Preprint, 11th IFAC Symposium on Dynamics and Control of Process Systems,

including Biosystems
June 6-8, 2016. NTNU, Trondheim, Norway

Hydrocyclone cut-size estimation using artificial neural networks

S. van Loggenberg*, G van Schoor*, K.R. Uren* and
A.F. van der Merwe**


* School of Electrical, Electronic and Computer Engineering, North-West University,

Potchefstroom, South Africa (e-mail:[email protected]/[email protected]/[email protected]).
** School of Chemical and Mineral Engineering, North-West University,
Potchefstroom, South Africa (e-mail:[email protected]).

Abstract: The hydrocyclone is widely used throughout the mineral processing industry when working with
slurries. It is either used for classifying, desliming or dewatering. Hydrocyclones are inexpensive,
application-efficient and relatively small to employ. In order to quantify its separation efficiency, models
are utilised to estimate the cut-size and sharpness of classification coefficient, usually in the form of a
partition curve. Most models are based on experimentally obtained data and are therefore not always
universally applicable. Over the last decade researchers have started employing Artificial Neural Networks
(ANNs) in order to obtain a dynamic model. This study endeavoured to use experimentally acquired data
to develop models that predict the cut-size. The models are discussed and evaluated in detail and the best
predicting model was compared to a conventional model from literature.
Keywords: Artificial Neural Network, Modelling, Hydrocyclone, Cut-size, Partition curve, Plitt-Flintoff.
quantitatively depict the separation efficiency of the
1. INTRODUCTION
hydrocyclone. Ideally a hydrocyclone is operated at
Ever since the hydrocyclone became popular in the mineral conditions where a specific cut-size and sharpness of
processing industry, there have been researchers that worked classification is achieved. These parameters cannot however
on developing a model describing its separation efficiency. be monitored in real-time (Frachon & Cilliers, 1999). This
Bradley (1965) published a book in which the known paper thus aimed at developing ANNs, based on
hydrocyclone fundamentals and research of that time were experimentally obtained data, which could predict the cut-size
detailed. He also compiled an extensive list of theoretical and sharpness of classification parameters at certain operating
equations estimating the cut-size and sharpness of conditions. An interesting and relatively new operating
classification, among others. These equations however were variable, called the angle of discharge, is included in some of
not always relevant to industrial hydrocyclones. The next the models.
important contributions made regarding the modelling of
hydrocyclones were by Lynch & Rao (1975). Their research A brief overview of a hydrocyclone is given in Section 2 in
was mainly focused around developing empirical equations for terms of what it is, where it is used, variables associated with
industrial hydrocyclones. In 1976, Plitt published a paper on it and how the performance is described. Section 3 focusses
his mathematical model of a hydrocyclone. His model is one on the ANN models that were developed, discussing their
of the most popular and most extensively referenced models. inputs, architecture and sample division. In order to check the
Plitt developed the model by utilising empirical data along adequacy of the developed ANN models, Analysis of Variance
with hydrocyclone variables that were deemed important in (ANOVA) studies were done and is shown in Section 4.1. To
describing the hydrocyclone’s operation. Flintoff et al. (1987) determine which one of the models performed the best, three
later reviewed the mathematical model and published a revised popular error metrics were utilised. Finally the best ANN
article on it where useful calibration factors were added to the model was compared to Plitt-Flintoff’s mathematical model
mathematical model to further improve estimations. estimations, in order to determine whether the ANN would
perform better than the mathematical model, given in Section
With the advance in computational capability, models now 4.2. The paper is concluded by outlining the work done and
include Computational Fluid Dynamics (CFD), System discussing the most important aspects of the study’s findings.
Identification and expanded statistical correlations. In 1997 H.
Eren et al. were some of the first researchers that employed
Artificial Neural Networks (ANNs) to predict the Particle Size 2. HYDROCYCLONE OVERVIEW
Distributions (PSD) and cut-size of various hydrocyclones (H
Eren, Fung, & Wong, 1997; Halit Eren, Fung, & Wong, 1997). 2.1 A general description

A hydrocyclone’s separation efficiency is mainly described by A hydrocyclone is a static, conical apparatus that is generally
the cut-size (𝑑50 ) and the sharpness of classification (𝑚). used within the mineral processing industry to separate solids
These two parameters are used within a partition curve to from water, better known as slurries. The separation of the

Copyright © 2016 IFAC 996

IFAC DYCOPS-CAB, 2016
June 6-8, 2016. NTNU, Trondheim, Norway

slurries is based on sedimentation, where the necessary swirl Table 1. Hydrocyclone variables
motion is generated by the slurry being fed into the
hydrocyclone by means of a pump (Bradley, 1965). Two Design variables
vortices form within the hydrocyclone referred to as the Hydrocyclone diameter 𝐷𝑐
primary vortex and the secondary vortex as depicted in Fig. 1 Feed inlet diameter 𝐷𝑖
(a). The primary vortex moves downwards and carries the Vortex finder diameter 𝐷𝑜
coarse particles to an opening called the underflow. The Spigot opening diameter 𝐷𝑢
secondary vortex carries the lighter particles, along with most Cone angle 𝜃
of the water, upwards to an opening called the overflow Free vortex height ℎ
(Frachon & Cilliers, 1999).
Operating variables
2.2 Hydrocyclone variables Inlet flow rate 𝑄𝑖
Overflow flow rate 𝑄𝑜
When working with hydrocyclones, two groups of variables Underflow flow rate 𝑄𝑢
are observed. The design variables (shown in Fig. 1 (b)) Pressure 𝑃
include variables that are dependent on the hydrocyclone’s Volumetric solid concentration 𝜙
size and design proportions. The operating variables are
Solid density 𝜌𝑆
independent of the hydrocyclone’s design and solely relate to
the operating conditions of the hydrocyclone. It should Overflow density 𝜌𝑜
however be noted that these two groups of variables cannot be Angle of discharge 𝜔
considered separately because of interactions that occur
between them. Table 1 summarises the main variables
allocated to the two groups.

Tangential inlet
Overflow
Overflow

Feed Cylindrical
Feed
section
Vortex finder

Secondary
vortex

Primary Conical section

vortex

Spigot

Underflow Underflow Fig. 2. A partition curve displaying the cut-size and sharpness
(a) (b) of classification coefficient.
Fig. 1. A graphical representation of a hydrocyclone
2.4 Experimental setup
depicting the (a) vortices and (b) relevant design variables
(adapted) (Frachon & Cilliers, 1999). A refurbished hydrocyclone test-rig, shown in Fig. 3, was
instrumented and used in order to acquire the necessary
2.3 Performance of a hydrocyclone
experimental data. The hydrocyclone’s dimensions are
A hydrocyclone’s performance is mainly described by its tabulated in Table 2. An analog pressure gauge was utilised to
separation efficiency which is directly quantifiable by a measure the inlet pressure and a Doppler flow meter to
partition curve. The partition curve describes the weight (or measure the inlet flow rate.
percentage) fraction of each particle size in the feed that might Table 2. The hydrocyclone dimensions
report to the underflow on the y-axis, to the specific particle
size, on the x-axis. An example of a partition curve is given in Design variable dimensions
Fig. 2 showing how the two efficiency indicating parameters, Hydrocyclone diameter 100 mm
cut-size and sharpness of classification coefficient, are related Feed inlet diameter 33.4 mm
to it. The cut-size, indicated as 𝑑50 , is defined as the size of Overflow diameter 34.0 mm
the particle in the Particle Size Distribution (PSD) that has a Free vortex height 531 mm
50% probability of reporting to either the underflow or the
overflow of the hydrocyclone (Bradley, 1965). The sharpness
of classification coefficient (𝑚) is a parameter that supplies a 3. ARTIFICIAL NEURAL NETWORKS
measure for the gradient of the partition curve. Ideally an 𝑚
> 3 is required in order to obtain sufficiently sharp separation 3.1 Artificial Neural Network overview
(Plitt, 1976).

997
IFAC DYCOPS-CAB, 2016
June 6-8, 2016. NTNU, Trondheim, Norway

To ensure that an accurate and comprehensive model was

developed for this study, three different ANNs were created
Pressure gauge
Flowmeter and evaluated. For each one of the three models the variables
used as inputs were adjusted and the performance of the ANN
evaluated. Model 1 used only the 3 inputs that were set during
Hydrocyclone
the experimental runs. Model 2 had two additional operating
Feed valve
variables, the inlet flow rate and the angle of discharge, in
order to evaluate whether the angle of discharge might
improve the ANN’s performance. The final model, Model 3,
Pump incorporated some of the unusual variables that were relevant
Underflow bin
to this study as suggested by Halit Eren et al. (1997). Table 3
Tank summarises the models’ specifications in terms of their inputs,
outputs, network architecture and sample division. All three
Overflow bin
the models were based on the same Artificial Neural Network
architecture as shown in Fig. 4 varying only the number of
neurons. The initial number of neurons used in training the
ANNs was determined using Klimasauskas’ (1991) general
rule as given by
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑠𝑎𝑚𝑝𝑙𝑒𝑠−5𝜊
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑛𝑒𝑢𝑟𝑜𝑛𝑠 = , (1)
5(𝜄+𝜊+1)
Fig. 3. The schematic of the hydrocyclone test-rig.
where 𝜊 denotes the number of outputs and 𝜄 the number of
Artificial Neural Networks (ANNs) are generally employed to
inputs. If the initial number of neurons did not give
describe input-output relationships of systems. Usually when
satisfactory results, the ANN was retrained using a different
working with ANNs no a priori knowledge is needed of the
number of neurons within the same order (Klimasauskas,
system. ANNs are therefore an ideal modelling method to use
1991). The three models were all initially trained using 2
for the hydrocyclone system because of the complex dynamics
neurons. For Model 2 and Model 3 this was not sufficient and
that are involved. Previous works that incorporated ANNs
it was found that 4 and 3 neurons, respectively, resulted in
found that the models could deliver comparable, and in some
improved ANNs.
cases even better, predictions to the conventional models. The
ANN models proved especially useful when alternative
variables were included such as overflow and underflow flow
rates and overflow density (H Eren, Fung, & Gupta, 1996;
Halit Eren et al., 1997). Some advantages of employing
ANNs, specifically to hydrocyclone modelling, include
incorporating any number of additional variables as inputs,
revisiting the training considerations and structures and its
Fig. 4. The Artificial Neural Network architecture.
utilisation within control applications.
It was found that the sharpness of classification ANNs were all
3.2 Developed Artificial Neural Networks inadequate when evaluating their ANOVA results, delivering
mostly unusable predictions. When examining the literature,
The main goal of this study was to develop an Artificial Neural very little information is available on the modelling of the
Network, using experimentally obtained data, which could sharpness of classification coefficient. This might indicate that
predict the hydrocyclone’s cut-size (𝑑50 ) and sharpness of the 𝑚 cannot be modelled comprehensively using the chosen
classification coefficient (𝑚). In order to obtain the variables, especially when examining the variables Plitt (1976)
experimental data, 41 experiments and their conditions were used to model the sharpness of classification. The results
defined by utilising the Centrally Composite Rotatable Design discussed in this paper will therefore only reflect the cut-size
(CCRD) as investigated by Cilliers et al. (1992). These models and their relevant outcomes.
experimental runs were conducted and the samples were
analysed accordingly, by using a Malvern Mastersizer 2000 4. EVALUATION
particle size analyser. In investigating the conventional
models it became evident that some hydrocyclone variables 4.1 Model adequacy and comparison
are considered more influential than others in estimating the
The first step of the model evaluation was to check whether
separation efficiency. Plitt (1976) deemed the design
the developed models could be deemed adequate or not. To
variables, the inlet flow rate and the solid density as most
achieve that, an Analysis of Variance (ANOVA) was
influential and developed a mathematical model around these
completed for each of the models. The calculated F-value of
variables. When studying Eren & Gupta (1988) the spigot
the models were compared to the appropriate critical F-value
opening diameter is said to affect the performance the most.
at an 𝛼 = 0.05. The 𝛼 = 0.05 implies that a confidence level
Unusual variables, such as overflow and underflow flow rates
of 95% is evaluated and it is expected that about 5% of the
and overflow density were additionally incorporated in the
samples might yield erroneous results. Should the calculated
models developed by Halit Eren et al. (1997).
F-value be larger than the critical F-value, the model was said

998
IFAC DYCOPS-CAB, 2016
June 6-8, 2016. NTNU, Trondheim, Norway

to be significant and thereby considered adequate (Devore & of all three models are satisfactorily higher than the identified
Farnum, 2005). The ANOVA of the models are tabulated in critical F-value and it can thus be concluded that all three
Table 4. When evaluating the table it is seen that the F-values models are adequate.

Table 3. Model details and specifications

Sample division Number
Inputs Output Hidden
Name Training Validation Testing of
layers
# Variables # Variable % Samples % Samples % Samples neurons

Model 1 3 𝑃, 𝜙, 𝐷𝑢 1 𝑑50 60% 25 20% 8 20% 8 1 2

𝑃, 𝜙, 𝐷𝑢 ,
Model 2 5 1 𝑑50 60% 25 20% 8 20% 8 1 4
𝑄𝑖 , 𝜔
𝑃, 𝜙, 𝐷𝑢 , 𝑄𝑖 ,
Model 3 8 1 𝑑50 60% 25 20% 8 20% 8 1 3
𝑄𝑢 , 𝑄𝑜 , 𝜌𝑜 , 𝜔

Table 4. Summary of ANOVA for Model 1, Model 2 and Model 3

Source dfa SSb MSc F test Fd
Model 1
Model 3 89.469 29.823 34.656 𝐹0.05 (3,37) = 2.92 < 34.656
Error 37 31.841 0.861 Significant @ the level 95%
Total 40 121.31 ∴The model is deemed adequate
Model 2
Model 5 83.291 16.658 15.335 𝐹0.05 (5,35) = 2.53 < 15.335
Error 35 38.019 1.086 Significant @ the level 95%
Total 40 121.31 ∴The model is deemed adequate
Model 3
Model 8 102.699 12.837 22.073 𝐹0.05 (8,32) = 2.27 < 22.073
Error 32 18.611 0.582 Significant @ the level 95%
Total 40 121.31 ∴The model is deemed adequate
a df - degrees of freedom; b SS - Sum of Squares; c MS - Mean Square; d 𝐹𝛼 (df Model, df Error).

With all three developed models found adequate, the models 40

r = 0:862
(a) Model 1
40
r = 0:841
(b) Model 2
40
(c) Model 3
r = 0:921
were compared with one another to determine which one R2 = 0:684
2
R = 0:552 R2 = 0:805
38 38 R2 = 0:488 38
would better predict the cut-size. The first measures employed R2 = 0:658 R2 = 0:756
Predicted d50 (7m)

were to calculate the linear correlation coefficient (𝑟), the 36 36 36

coefficient of determination (𝑅2 ) and the adjusted coefficient 34 34 34

of determination (𝑅̅2 ). This was done by plotting the ANN
predicted cut-size, denoted as 𝑦̂𝑖 , against the actual cut-size 𝑦𝑖 . 32 32 32

The process was repeated for all three models and the results 30 30 30
30 32 34 36 38 40 30 32 34 36 38 40 30 32 34 36 38 40
are shown in Fig. 5. The dashed line represents the best fit Actual d50 (7m) Actual d50 (7m) Actual d50 (7m)

linear regression line between the actual and predicted cut- Fig. 5. The actual cut-size versus the predicted cut-size for (a)
sizes. The higher the coefficient-values the stronger the Model 1, (b) Model 2 and (c) Model 3.
relationship is expected to be. By comparing the coefficients
of the three models, one can observe that Model 3 performed The predicted cut-size per sample was then plotted on the same
the best. This was however the first assessment and some graph. The models’ predicted cut-size were expected to at
additional evaluation is needed, as these coefficients should least follow the trend of the actual cut-size. The graphs
never be the only measures examined (Devore & Farnum, obtained for the three models are given in Fig. 6. showing all
2005). the samples as used for training, validation and testing
(unknown samples). Very small differences are observed
The second measure used to evaluate the models was to between the three models but it is noted that the Model 2 shows
visually compare the predicted cut-sizes. The actual cut-size the worst correlation of the three. In order to better view the
was plotted per sample depicting the expected experimental testing samples’ actual and predicted cut-sizes, they were
error, calculated as 2.95%, as error bars.

999
IFAC DYCOPS-CAB, 2016
June 6-8, 2016. NTNU, Trondheim, Norway

separately plotted and are shown in Fig. 7. It was found once and examined. After computing the calibration factor for
more that the correlation of Model 2 was the worst. known samples, (5) was employed to estimate the cut-size for
the same unknown testing samples used to evaluate Model 3.
To further investigate the models’ performance, three popular
error metrics were calculated, assessed and compared. The 42
(a) Model 1 Actual vs Predicted d50
Actual
metrics used in this study was the Mean Square Error (MSE), 40
38
Model 1

d50 (7m)
the Root Mean Square Error (RMSE) and the Mean Absolute 36
34
Error (MAE). When comparing the three models in terms of 32
30
the three metrics, one would like to see the same ranking order 28
10 12 21 25 28 29 31 37
of the models for each one of the metrics. The metric error (b) Model 2 Actual vs Predicted d50
42
equations are given in (2), (3) and (4), respectively. 40 Actual
Model 2
38

d 50 (7m)
36
1 34
𝑀𝑆𝐸 = ∑𝑛𝑖=1(𝑦𝑖 − 𝑦̂𝑖 )2 , (2) 32
𝑛 30
28
1 6 7 12 19 25 26 31 39
𝑅𝑀𝑆𝐸 = √ ∑𝑛𝑖=1(𝑦𝑖 − 𝑦̂𝑖 )2 , (3) (c) Model 3 Actual vs Predicted d50
𝑛 42
Actual
1 40
𝑀𝐴𝐸 = ∑𝑛𝑖=1|𝑦𝑖 − ŷ𝑖 |, (4) 38
Model 3

d50 (7m)
𝑛 36
34
32

𝑦𝑖 denotes the actual cut-size, 𝑦̂𝑖 the predicted cut-size and 𝑛 30

28
the number of observations evaluated (Devore & Farnum, 1 2 11 16 23
Sample number
24 31 36

2005). Fig. 7. The actual and predicted cut-size of unknown testing

The metric assessment results computed for all the samples samples for (a) Model 1, (b) Model 2 and (c) Model 3.
and unknown testing samples are tabulated in Table 5. When Table 5. Summary of error metrics when evaluating all
comparing the metrics all three indicate that the models, and unknown samples for Model 1, Model 2 and Model 3
ranking best performing to worst performing, are Model 3,
Model 1 and Model 2. Model 3 only just surpass Model 1 Error functions
Name
when unknown samples are considered. When reviewing the MSE RMSE MAE
model evaluations, it is seen that Model 3 performed the best All samples
in terms of the coefficient-values, the visual comparison and Model 1 0.7766 0.8812 0.6460
the error metrics. It is thus concluded that Model 3 should be Model 2 0.9273 0.9630 0.7672
used to ensure sufficient predictions of the cut-size. Model 3 0.4539 0.6737 0.5170
Unknown samples
45
(a) Model 1 Actual vs Predicted d50 Model 1 0.5774 0.7599 0.5826
40
Actual Training Validation Testing
Model 2 0.9200 0.9592 0.8166
d50 (7m)

35
Model 3 0.5508 0.7422 0.5757
30

25
The conventional model that was investigated was the
1 2 4 5 6 7 8 9 111315 181922 232427 303233 343539 4041 3 141617 202636 381012 212528 293137
(b) Model 2 Actual vs Predicted d50
mathematical model Plitt and Flintoff revised in 1987.
45
Actual Training Validation Testing Equation (5) shows the mathematical model that was used to
40 calculate the cut-size (𝑑50 ) when the hydrocyclone design
d50 (7m)

35 variables and operating conditions are known. F1 is a

30 calibration factor which is used to improve the cut-size
25
2 3 4 5 9 10 13141516 182022 23242729 303334 35363740 41 1 8 11172128 3238 6 7 121925 263139
estimation (Flintoff et al., 1987). It is usually computed by
(c) Model 3 Actual vs Predicted d50
finding the factor-ratio between the actual measured cut-size
45
Actual Training Validation Testing and the mathematical model calculated cut-size.
40
d 50 (7m)

35 39.7𝐷𝑐0.46 𝐷𝑖0.6 𝐷𝑜1.21 𝜂 0.5 𝑒 0.063𝜙

d50 = F1 𝑘 , (5)
30 0.71 ℎ0.38 𝑄 0.45 [(𝜌𝑠 −1)]
𝐷𝑢
1.6
25
3 4 5 6 8 9 12141517 182125 26272829 303233 34353738 40 7 10 13192022 3941 1 2 111623 243136
Sample number
Fig. 8 shows the actual and predicted cut-sizes for the
Fig. 6. The actual and the predicted cut-size for all samples: unknown samples. It is difficult to distinguish exactly which
(a) Model 1, (b) Model 2 and (c) Model 3. model might perform better as some samples of Model 3 seem
better than the Plitt-Flintoff estimation and vice versa. In order
4.2 The best ANN model versus a conventional model to differentiate which model performs better the same three
error metrics were calculated and compared. The metric
In order to evaluate whether the developed ANN model could results, shown in Table 6, indicate that Model 3 predicts
substitute the conventional model, a comparison was drawn significantly better than the Plitt-Flintoff mathematical model.

1000
IFAC DYCOPS-CAB, 2016
June 6-8, 2016. NTNU, Trondheim, Norway

It is therefore concluded that not only can an ANN model be to see which variable(s) and/or ANN attributes used with
developed and used to predict the cut-size at specific operating Model 2 were responsible for the poor predictions.
conditions, but that an ANN model could substitute the
ACKNOWLEDGEMENTS
conventional mathematical model.
Actual vs Predicted d50 This work is based on the research supported in part by the
38
Actual National Research Foundation (NRF) of South Africa (Grant
Model 3
37 Plitt ! Flinto, Number 91093). Any opinion, finding and conclusion or
36 recommendation expressed in this material is that of the
author(s) and the NRF does not accept any liability in this
d50 (7m)

35
regard.
34

33
REFERENCES
32 Bradley, D. (1965). The Hydrocyclone - Bradley. London:
31
Pergamon Press Ltd.
1 2 11 16 23 24 31 36
Sample number
Cilliers, J. J., Austin, R. C., & Tucker, J. P. (1992). An
Fig. 8. Comparing the actual and predicted cut-size of evaluation of formal experimental design procedures. In
unknown samples employing Model 3 and Plitt-Flintoff’s International Conference on Hydrocyclones (pp. 31–49).
mathematical model.
Table 6. Summary of error functions when evaluating the Devore, J., & Farnum, N. (2005). Applied Statistics for
unknown samples for Model 3 and Plitt-Flintoff model Engineers and Scientists (2nd ed.). Brooks/Cole, Cengage.

Error functions Average Eren, H., Fung, C. C., & Gupta, A. (1996). Application of
Name of artificial neural network in estimation of hydrocyclone
MSE RMSE MAE errors parameters. In Diversity , the key to prosperity : the AusIMM
Model 3 0.5508 0.7422 0.5757 0.6229 1996 Annual Conference (pp. 225–229). Perth, Western
Plitt-Flintoff 0.8292 0.9106 0.7073 0.8157 Australia.

Eren, H., Fung, C. C., & Wong, K. W. (1997). Artificial

5. CONCLUSION Neural Networks in Estimation of Hydrocyclone Parameter
This paper found that it was possible to develop Artificial 50 with Unusual Input Variables. IEEE Transactions on
Neural Networks (ANNs) comprising of different Instrumentation and Measurement, 46(4), 908–912.
hydrocyclone variables as inputs. Using variables that were
deemed important in literature, three different models were Eren, H., Fung, C., & Wong, K. (1997). An application of
developed. All three these models were found to be adequate artificial neural network for prediction of densities and
in predicting the cut-size at various viable operating particle size distributions in mineral processing industry. In
conditions, falling mostly within the acceptable experimental IEEE Instrumentation and measurement Technical
error of 2.95%. When comparing their performance however, Conference (pp. 1118–1121).
Model 2 with the inlet flow rate and the angle of discharge as
additional inputs, seems to perform the worst of the three. This Eren, H., & Gupta, A. (1988). Instrumentation and on-line
could be indicative that the combination of the chosen inputs control of hydrocyclones. In International Conference on
do not deliver favourable predictions. Model 3 with its 8 Control (pp. 301–306). Oxford: IET.
inputs seemed to deliver the most accurate predictions. It
should be noted that it only just performs better than Model 1. Flintoff, B. C., Plitt, L. R., & Turak, A. A. (1987). Cyclone
modelling: a review of present technology. CIM Bulletin,
In wanting to investigate whether Model 3 might be able to 80(905), 39–50.
substitute the conventional mathematical model of Plitt and
Flintoff (1978), the two models’ prediction capabilities were Frachon, M., & Cilliers, J. J. (1999). A general model for
compared by presenting them with unknown samples. The hydrocyclone partition curves. Chemical Engineering
error metrics that were evaluated, implied that Model 3 would Journal, 73(February), 53–59.
be better at predicting cut-size than the Plitt-Flintoff model;
resulting in a 23.6 % smaller error on average. It could Klimasauskas, C. C. (1991). Applying Neural Networks.
therefore be concluded that Model 3 would be able to replace PCAI, 5(3), 20–24.
the conventional mathematical model and that it would deliver
more accurate predictions. Future work could involve Plitt, L. R. (1976). A Mathematical Model of the
investigating whether all the unusual variables are needed to Hydroyclone Classifier. CIM Bulletin Mineral Processing,
improve the ANN. In other words determine if less inputs 69(December), 114–123.
could deliver the same or better predictions. This could
simplify the ANN as well as minimise the data acquisition
effort (measuring less variables). It would also be interesting

1001