ON-LINE MONITORING OF HYDROCYCLONES

BY USE OF IMAGE ANALYSIS

Leeuwner, M.J.*, Aldrich, C., Auret, L., and Bezuidenhoudt, C.

Department of Process Engineering, University of Stellenbosch,

Private Bag X1, Matieland, 7602, Stellenbosch, South Africa,
Tel: +27(21) 808 4728, Fax: +27(21) 808 2059
E-mail: [email protected]

In the past, numerous techniques have been used for hydrocyclone monitoring, with varied
success. Industrial implementation of these techniques, however, is limited as these methods
often are intrusive, economically unfeasible and lack robustness. Thus, the need arises to
investigate a viable alternative monitoring technique. Previous studies have shown that there
exists a relationship between the shape of the underflow and the operating state. The
evaluation of hydrocyclone dynamic behaviour, by means of image analysis of the
underflow, is therefore considered.

Continuous video recordings of the underflow stream are used to quantify the underflow
width. Time series analysis is employed to visualize the data in terms of a lag plot and
reveal that there exists identifiable structures. Two data groups are identified: a large cluster
relating to normal operating conditions and smaller clusters representing the oscillatory
behaviour of incipient roping. Identification of these operating states supports the use of
image analysis for error-detection of hydrocyclones on industrial scale.

Keywords: Hydrocyclone; image analysis; on-line analysis; classification; underflow monitoring

1. INTRODUCTION

The hydrocyclone is a separation device widely applied in the chemical and mineral processing industries. Its
popularity is promoted by its simple design and operation, compact structure, low running costs and versatility
(Narasimha et al., 2006; Chu et al., 2002). Applications of the hydrocyclone include liquid clarification, slurry
thickening, solids washing and solids classification (Svarovsky, 1984). With any of these operations, the overall
profitability of the process relies on the effective control of the process equipment. In practice, however,
hydrocyclones lack the required controllability and monitoring. This could be attributed to the complex internal flow
features and the economical unfeasibility of measuring them (Petersen et al., 1996).

Previously, various attempts have been made to monitor hydrocyclones. Neesse et al. (2004) summarize and give
examples of the most common methods used thus far. These include the determination of the operating state by
weighing the hydrocylone (Schweitzer, 1972), applying X-ray and electrical impedance tomography (Galvin and
Smitham, 1995; Dyakowski et al., 1996), and using acoustic emission sensing (Leinemann, 1990; Williams et al.,
1996). Other methods entail the determination of the underflow discharge shape by use of mechanical detection
(Hubert, 1993) and electrical capacitance tomography (Williams et al., 1995). As most of the methods are intrusive,
economically unfeasible or lacking robustness, none has found broad adoption yet. However, laser-optical
monitoring of the underflow, as applied by Neesse et al. (2004), proves beneficial as the method is non-intrusive and
has relatively low costs. The work of Neesse et al. (2004) prompted the investigation of another monitoring
technique based on the underflow discharge shape.

Petersen et al. (1996) proposed the use of image analysis of the hydrocyclone underflow discharge as a diagnostic
tool to assess its performance. This technique, like other underflow monitoring techniques, is based on the fact that
the underflow of hydrocyclone is indicative of the prevailing operating state (Neesse et al., 2004; Petersen et al.,
1996). The discharge spray angle therefore serves as a visual variable, and represents the excessively dilute, normal
or roping operating state as defined by Wills et al. (2006). A schematic portrayal of the three operating states is
provided in Figure 1. Since the Petersen et al. study (1996), powerful advances in visual and computer technology,
such as image recognition and computerized interpretation of data, have been made. This allows for dealing with the
computationally intense nature of these types of monitoring problems. In light of this, an investigation into the
dynamic behaviour of a hydrocyclone, by use of image analysis of the underflow width, is made.

Figure 1: Correlation between discharge spray angle and hydrocyclone operating state

2. EXPERIMENTAL WORK

2.1 Experimental Setup and Procedure

A schematic diagram of the experimental setup used during the investigation, is illustrated in Figure 2. The
hydrocyclone system consisted of a 36 mm diameter hydrocyclone placed above an open mixing tank. Both the
overflow and underflow streams were discharged into the mixing tank and the combined stream was recirculated by
use of a centrifugal pump. To capture video data, a Sony DCR-SR300 HDD camera was mounted on the side of the
mixing tank. A light source was directed towards the underflow discharge to provide clear definition between the
spray particles and the background features. Pressure measurements of the feed stream were also taken by use of a
pressure sensor and data logger situated on the feed line.

Experiments were initiated under excessively dilute conditions. To simulate varying inlet concentrations, small
amounts of ore were gradually added to the feed stream until roping conditions were achieved. Figure 3a-3c depicts
recorded images, obtained during each of these conditions.

2.2 Material

Gold ore from an old mining operation was used as material due to its favourable properties: The ore density (1750
kg/m3) was sufficiently low to ensure that particles stay suspended in the mixing tank. Furthermore, the light colour
of the mineral ensured that the underflow stream was clearly distinguishable from the surrounding environment.
Figure 2: Schematic diagram of hydrocyclone monitoring setup

(a) (b) (c)

Figure 3: Recorded images of (a) excessively dilute, (b) normal operating and (c) roping conditions.

3. ANALYTICAL METHODOLOGY

Recorded images were transmitted from the camera to a computer, where analysis of the images was performed in
the software package MATLABTM. Data analysis comprised of various steps, as illustrated in Figure 4. To analyse
the variation in underflow width (Fig. 4a), a single horizontal line was extracted from each video frame.
Thresholding was applied as filter to differentiate between the colour intensities of the line. By choosing the
appropriate thresholding level, the underflow was distinguished from the background (Fig. 4b). The undesirable
points were eliminated by normalizing the image data and making use of 2.5% and 97.5% quantiles (Fig. 4c). Data
in the region between these two quantiles were assigned a value of one. The total of the data vector (or pixel value)
corresponds to the ‘underflow width’ of the frame. Data relating to the underflow width is illustrated by a blue arrow
in Figures 4a-4c.

By repeating the procedure, a pixel value corresponding to the underflow width was obtained for each frame. In
order to determine a data point for every second, the value for every 30 frames was taken (videos were created at 30
frames per second). Consequently, the variation in underflow width could be illustrated against time as in Figure 4d.
Additional analyses were performed in the software package STATISTICATM and are briefly discussed in the
Results section, where relevant.
 (a) (b)

(c) (d)

Figure 4: (a) Extraction of horizontal line, (b) Application of thresholding limit, (c) Removal of outliers by use of
quantile distribution and (d) Resultant variation of underflow width with time

4. RESULTS

Variations in the underflow width were investigated by performing a time series analysis. A lag value of 250
(equivalent to approximately 8.33 seconds) was obtained and the resultant lag plot is illustrated in Figure 5.

0.45

0.40
Underflow width (t-250)

0.35

0.30

0.25

0.20

0.15

0.10

0.05

0.00
0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45
Underflow width (t)
Normal operating conditions Fully developed roping Incipient roping

Figure 5: Lag plot of underflow width variation

Various clusters can be identified in Figure 5, signifying the existence of different data groups. Each data group, in
turn, relates to the different conditions in which the underflow might occur. Normal operating conditions, for
instance, are demonstrated by the largest cluster (triangles). A series of smaller clusters (crosses) are also detected.
The oscillatory behaviour associated with these clusters typically represents incipient roping. Fully developed roping
is illustrated by the cluster in the bottom left corner of the figure (circles).

Further interpretations were made by considering how well the experimental data could be classified into the three
classes identified in Figure 5 (normal operating conditions, incipient roping and fully developed roping). The
predictive potential of the three measured variables (pressure, concentration and underflow width) could then be
investigated by constructing classification models. After dividing the dataset into equal-size training and test sets, a
variety of classification models were fitted. The resulting generalization errors on the test set for the best model from
each type is summarised in Figure 6. More details on the construction of the models are given in Table 1 in appendix
A.

Figure 6: Comparison of generalization accuracy for different classification models (MV = majority vote; SVM =
support vector machines; NBC = naive Bayes classifier; kNN = k-nearest neighbours; RF = random forest; BT =
boosted trees; CT = classification tree; NN = neural network).

The high generalization accuracies for most models indicate strong predictive power of the measured variables.
However, the effect of the large proportion of the normal operating conditions class may conceal misclassification of
the incipient and fully developed roping classes. As illustration, an inspection of the best model (multilayer
perceptron) shows the following class-based generalization accuracies: 98.5 % for normal operating conditions, 74.2
% for incipient roping and 95.3 % for fully developed roping.

Variable importance measures for the neural network, classification tree, boosted trees and random forest models
consistently ranked underflow width as the most important variable, with concentration second and pressure last.
The neural network model reported relative variable importance for underflow width, concentration and pressure of
0.648, 0.280 and 0.071 respectively. None of the input variables has a Pearson correlation of greater than 0.4,
indicating that none of the variables has a strong correlation with any other variable.

The classification modelling exercise shows high predictive power for underflow width, concentration and pressure
(in that order) when discerning between normal, incipient roping and fully developed roping conditions. However,
care must be taken with the development and application of these models, as the specification of classes from the lag
plot can be considered as somewhat subjective.

5. CONCLUSIONS

Results of the time series analysis show how a correlation might be drawn between the underflow width variation
and the hydrocyclone operating conditions. The data also can be classified fairly accurately, as indicated by the high
accuracies obtained for the various models. Image analysis therefore proves to be an appropriate tool for the
identification of a troublesome operating state, such as roping. Further development of the image based method as
an effective monitoring tool is supported by the fact that the method is relatively inexpensive, robust and non-
intrusive.

REFERENCES

Chu, L., W. Chen and X. Lee (2002). Enhancement of hydrocyclone performance by controlling the inside
turbulence structure. Chem. Eng. Sci., Vol. 57, pp. 207–212.
Dyakowski, T., J.A. Guttierrez, M.S. Beck and R.A. Williams (1996). Use of impedance tomography for control of
a dewatering hydrocyclone. In: Hydrocyclones 96 (Claxton, D., L. Svarovsky and M.T. Thew), pp. 215–228.
Mechanical Engineering Publications Ltd., London.
Galvin, K.P. and J.B. Smitham (1994). Use of X-rays to determine the distribution of particles in an operating
cyclone. Min. Eng, Vol. 7, pp. 1269–1280.
Hubert, D. (1993). Measurement method and apparatus for hydrocyclones. Pat. EP0522215A2. Randburg.
Leinemann, J. (1990).Verfahren und einrichtung zur überwachung der teileabtrennung im hydro- und aerozyklonen,
Pat. DE4026460. Berlin.
Narasimha, M., M. Brennan and P.N. Holtham (2006). Large eddy simulation of hydrocyclone – prediction of air-
core diameter and shape. Int. J. Min. Proc., Vol. 80, pp. 1–14.
Neesse, Th., M. Schneider, V. Golyk and H. Tiefel (2004). Measuring the operating state of the hydrocyclone. Min.
Eng., Vol. 17, pp. 697–703.
Petersen, K.R.P., C. Aldrich, J.S.J. Van Deventer, C. McInnes and W.W. Stange. Hydrocyclone underflow
monitoring using image processing methods. Min. Eng., Vol. 9, pp. 301–315.
Schweitzer, K. (1972). Hydrozyklon mit selbst regelnder Unterlaufdüse, Pat. DE-GM 7201703. Hochdahl.
Svarovsky,L. (1984). Hydrocyclones. Holt, Rinehart and Winston Ltd., London.
Williams, R.A., T. Dyakowski, C.G. Xie, S.P. Luke, P.J. Gregory, R.B. Edwards and L.F. Gate. (1995). In: Process
tomography – Implementation for industrial processes (Beck, M.S. et al.), pp. 3–15.EU/UMIST, Manchester.
Williams, R.A., S.J. Peng, D. Brown, N. Parkinson and Pl. Jannes. (1996). On-line measurement of hydrocyclone
performance using acoustic emission. In: Hydrocyclones 96 (Claxton, D., L. Svarovsky and M.T. Thew).
pp.241–252. Mechanical Engineering Publications Ltd., London.
Wills, B. and T.J. Napier-Munn. (2006). Wills’ mineral processing technology: An introduction to the practical
aspects of ore treatment and mineral recovery, 7th Edition. Elsevier, Oxford.

APPENDIX A

Table 1: Model specifications

Model Model description Model selection

Prediction based on majority class of
Majority vote (MV) Not applicable
training data
Radial based function SVM, γ = 1, nd
Tested 2 order polynomial and RBF,
Support vector machines (SVM)
1000 iterations γ = 0.333 & 1
Assumed independent variables
Naive Bayes classifier (NBC) Tested prior and equal class distribution
distribution: Gaussian
Standardized Euclidian distances, Tested Euclidian and Manhattan
k-nearest neighbours (kNN)
k = 20 distances, k = 1,5,10,20
Maximum 10 levels and 100 nodes,
Random forest (RF) Tested mtry = 1 & 2
mtry = 1, ntree =200
Maximum 3 nodes, 200 trees,
Boosted trees (BT) One model
learning rate = 0.1
Maximum 1000 nodes, Gini goodness
Classification tree (CT) Lowest cost
of fit
StatisticaTM Intelligent Problem Solver
Neural network (NN) Multilayer perceptron, 3-3-10-3-1
(best of 10)
(All models, unless stated otherwise, applied estimated prior class probabilities, as opposed to equal class probability)