
took about 3 min to generate an optimized ore blending scheme on an ordinary computer.
The original method takes at least half an hour to calculate a plan by trial and error.
The comparison results are shown in the Table 6:

Table 6. Ore blending scheme.

Modeling Method Mining No.1 Mining No.2 Mining No.3 Mining No.4 Ore Dressing Index
Original method 16 24 32 28 79.45
This method 27 27 31 15 82.35

It can be seen from Table 6 that the ore dressing index of the ore dressing results
obtained by this modeling method is higher than that of the original method because the
effect of the ore dressing results on the blending is directly considered by this method.
Ore dressing index is a model used to evaluate the quality of mixed mineral products.
After substituting the relevant parameters into the function, the output index represents
the degree to which the mixed ore is suitable for beneficiation. The model is related to the
ore’s composition, the particle size of the ore, and the grinding time. When comparing
methods in this study, it is assumed that the ore size and grinding time are fixed values.
The degree of ore suitable for beneficiation can be found in Table 7.
By looking up the table, it can be found that the ore blending scheme generated by
this research can improve the quality of the mixed ore from “good” to “excellent”, and it
takes less time. In the production process, the ore dressing index can be used to predict the
nature and production cost of the beneficiation products, as shown in Table 8.
Appl. Sci. 2021, 11, 5092 13 of 16

Table 7. Quality evaluation method of mixed ore.

No. Ore Dressing Index Result

1 >80 Excellent
2 70–80 Good
3 60–70 Qualified
4 50–60 Relatively poor
5 40–50 Poor
6 <40 Unavailable

Table 8. The nature and production cost of beneficiation products.

ODI Concentrate Grade Tailing Grade Concentration Ratio Cost (CNY/t)

Excellent 67.5% 7.5–8.5% 2.93–3.03% 383
Good 67.5% 8.5–9.5% 3.03–3.14% 407
Qualified 67.5% 9.5–10.5% 3.14–3.26% 447
Relatively poor 67.5% 10.5–11.5% 3.26–3.39% 511
Poor 67.5% 11.5–13.5% 3.39–3.72% 609
Unavailable 67.5% More than 13.5% More than 3.72% 796

This concentrator produces about 7 million tons of concentrate per year. Using this
solution can reduce the production cost per ton by at least 2 CNY (increase DOI by 1%),
which can save 14 million CNY (2.15 million USD) per year.
The results of industrial trials are encouraging. The operator immediately put the
system into production and applied it for four months, as shown in Tables 9–12.

Table 9. Ore blending scheme in June.

Mining Mining Mining Mining Mining Mining Mining Mixed Ore Grade Profit
Scheme
No.1 No.2 No.3 No.4 No.5 No.6 No.7 (%) (CNY)
Method 1 41 24 15 6 14 2 5 54.30 46.00
Method 2 50 24 6 6 15 2 5 54.08 50.24
Method 3 41 24 15 6 14 2 5 54.56 45.04
Method 4 50 24 6 6 15 2 5 54.36 49.64
This method 50 20 5 5 13 2 4 54.22 58.11

Table 10. Ore blending scheme in July.

Mining Mining Mining Mining Mining Mining Mining Mining Mining Mixed Ore Profit
Scheme
No.1 No.2 No.3 No.4 No.5 No.6 No.7 No.8 No.9 Grade (%) (CNY)
Method 1 41 24 0 15 0 6 14 2 5 54.30 94.12
Method 2 50 24 0 6 0 6 15 2 5 54.08 102.71
Method 3 41 0 24 15 0 6 14 2 5 54.56 93.49
Method 4 50 0 24 6 0 6 15 1 5 54.36 102.35
Method 5 41 24 0 0 15 6 13 1 5 55.28 83.68
Method 6 50 24 0 0 6 6 14 2 5 54.51 98.85
This method 46 23 3 6 3 6 14 2 4 54.18 112.10

Table 11. Ore blending scheme in August.

Mining Mining Mining Mining Mining Mining Mining Mining Mining Mixed Ore Profit
Scheme
No.1 No.2 No.3 No.4 No.5 No.6 No.7 No.8 No.9 Grade (%) (CNY)
Method 1 42 23 0 15 0 6 15 2 5 53.71 54.12
Method 2 42 32 0 6 0 6 15 2 5 54.03 59.83
Method 3 42 18 5 0 15 6 15 2 5 54.00 49.35
Method 4 42 29 3 0 6 6 15 2 5 54.01 57.94
This method 36 26 3 13 2 5 14 2 4 54.15 67.36
Appl. Sci. 2021, 11, 5092 14 of 16

Table 12. Ore blending scheme in September.

Mining Mining Mining Mining Mining Mining Mining Mining Mining Mixed Ore Profit
Scheme
No.1 No.2 No.3 No.4 No.5 No.6 No.7 No.8 No.9 Grade (%) (CNY)
Method 1 45 20 0 0 15 6 14 2 5 54.40 56.41
Method 2 45 0 20 0 15 6 14 2 5 54.63 56.09
Method 3 45 20 0 15 0 6 15 2 5 53.83 69.35
Method 4 45 0 20 15 0 6 15 2 5 54.04 68.92
This method 35 27 1 13 5 6 13 2 4 54.21 80.27

This method’s profit was 7.87 CNY (1.22 USD) per ton, higher than the original
method’s maximum value. The monthly output was 360 kilotons. Under the premise that
other production conditions remain unchanged, this method can increase the profit by
2.83332 million CNY (440,000 USD). Starting from the second month, the number of mining
locations changed from 7 to 9. At this time, the number of calculations increased greatly,
and it became more difficult to generate a ore blending plan with the original method,
whereas this new method could generate an reliable plan.
In the last three months, compared with the original method, the profit created by this
new method is at least 8.36, 7.53, and 10.92 CNY (1.30, 1.17, and 1.70 USD), respectively.
The effect of industrial application shows the effectiveness of this research.

4. Conclusions
The current common iron ore blending methods only refer to the mixed ore indicators
and the original ore’s nature, but ignore the impact of the mixed ore properties on the
beneficiation production. They often regard ore blending and beneficiation as two separate
processes, rather than using the former as an auxiliary link to the latter to ensure the
quality of the beneficiation products. This has led to a disconnect in the whole process
of beneficiation production. This paper proposes a method of ore blending based on the
quality of beneficiation.
This method uses the ABC-BP neural network to fit the blended product’s properties
and the total concentrate recovery rate and establishes a prediction model. The results
show that the improved BP neural network based on the ABC algorithm has a good
application effect. This research also proposes not only to increase the recovery rate of
beneficiation as the goal, but also to focus on long-term development, and to improve the
resource utilization rate of the mine as another goal. In this way, the mining farm and the
concentrator obtained the linkage through the ore grade, which can achieve a win-win
effect. Based on this, a multi-objective optimized ore blending model was established, and a
reasonable ore blending plan could obtained after calculation. Finally, a set of ore blending
scheme generation system was developed, and it operated stably in a concentrator.
Scholars should continue to collect data on mixed ore and beneficiation concentrates
and establish a more comprehensive concentrate quality model in further research. This is
because the accuracy of the objective function in the multi-objective optimization ore
blending model determines the ore blending scheme’s reliability. Moreover, the constraint
conditions for optimizing ore blending should keep pace with the times. Taking this
research as an example, when the ODI calculation rules are changed, the operator must
actively change the constraint conditions for ore blending. In future research, a system
linked to an economic benefit calculation should be developed to further reduce manual in-
tervention.

Author Contributions: Conceptualization, B.L. and D.Z.; methodology, D.Z.; software, D.Z.; valida-
tion, B.L., D.Z. and X.G.; formal analysis, B.L. and D.Z.; investigation, B.L. and D.Z.; resources, X.G.;
data curation, D.Z.; writing—original draft preparation, D.Z.; writing—review and editing, B.L. and
D.Z.; visualization, B.L. and D.Z.; supervision, B.L. and D.Z.; project administration, X.G. All authors
have read and agreed to the published version of the manuscript.
Appl. Sci. 2021, 11, 5092 15 of 16

Funding: This work was supported by the National Natural Science Foundation of China (No. 61573088).
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Nava, J.V.; Llorens, T.; Menéndez-Aguado, J.M. Kinetics of Dry-Batch Grinding in a Laboratory-Scale Ball Mill of Sn–Ta–Nb
Minerals from the Penouta Mine (Spain). Metals 2020, 10, 1687. [CrossRef]
2. Van Tonder, E.; Deglon, D.A.; Napier-Munn, T.J. The effect of ore blends on the mineral processing of platinum ores. Min. Eng
2010, 23, 621–626. [CrossRef]
3. Lamghari, A.; Dimitrakopoulos, R. A diversified Tabu search approach for the open-pit mine production scheduling problem
with metal uncertainty. Eur. J. Oper. Res. 2012, 222, 642–652. [CrossRef]
4. Consuegra, F.R.A.; Dimitrakopoulos, R. Algorithmic approach to pushback design based on stochastic programming: Method,
application and comparisons. Min. Technol. 2010, 119, 88–101. [CrossRef]
5. Ramazan, S.; Dimitrakopoulos, R. Production scheduling with uncertain supply: A new solution to the open pit mining problem.
Optim. Eng. 2013, 14, 361–380. [CrossRef]
6. Benndorf, J.; Dimitrakopoulos, R. Stochastic long-term production scheduling of iron ore deposits: Integrating joint multi-element
geological uncertainty. J. Min. Sci. 2013, 49, 68–81. [CrossRef]
7. Leite, A.; Dimitrakopoulos, R. Stochastic optimisation model for open pit mine planning: Application and risk analysis at copper
deposit. Min. Technol. 2013, 116, 109–118. [CrossRef]
8. Goodfellow, R.C.; Dimitrakopoulos, R. Global optimization of open pit mining complexes with uncertainty. Appl. Soft Comput.
2016, 40, 292–304. [CrossRef]
9. Lambert, W.B.; Newman, A.M. Tailored Lagrangian Relaxation for the open pit block sequencing problem. Ann. Oper. Res. 2014,
222, 419–438. [CrossRef]
10. Boland, N.; Dumitrescu, I.; Froyland, G.; Gleixner, A.M. LP-based disaggregation approaches to solving the open pit mining
production scheduling problem with block processing selectivity. Comput. Oper. Res. 2009, 36, 1064–1089. [CrossRef]
11. Goodfellow, R.; Dimitrakopoulos, R. Algorithmic integration of geological uncertainty in pushback designs for complex multipro-
cess open pit mines. Min. Technol. 2013, 122, 67–77. [CrossRef]
12. Ferland, J.A.; Amaya, J.; Djuimo, M.S. Application of a Particle Swarm Algorithm to the Capacitated Open Pit Mining Problem; Springer:
Berlin/Heidelberg, Germany, 2007; Volume 76, pp. 127–133.
13. Shishvan, M.S.; Sattarvand, J. Long term production planning of open pit mines by ant colony optimization. Eur. J. Oper. Res.
2015, 240, 825–836. [CrossRef]
14. Kumral, M. Optimizing ore-waste discrimination and block sequencing through simulated annealing. Appl. Soft Comput. 2013,
13, 3737–3744. [CrossRef]
15. Khan, A.; Niemann-Delius, C. Production Scheduling of Open Pit Mines Using Particle Swarm Optimization Algorithm.
Adv. Oper. Res. 2014, 2014, 1–9. [CrossRef]
16. Montiel, L.; Dimitrakopoulos, R. Stochastic mine production scheduling with multiple processes: Application at Escondida Norte,
Chile. J. Min. Sci. 2013, 49, 583–597. [CrossRef]
17. Dewil, R.; Vansteenwegen, P.; Cattrysse, D.; Van Oudheusden, D. A minimum cost network flow model for the maximum
covering and patrol routing problem. Eur. J. Oper. Res. 2015, 247, 27–36. [CrossRef]
18. Dominy, S.; Connor, L.O.; Parbhakar-Fox, A.; Glass, H.; Purevgerel, S. Geometallurgy—A Route to More Resilient Mine Operations.
Minerals 2018, 8, 560. [CrossRef]
19. Navarra, A.; Menzies, A.; Jordens, A.; Waters, K. Strategic evaluation of concentrator operational modes under geological
uncertainty. Int. J. Min. Process 2017, 164, 45–55. [CrossRef]
20. Tavares, L.M.; Kallemback, R.D.C. Grindability of binary ore blends in ball mills. Min. Eng. 2013, 41, 115–120. [CrossRef]
21. Birat, J. Society, Materials, and the Environment: The Case of Steel. Metals 2020, 10, 331. [CrossRef]
22. Phull, J.; Egas, J.; Barui, S.; Mukherjee, S.; Chattopadhyay, K. An Application of Decision Tree-Based Twin Support Vector
Machines to Classify Dephosphorization in BOF Steelmaking. Metals 2020, 10, 25. [CrossRef]
23. Karaboga, D.; Basturk, B. A powerful and efficient algorithm for numerical function optimization: Artificial bee colony (ABC)
algorithm. J. Glob. Optim. 2007, 39, 459–471. [CrossRef]
24. Karaboga, D.; Akay, B. A survey: Algorithms simulating bee swarm intelligence. Artif. Intell. Rev. 2009, 31, 61–85. [CrossRef]
25. Li, B. Research on WNN Modeling for Gold Price Forecasting Based on Improved Artificial Bee Colony Algorithm.
Comput. Intel. Neurosci. 2014, 2014, 1–10. [CrossRef]
26. Jafrasteh, B.; Fathianpour, N. A hybrid simultaneous perturbation artificial bee colony and back-propagation algorithm for
training a local linear radial basis neural network on ore grade estimation. Neurocomputing 2017, 235, 217–227. [CrossRef]
27. Chen, C.; Tsai, Y.; Jhang, R. Approximation of the Piecewise Function Using Neural Fuzzy Networks with an Improved Artificial
Bee Colony Algorithm. J. Autom. Control Eng. 2015, 3, 18–21. [CrossRef]
28. Anuar, S.; Selamat, A.; Sallehuddin, R. Hybrid artificial neural network with artificial bee colony algorithm for crime classification.
In Computational Intelligence in Information Systems; Springer: Berlin/Heidelberg, Germany, 2015; pp. 31–40.
Appl. Sci. 2021, 11, 5092 16 of 16

29. Ghanem, W.A.H.M.; Jantan, A. Swarm intelligence and neural network for data classification. In Proceedings of the 2014 IEEE
International Conference on Control System, Computing and Engineering (ICCSCE 2014), Penang, Malaysia, 28–30 November
2014; IEEE: Piscataway, NJ, USA, 2014; pp. 196–201.
30. Cui, L.; Zhang, K.; Li, G.; Wang, X.; Yang, S.; Ming, Z.; Huang, J.Z.; Lu, N. A smart artificial bee colony algorithm with
distance-fitness-based neighbor search and its application. Future Gener. Comput. Syst. 2018, 89, 478–493. [CrossRef]