Minerals Engineering, Vol. 9, No. 1, pp.

55-72, 1996
Pergamon Copyright © 1995 Elsevier Science Ltd
Printed in Great Britain. All fights reserved
0892-6875(95)00131-X 0892-6875/96 $15.00+0.00

D E V E L O P M E N T O F A S O F T S E N S O R F O R P A R T I C L E SIZE M O N I T O R I N G

R. G. DEL VILLAR§*, J. THIBAULTt and R. DEL VILLAR§

Depart. of Mining & Metallurgy, Laval University, Quebec (QC), Canada, G1K 7P4
Depart. of Chemical Engineering, Laval University, Quebec (QC), Canada, G1K 7P4
* Presently at Barrick Gold Co., Bousquet Division, 2 chemin Bousquet,
Route 395, Preissac (QC), Canada JOY 2E0
(Received 26 July 1995; accepted 3 September 1995)

ABSTRACT

Some key control variables of industrial processes, associated with product quality, often
cannot be measured directly or frequently enough to establish adequate control. In such
cases, it is possible to use available measurements to provide a prediction for these
process variables and use them in a control strategy, thereby giving rise to what is now
commonly called a softsensor.

In some industrial grinding circuits, the on-line particle size analyzer is shared between
various sampling points. Therefore, for a given location, the actual measurement is only
available every 10 to 20 minutes, a delay which is unacceptable for automatic control
purposes. To alleviate this problem, a soflsensor based on an artificial neural network has
been investigated. First, the structure of the neural network and different schemes for the
training process are analyzed. Then, the performance of the neural network softsensor is
compared with other inferential methods such as ARMA models and Kalman filters.

Keywords
Neural networks; particle size; process control; on-line analysis; process instrumentation

INTRODUCTION

GRAIIM, French acronym of the Research Group on Computer Applications in the Mineral Industry at
Laval University, began in 1990 a research project on Knowledge Based Automatic Control (KBAC). This
project, sponsored by a consortium of eight mining companies and the federal and provincial governments,
seeked to screen advanced (knowledge-based) methods for data processing (filtering, mass balancing, etc)
and automatic control (model-based) for their use in the mineral processing operations and the transfer of
the project results to the sponsor companies.

In parallel to more academic research work, a number of case-studies have been undertaken with the
cooperation of partner companies. Among such works, a very successful one has been the design and
installation of a new control strategy for one of the Kidd Creek (Falconbridge Ltd.) grinding circuits. This
project was related to the evaluation of a number of control strategies and the implementation of one of
them, as selected by the plant management, in the concentrator.

Presented at Minerals Engineering '95, St. Ives, Cornwall,England,June 1995

9:l-t 55
56 R.G. Del Villar et al.

A total of sixteen control schemes were evaluated on one of the dynamic phenomenological simulators
developed at GRAIIM, a trademark of this research group [1]. The calibration of the simulator models was
done based on industrial data gathered by GRAIIM and Kidd Creek personnel. Then, the simulator was used
to test the performance of various PID-based control schemes, designed to achieve the control objectives
set by the Kidd Creek personnel. The best scheme proposed called for controlling the cyclone overflow
particle size by manipulating the water addition rate to the cyclone feed box and the circulating load by
manipulating the circuit ore feed. The particle size analyzer (PSI-200) available for monitoring the product
size, was unfortunately shared (multiplexed) among various streams of the plant, which meant a 14 min
delay between two consecutive readings for any particular stream. This induced the plant personnel to
choose an alternative control strategy, one which controlled the cyclone feed density instead of the cyclone
overflow particle size.

This situation motivated us to evaluate the possibility of using a different approach to ensure a greater
availability of particle size values for control purposes. One alternative are the so called softsensors or
model-based sensors. A softsensor is an algorithm or a model that makes use of existing information to infer
the value of a variable which cannot be obtained by a physical sensor.

Inferred values obtained through a softsensor can be used reliably for control purposes and can provide a
level of insight into process operations that was not previously available. As such, these softsensors have
attracted considerable interest over the last decade. For instance, due to stringent regulation of emissions
by the U.S. Environmental Protection Agency, a large number of companies are opting for predictive
emission monitoring software in which future values of NO x ,SO x and particulate emissions are foreseen
from present operating conditions using a predictive model that has been previously calibrated on site [2].
Other examples include the estimation of the digester quality control in a batch sulphite pulping process [3]
and the estimation of the volatility of bottom atmosphere gas oil [4].

In the mineral processing industry, achieving an optimal particle size and therefore an optimal liberation
of valuable minerals is an extremely important goal. As such, work has been conducted by some researchers
to develop on-line particle size estimators capable of producing measurements that are both accurate and
frequent enough to establish automatic control strategies. One of the earliest attempts to provide model-
based estimates of cyclone overflow size distributions for control purposes was performed in Australia [5]
where a variant of Rao's equation [6], i.e. a simple linear relationship, was used. Wills [7] reports another
use of the same type of model-based estimator of grinding product size distribution at the Vihanti
concentrator in Finland. More recently, Chilean researchers [8,9,10] have developed particle size estimators
for one of Codelco's concentrators in Chile, using both static (with dynamic compensation) and dynamic
models. These models were based on measurement such as the fresh ore rate, pumpbox level, total water
addition rate and percent solids at the cyclone feed, and were validated within existing control strategies.

Among the many uses of a softsensor some of the most important are: a) the generation of more frequent
values in the case of infrequent measurements resulting from sensor sharing (the case which originated this
work) or simply from a lengthy measurement time, b)the replacement of an existing sensor (by
reconstructing the values of the measured variable from other available readings or historical information)
when the sensor fails or is under maintenance, c) the replacement of a physical sensor for economic or
others reasons, in which case the physical sensor is still necessary at least during the process modelling
period and for the validation of the softsensor, d) the estimation of a given process variable for which no
on-line sensor exists.

To produce such a model-based sensor, various types of mathematical tools can be used. In the present case
we have chosen three well known methods: a neural network algorithm, an auto-regressive moving-average
model (ARMA), and a Kalman filter. The soflsensor was designed to estimate the cyclone overflow particle
size (percent passing 45 pm) based on available information (present and past values of pulp flow rates and
densities of cyclone feed and overflow and past values of cyclone overflow particle size). The obtained
softsensor is presently being implemented in the dynamic grinding circuit simulator for its use within the
control strategy. Based on the simulation results that will be produced, the implementation of such an
algorithm in the Kidd Creek control system will be considered.
 A softsensorfor particle size monitoring 57

In the first section of this paper, the industrial grinding circuit and the control problems which originated
the idea of using a softsensor for particle size estimation will be presented. Then, the dynamic simulator
used in this study (unit models and main features) and its calibration with industrial data will be detailed.
In the second section, the various mathematical tools used for developing the softsensor will be presented:
neural networks, ARMA models, and Kalman filters. In each case, the model structure, the calibration
procedure and the possibility for adaptation will be discussed. The third section is devoted to the
presentation and analysis of the simulation results. Finally some conclusions are drawn.

BACKGROUND

Grinding circuit description

The grinding circuit being considered in this study corresponds to that of the Kidd Creek concentrator
(Falconbridge Ltd.) located near Timmins, Ontario. It processes 3.5 MM t/y of a Cu-Zn ore. The present
work is related to one of their grinding circuits (line B), a classical rod mill (Allis Chalmers 3.2 m x 4.9 m)
- ball mill (Allis Chalmers 3.7 m x 5.5 m) operating in closed-loop with a cluster of eight 380 m m Krebs
hydrocylones. Both mill discharges are mixed in a 12 m 3 non-linear pump-box and the resulting pulp is fed
to the hydrocyclones using a 12"x10" Linatex fixed-speed-drive centrifugal pump (Figure 1).

I)(': Cycl. Ovclflow I)¢lisily('oillrollcr (PI)

I)I': Ikenslly'l'lllllsn|hler
I:I(C: Walcr P'low("mllloiler COFD
lq': l'Iow"llilllSllllllCf ~/¢1l)iJilil
M: ('onvcy(ir"i I)rlv¢
II: Ralio f:olllloller
water W R C : I:¢cd Ralc ('Olllroller
W'i': Wcighl onlelcl"
I
!

i I

~ : ~ .... lilio ICl

Mill
@.i,, Mill
OIC
lind

I ''+',,,.
!
I liMP
$cl poln!
Fig.1 Kidd Creek grinding circuit

When the process control study was undertaken by GRAIIM, the available instrumentation consisted of a
Milltronics Compuscale A weightometer on the rod mill feed belt, three Rosemont magnetic flowmeters to
measure the water flow rate to the rod mill (ratio-controlled), the ball mill and the cyclone pump-box.
Cyclone overflow pulp density is measured by a Texas-Nuclear ~/-ray gauge, whereas ball mill discharge
pulp density is measured several times during a shift using a Marcy scale. Product particle size is obtained
from an Outokumpu PSI-200 particle size analyzer located on the cyclone overflow line. This device, whose
response time is 2 minutes, was shared between seven points in the concentrator implying that each stream
58 R.G. Del Villar et al.

is sequently analyzed every 14 rain. Since the implementation of the new control strategy, instruments for
monitoring the cyclone feed flow rate and pulp density were installed and their readings were considered
for the design of the softsensor.

Before the GRAIIM control project was initiated, Kidd Creek grinding line B was operated using the
following control strategy (Figure 1):

The rod mill feed tonnage was kept constant by manipulating the speed of the feeder conveyor,
to achieve a daily average tonnage.

The cyclone overflow pulp density was kept constant by manipulating the water flow rate to the
pump box, to avoid an increase in pulp density which might cause pipe blockage.

The rod mill pulp density was regulated by maintaining a constant water/ore ratio to achieve good
milling efficiency.

The ball mill discharge pulp density was kept constant by manipulating the water flow rate at the
mill feed box, to ensure a good milling efficiency. Since the ore feed rate was not measured, only
the mill percent solids was available, this control was done manually.

To respond to abnormal operating conditions (e.g. stoppage for rod addition, extreme pump-box pulp level,
plugged cyclones), operators normally manipulate the circuit tonnage set-point, the number of cyclones in
operation or the cyclone overfow density set-point.

The grinding dynamic simulator

Data used to develop the softsensor was generated using a grinding circuit dynamic simulator (Dynafrag)
[11,12], whose flowsheet configuration corresponds to the Kidd Creek grinding circuit. The simulator
includes the following process unit models: rod mill, ball mill, hydrocyclone, pump box, centrifugal pump,
piping and conveyor belt. The ball mill is modelled using a population mass balance model [13], which
requires the evaluation of the breakage function (characteristic of the ore), the selection function and the
residence time distribution (both dependent on the mill environment). Since rod mill operation also implies
some form of classification (some particle size classes disappear), the standard population mass balance
model is not sufficient, a situation which has been successfully overcome in Dynafrag by modelling the rod
mill as two ball mills in series. The hydrocyclone performance is reproduced by a variant of Plitt's model
[14]. The whole process dynamic model then consists of a set of mass-balance equations (a hundred
differential equations) for pulp, water and particles of differing hardness.

A certain number of sensors and actuators are also reproduced in the simulator since they contribute to the
system dynamic response (Figure 2). Five different actuators are implemented: the belt conveyor drive to
control the fresh ore feed rate (Ul), the valve of the water addition to the rod mill feed (u2), the pump box
water valve (u4), and the ball mill water valve (Us). The actuator u 3 , the variable speed drive to control the
pump speed, was transformed in the present study into a fixed speed drive to reproduce the Kidd Creek
circuit. In addition, seven sensors, with optional corresponding noise patterns, are simulated to measure: the
rod mill ore feed rate (ml), the pump box pulp level (m2), the cyclone feed pulp flow rate (m3), the cyclone
feed pulp density (m4), the cyclone overflow pulp flow rate (ms), the cyclone overflow pulp density (m6),
and the cyclone overflow percent passing 45 pin (mT).

Typical grinding process disturbances can be simulated through the following internal variables: the ore
hardness (Pl), the rod mill feed size distribution (P2), the ore specific gravity (P3), the cyclone pump
performance (P4), and the cyclone operation feature (P5) which can accommodate various types of cyclone
problems (apex blockage, apex changes, addition/removal of units). In this study only Pl, P2 and P5 were
used. Actuators and sensors can be simulated using additive noises: ai to the manipulated variables (u i ) and
n i to the simulated measurement values (m i ). Various types of noise pattern are implemented in Dynafrag:
white noise, step, ramp, AR, MA, ARMA model, PRBS and PRTS.
 A softsensorfor particle size monitoring 59

u d-[ u2 u4 /
al. ]/°ill, [}l-a2-- ~~- a4 a5~ d5~

~~n l~m!n~,~,~1a3~__~n3[~
[ ~[~~
04
m2 i d4 ~
6miii
u3 m3 m4
Fig.2 Dynafrag, grinding circuit dynamic simulator

The processed ore is characterized in Dynafrag in terms of three properties: size distribution, hardness and
specific gravity. An adequate mixture (flow rate) of two key components, each having its own size
distribution, hardness and specific gravity, allows the reproduction of any type of ore. By changing the
proportion of both key-components, the effect of disturbances in hardness, size distribution and specific
gravity on the final product quality or the process behaviour can be easily studied. The ore hardness is
implemented in the grinding model through the mill selection function whereas the circuit feed size
distribution is directly used in the calculations. Ore hardness is categorized in three levels, hard, medium
and soft, whereas size distribution can be depicted as coarse, medium and fine. Desired disturbances can
be pre-programmed (type and magnitude) and then used at anytime during the simulation.

Before using Dynafrag for simulation purposes, a calibration on the stated circuit was performed. Both the
steady-state and dynamic behaviour of the industrial grinding circuit were identified and accommodated into
the mathematical models implemented in the simulator. The steady-state part of Dynafrag was calibrated
through grinding and residence-time data obtained from carefully planned and completed sampling
campaigns [12]. A total of nine steady state grinding tests were completed at different operating conditions.
Samples were collected at the rod mill feed and discharge, at the ball mill feed and discharge, and at the
cyclone overflow. Tracer tests were performed at both rod and ball mills to determine the residence time
distribution. In the case of the rod mill, only tracer concentration data at the mill discharge were necessary
since it operates in open loop. In the case of the ball mill, tracer data at both mill ends, feed and discharge,
were required to deconvolute the open-loop RTD model (in this case the mill is continuously submitted to
new tracer impulses as it recirculates to the mill via cyclone underflow). The gathered data was reconciliated
using BILMAT, a material balance program [ 15]. Residence time distribution data (Li + concentration) were
processed using the SPOC software [16]. Then the methodology proposed by the SPOC project was
followed to determine the steady state model of both mills [17] and that of the hydrocyclones [18].

SOFTSENSORS FOR PARTICLE SIZE MONITORING

Since the PSI-200 measurements were not sufficiently frequent for control purposes, the concentrator
management decided not to retain the control scheme based on product size distribution envisaged by the
GRAIIM in its study. Instead, a strategy based on the cyclone feed density was chosen. This situation
motivated GRAIIM to explore some model-based methods which would allow a greater availability of
particle size values for control.
60 R.G. Del Villar et al.

To predict the cumulative percent passing 45 lam (325 mesh) of the cyclone overflow, three different
approaches were used. In this section, they will be described in turn. The process variables used in each
model as input variables in view of predicting the percent passing 45 ~tm are variables that are measured
on-line at each sampling instant.

Neural Network Models

Feedforward neural networks have attracted considerable interest in the past decade for their ability to
capture with relative ease underlying phenomena between input and output variables of a process. A
feedforward neural network can be viewed as a general nonlinear model. It is made of a series of processing
units called n e u r o n s , organized in layers. Typically, a neural network comprises three types of layers: an
input layer, one or more intermediate or hidden layers and an output layer. The neurons of the input layer
are simply used to scale each of the input variables and to distribute the scaled inputs to each neuron of the
first intermediate layer. Each neuron of this first intermediate layer performs a weighted sum of the outputs
from all the neurons of the previous layer, processes this sum through a nonlinear transfer function to
produce the output of that neuron, which in turn is transmitted to each of the neurons of the next layer. A
constant-value neuron, called the bias, is included in the input layer and in all hidden layers, its role being
similar to that of the constant term in a standard linear regression model. A neural network is thus
characterized by the number of neurons and layers, the weights of the interconnections between neurons of
adjacent layers and the nonlinear transfer function.

Before using the neural network for predicting values of a given process variable, it must be trained, that
is the internodal weighting factors (parameters of the model) must be determined. Network training is
performed on a set of data by comparing the predicted output values to known target values and calculating
the corresponding error for each of the output neurons. The errors are then used to adjust the weights of
the connections, thus improving the quality of the output variable prediction. At the same time, the ideal
number of intermediate layers and neurons (best network architecture) must be determined. Great care must
be exercised in the choice of target values to be used in the training process to ensure the quality of the
network predictions. Values should be chosen proportionally to the likelihood of their occurrence in the
process. Preprocessing of the samples using domain knowledge is a mandatory step if the performance of
the network is to be improved by reducing the number of neurons and therefore reducing the time necessary
for the training. It is then clear that the use of this procedure is conditioned to the feasibility of the neural
network training. For a more detailed description of neural networks, the reader is referred to the numerous
and excellent textbooks on the subject [19,20].

In this investigation, a typical neural network is presented in Figure 3a where the cyclone feed flow rate
(CFF), the cyclone overflow flow rate (COFF), the cyclone feed density (CFD) and the cyclone overflow
feed density (COFD) are used as inputs to the network to predict the percent passing 325 mesh (M325).
Numerous neural network architectures were investigated in view of selecting the architecture that will
predict with adequate robustness and accuracy the percent passing 325 mesh. The architecture that was
selected, as will be discussed in the results section, is presented in Figure 3b. In this architecture, the current
and immediate past values of the four input variables along with the past percent passing 325 mesh were
used to predict the current percent passing 325 mesh. At first glance, this structure may be surprising since
the dynamics of the hydrocyclone is essentially instantaneous compared to the dynamics of the grinding
circuit. It is reasonable to argue that past values should not have a direct incidence on the current particle
size distribution at the hydrocyclone. However, the use of past values of the percent passing 325 mesh and
of the flow rate and density of the inlet flow and the overflow of the hydrocyclone in addition to their
current values allows capture of some of the dynamics of the overall grinding circuit.

It is possible that the dynamics of the grinding circuit varies from the time when the data for training the
network was gathered, leading to a deterioration of the particle size estimation. To remedy this problem,
continuous adaptation is possible whereby, at each sampling instant, the most current input/output
information is used to modify the weights of the network in order to produce a better fit. Doing a few
backpropagation iterations using the most current information or to use a moving window comprised of a
number of past measurements are two ways to achieve this objective [21].
 A softsensor for particle size monitoring 61

CFD (t)-'~[-], (b)

COFF(t)-'~[-], 0
CFD "~D O~ COFD(t)--~[-], '
COFF ~ ] - - ] " ~ 0 7 0 "-'~ M325 CFD
CFF (t-1)---~["],
l)---~["], ~ 0----~ M
COFD~E]~0~ COFF (t-1)-~[31
COFD(t- 1)--~[-]J
1 [7 1 "-~0 / M325 (t-1)--~r-]J
Input Hidden Output 1--~[-1 0
Fig.3 Neural network structures

ARMA Models

The linear counterpart of the neural network model is the auto-regressive moving average (ARMA) model.
An ARMA model can be viewed as a neural network with an input layer, no hidden layers and an output
layer, using a linear transfer function, as depicted in Figure 4. Among the advantages of an ARMA model
are its simplicity and the ease of model calibration. The model parameters can be calculated using well-
known linear regression techniques to minimize the sum of squares of the errors. The main disadvantage
is the usually narrow range of applicability for nonlinear processes. To circumvent this disadvantage and
to compensate for process changes, the parameters of the ARMA model can be updated on-line at each
sampling interval using the recursive least squares algorithm. This algorithm uses the most current
information to calculate the estimation error and to change each of the parameters of the model to lead to
a better process output estimation [22].

CFF (t)
CFD (t)
Output
COFF (t) [--7 layer
COFD (t) M325 (t)
CFF (t-l)
CFD (t- 1)
COFF (t- 1)
COFD (t- l)
M325 (t- 1) ~ ~ Inputlayer
Fig.4 ARMA model structure

As in the case of the neural network models, a number of structures have been tested ranging from a purely
static model to models with a large number of past input and output values. The structure that was retained
in this investigation uses identical information as the one used for the neural network model (Figure 4).
62 R.G. Del Villar et al.

Kalman Filter

The Kalman filter is used to estimate, using noisy measurements, the values of the state variables of a
process subject to stochastic input disturbances [23]. To perform an estimation, a reasonably good model
is required in order to relate measurements with state variables. A popular model commonly used with
Kalman filters is the state-space model which is a representation of the dynamic behaviour of the process
using a set of simultaneous first order differential equations and an auxiliary state vector. A schematic
representation of the Kalman filter is provided in Figure 5.

Model Filtered
meassured
/ variables
Kalmaa
Filter am easured
variable

Process
Measurements
T estimate

Fig.5 Kalman filter representation

In this investigation, a discrete state-space model has been used to estimate the percent passing 325 mesh.
This model was obtained from a simple ARMA model and it is given by the following set of equations:

x,, i "- A x t + B u t + w (1)

Yt : C x t ÷ Sit
(2)

where x is the vector of state variables, y the vector of outputs variables, u the vector of input variables,
w the process noise acting on the states, v the measurement noise, A the system matrix, B the control matrix
and C the output matrix. In the present case, the x vector corresponds to the cyclone overflow rate and
density and the percent passing 325 mesh, the y vector to the cyclone overflow rate and density and the u
vector to the cyclone feed rate and density. The C matrix is defined so as to solve the relationship presented
in equation 2, (the third row of the matrix, corresponding to the percent passing 325 mesh has a zero value).

While it is not the intention in this paper to present a derivation of the Kalman filter, it should be sufficient
to say that the conditional expectation of Yt, given past input and output data and the state-space model, is
given by

<,: 8u,. <,[y,.,- (3)

 A softsensorfor particle size monitoring 63

In other words, the state vector x is modified in order to reduce the model prediction error. K is the gain
matrix that weighs this error vector. It is calculated at each sampling instant and depends on the covariance
matrix of the state estimate error and the variances of the two noise matrices w and v. For a more detailed
description, the reader is to referred to Ljung's book [24].

A phenomenological model of the hydrocyclone could have also been used instead of a state-space model
for the estimation of the percent passing 325 mesh similarly to what has been done in the case of a
bioreaction [25] and an exothermic chemical reaction [23].

RESULTS AND DISCUSSION

The evaluation of the performance of each of the three different types of particle size softsensors developed,
neural network (NN), ARMA model (AM), and Kalman filter (KF), was performed on data obtained from
Dynafrag by introducing step changes to the set-points of some of the manipulated process variables and
by modifying the state of some of the disturbance variables.

To quantify the influence of the different operating variables, various individual step changes were made
in Dynafrag while in steady-state operation:

___30% step changes in circulating load set-point (768 t/h);

up to _+20% step changes in cyclone feed percent solids set-point (75%);

step changes in ore composition, various feed mixtures were used ranging from 100% of one type
of ore (hardness) to 100% of a different type of ore;

sudden blockage and/or closure of one cyclone (out of a total of eight);

preprogrammed changes of the feed size distribution;

The simulation results showed that a continuous series of step changes of random but restricted duration
and amplitude in ore hardness and particle size provided a data set rich enough for the development of the
softsensor. The length of the data file thus obtained corresponds to 4V2 hours of process time (almost 200
data points). For comparison purposes two sets were generated, where applicable the first set was used for
model calibration (learning data set) while the second was used as a baseline validation set, data that was
not used during the model calibration. Each of the obtained data sets were then corrupted with different
levels of noise to produce three data sets called noise-free, noise level 1 and noise level 2. The noise level
selected to produce the first noisy data set is equivalent to that detected in Kidd Creek data records, whereas
the one used for the second set is twice the previous value. These three data sets permitted the evaluation
of the softsensor performance in noise-free and industrial-like environments.

Neural Network

As a first step, tests were conducted to establish the network structure that would result in both a simple
and accurate particle size sensor. To that effect, simulations were made for a varying number of input nodes
(each corresponding to the value of a particular operating variable at present and past sampling intervals),
hidden layers (1 and 2) and nodes (1 to 10, 15 and 20 neurones per layer), transfer functions (hyperbolic
tangent, sigmoidal and linear) and learning methods (quasi-Newton and standard back-propagation).

Given the results it was concluded that a network of the structure presented in Figure 3b produced accurate
results while at the same time limiting the complexity of the network. The retained structure included an
input layer made of 10 nodes (including the bias), fed respectively with present and past (one sampling
interval back) values of the flow rate and density of both cyclone feed and overflow and the inmediate past
64 R . G . Del Villar et al.

value of the cyclone overflow particle size, a three-node hidden layer and using a standard back-propagation
method for training.

Figures 6 present a comparison between the softsensor predictions (solid line) and the data sets generated
by Dynafrag for: (a) the training data set without any noise, (b) the validation data set (not used in the
learning process) without any noise, (c) the validation data set corrupted with noise level 1, and (d)
corrupted with noise level 2. It can be seen that in all cases the neural network predictions are in good
agreement with the actual size distribution "measurements" (percent passing 325 mesh values provided by
Dynafrag).

49
L
~ 48
(a)
m

47 ,¢,

•~ 46 tp

m., 45 tl,

44
~ 43 I I I I I I I I

0 1 2 3 4 5
time (h)
~ 49
L • ° • •
(b)
t"q 48
t"Q
e~O
47

46

45 I I I I I I I I I

0 1 2 3 4 5
time (h)
Fig.6 Neural network softsensor predictions
(a) training data set (noise-free); (b) validation data set (noise-free)
 A softsensorfor particle size monitoring 65

r/) 49
E . (c)
tt')
t"q 48
e~t)
• ~,,,I
47
""" • .V.'--
46 4 • •

I I ! I ! I I I I
45
0 1 2 3 4 5
time (h)
49
(d)
¢q 48 4' • 0 ~

=
.- 47 .: k'"

46
tD

I I I I I I I I I
45
0 1 2 3 4 5
time (h)
Fig.6 (cont.) (c) validation data set (noise level 1); (d) validation data set (noise level 2)

A series of simulations were then performed to study the influence of the sampling horizon, o r in other
words to evaluate the problem of an instrument with a long measurement rate. The sampling horizon refers
to the number of sampling intervals without actual instrument readings, as a result of a long measuring time
and/or the sharing of the instrument (increased number of sampling points). For longer sampling horizons,
less information is available to calibrate the neural network model and, as a result, the cumulative percent
passing 325 mesh is estimated with a reduced accuracy. In addition, in those cases where the neural network
model requires past values of the particle size, past estimations must be used with their corresponding level
of error.

In the case of recursive modelling methods, a longer sampling horizon also means that less information is
available for the adaptation of the model. This means that both initial adaptation and adaptation related to
changing process conditions can only be performed at longer intervals, thereby deteriorating the overall
performance of the model. Also in the case of an error in adaptation, less frequent measurements mean that
66 R.G. Del Villar et al.

the eventual corrections will take longer and therefore any single error in adaptation will have a greater
impact on estimation. Obviously this behaviour is common to all three modelling methods, and an example
will be discussed in the following section of this paper.

To simulate the effect of a changing sampling horizon, only certain points from the data set were provided
to the network for training (e.g. for a sampling horizon of 2, every second point was provided). Sampling
horizons of lenght 1, 2, 4, 10 and 20 were simulated and the results are summarized in Figure 7. It can be
seen that for the training data set (Figure 7a, squares), the network is able to come up with a reasonable
model that does not deteriorate significantly with the sampling horizon. However, if the network is provided
with less information (longer sampling horizon) it is increasingly difficult to produce a generalised model
for the prediction of the particle size, as evidenced by the results obtained with the validation data set
(Figure 7b, squares).

When the network is adapted at each sampling instant, better estimation performance is achieved (Figures
7, triangles). Contrary to the standard neural network learning procedure (limited to the information
contained in the training data set), the adaptive neural network is continuously trained (the interconnection
weights are calculated) as new information is available. To this effect the network was first trained with a
third data set and then it was adapted using both the training and the validation data sets.

20
LI~
t..,
© • slalldard
15 I,dd,l[,tivc'l
(D

• m.
¢-J 10 . . . . . . . . . . . . . . . -" . . . . . --~77-= ........ . . . . . . . . . . .

c.t) .. .- -

5 (a)
Oq

I I I I
0
1 2 4 10 20 50

sampling horizon
50
[~ ..,,,d.,'d [ .....
40 adal~tiv e ..........

(D
30 ..,-
..,.

• . ..." •

• .," Ii" •

= 20 . . . . . . . --" ,-..~.7" ..... ". . . . .

,...,.. •
. .,....'- "

E lO (b)
r/3

0 I I I I

1 2 4 10 20 50
sampling horizon
Fig.7 Neural network softsensor adaptation
(a) training data set; (b) validation data set
squares = no-adaptation ; triangles = adaptation
 A softsensor for particle size monitoring 67

ARMA models

Both the model structure and the simulation data described in the neural network section were used for the
ARMA model to allow a meaningful comparison between these two methods. Initial tests of an ARMA
model focused on simple static models, using a similar structure as that developed for the neural network
(Figure 4). Those simulations resulted in extremely accurate predictions for noise free data but performance
deteriorated for the noisy data. To improve these results a recursive calibration of the model was attempted,
resulting in accurate predictions of the particle size in the case of noise free data. Results for the training
and validation data sets are given in Figure 8 (first and third bars for an horizon of 1). While the
introduction of noise deteriorates the performance of the ARMA predictions (second and fourth bars), it
nonetheless gives better results than the static models previously used.

200
r.¢3 • training no noise
O - training noise 2
150
• validation no noise
II validation noise 2

100

r/3
50
,,:=:::: .......,
0
1 2 4 10 20
sampling horizon
Fig.8 Effect of sampling horizon on ARMA softsensor predictions

As the sampling horizon is increased, the ARMA model has to rely more heavily on its own to estimate
the percent passing 325 mesh at each control instant. Since only an ARMA model with recursive adaptation
was shown to be acceptable, an increase in the sampling horizon leads to a dramatic deterioration of the
particle size estimates as the ARMA model requires to be continuously adapted to perform adequately
(Figure 8). As was discussed in the previous section any adaptation error that occurs in a simulation using
long sampling horizons produces a greater estimation error, which due to the lack of actual measurements
lasts for a longer period of time before it can be corrected. A typical example of this behaviour is presented
in Figure 9b where the long sampling horizon produces a significant instability in the estimations, a situation
which is not seen when a short horizon is used (Figure 9a).

Kalman filter

As can be seen in equation 2, the process output estimates are calculated using the state vector which in
turn is calculated using two different pieces of information: the states predicted by the state-space model
and the most current measurements, expressed in terms of prediction errors. The proportion of each
contribution depends on the gain matrix of the Kalman filter. A null gain matrix implies that we fully trust
the model. As the gain is increased, more importance is given to the actual measurement.

The results presented in Figure 10a corresponds to the case where the gain matrix was initialized to zero.
This procedure is the easiest and the fastest as no analysis is necessary to determine precise initial values.
The results clearly show that the model by itself does not contain enough information to produce accurate
initial estimates of the percent passing 325 mesh. However, once the algorithm has used several measured
68 R.G. Del Villar et al.

values to calibrate the weighting matrix, very good estimates are obtained. If on the other hand, a more
reasonable initial gain matrix is provided, the Kalman filter is able to give accurate estimates in a relatively
low number of sampling periods as can be visualized in Figure 10b. In this case, the initial gain matrix
corresponds to the final weight matrix obtained in a secondary simulation.

.-= 50
E (a)
49
rr~
48
°~

47
,Ib ~,o
4-a

=
~D 46
G)
G) i l I T i I i I
c~ 45
0 1 2 3 4 5
time (h)
,..CZ
49
(b)
tr3 48
t',l
¢e)
47
rae~
46

45
¢..)

i I I I t I i I
44
0 1 2 3 4 5
time (h)
Fig.9 Recursive A R M A softsensor predictions validation data set, noise level 2, sampling horizon:
(a) 1, (b) 10

In its present form, the Kalman filter does not use the percent passing 325 mesh to correct the state vector,
it only uses the other process variables wihch are available at each sampling instant. Under these conditions,
the Kalman filter provides a particle size estimate at each sampling instant and the concept of the sampling
horizon is not applicable. The results presented in Figures 10a and 10b correspond to this type of model.
Presently, work is under progress to include the particle size measurement in the output vector instead, only
then can the influence of the sampling horizon be tested.
 A softsensor for particle size monitoring 69

tt~
¢,q
48 t
i:
" t •oO",,•
•
(a)
47

• ,-., 46
45
¢~ 44
~D
43 t I I I I I I I I

0 1 2 3 4 5
time (h)
,x2
r/3 49
~D
(b)
tt~
48
-° -o 0
ca) o~
47
r,t3 • • ~0•
• • •4 000 •
,¢,,,a • •IN,•
46
O• • 41.0

~D
45 I I I ; I I s I I

0 1 2 3 4 5
time (h)
Fig.lO Kalman filter softsensor predictions validation data set; sampling horizon = 1
(a) non-initialised; (b) initialised

If the sampling horizon is increased while the control time interval remains constant, the algorithm has to
rely more heavily on the process model to provide good estimates of the controlled variable. For instance,
if an actual measurement is available every five control periods, the algorithm has to estimate, for four
consecutive control time intervals, the controlled process output from the model and from other
measurements that are available at each control instant. The meaningful adaptation of the filter would only
be done at each process output sampling instant. Therefore, to insure that the filter will provide good
estimates of the process output, it is mandatory to initialize adequately the gain matrix of the filter. This
problem is akin to the imperfect adaptation seen with an ARMA model in the case of long sampling
horizons. This study has shown, however, that the Kalman filter is more robust than the ARMA model to
imperfect adaptation.
 70 R.G. De| Villar et al.

Figure 11 presents a comparison between the three different approaches for the estimation of the percent
passing 325 mesh as a function of the noise level and for two distinct sampling horizons. Estimation
algorithms based on a properly initialized Kalman filter or a neural network model are very efficient with
a slight edge to a properly initialized Kalman filter with respect to estimation accuracy and robustness.
However, the estimation algorithm based on a neural network is more straightforward to develop and to use
because it requires significantly less expertise.

140
120
© [] Kalman Filter
~,, K.F. initialized
100
• Adaptive ARMA
80 Neural Network

= 60
r~
40
= 20
0
no noise 1 2

noise level
Fig.11 Softsensor comparison (noise effect)

CONCLUSIONS

In this work three approaches have been considered for the development of a softsensor capable of
estimating the particle size (cumulative percent passing 45 ~tm) of the final product (cyclone overflow) of
a grinding circuit in view of its use within an automatic control strategy. The first approach is based on
neural network models, the second uses an ARMA model with recursive identification and the third is a
Kalman filter. The three methods exhibit good performance when the sampling horizon is relatively short
and/or the input variables are free of noise. In the presence of noisy data the ARMA model deteriorates
considerably and a recursive model must be used to maintain adequate predictive capabilities. Nontheless,
it lacks robustness for increased sampling horizons. The neural network, even in its simplest form (without
adaptation), provides excellent estimates of the cyclone overflow particle size. Moreover, it is reasonably
robust in terms of increasing sampling horizon and with respect to the level of noise of the input variables.
The adaptive version of the neural network increases dramatically the robustness of the approach in what
concerns the problem of sampling horizon. Also, this approach requires less expertise from the user and is
able to capture the underlying characteristics of the process over a wide range of operation. Finally, the
Kalman filter has shown excellent performance when it is properly initialised, otherwise it requires a certain
time before being able to produce adequate estimates of the output variable. It must be pointed out that this
model was used only for a case where all physical measurements used were available at each control instant.
A study is being pursued to use the Kalman filter with long sampling horizons by including the particle size
measurements. The neural network soflsensor will be implemented in a grinding circuit control strategy and
its performance will be evaluated in process control.
 A softsensorfor particlesize monitoring 71

ACKNOWLEDGMENTS

The authors wish to thank GRAIIM (Universit6 Laval) for its support in the execution of this project.
Funding has been provided through the KBAC project, in turn supported by a consortium of eight mining
companies (Cominco, Noranda, Falconbridge, Les Mines Selbaie, Mining Metall Co, Cie. Mini~re Qurbec-
Cartier, Brunswick Mining & Smelting, QIT-Fer et Titane), the Mining Technology Council of Canada
(MITEC), the Association Mini~re du Qurbec, CANMET, the Centre de Recherches minrrales and NSERC.
We would also like to acknowledge the support of Frrdrric Flament, research assistant at GRAIIM.

REFERENCES

. Flament, F., Desbiens, A., del Villar, R., Jakelski, D.M. & Pinard, S. Design of a control strategy
for Kidd Creek grinding circuit line B. 26th Annual Meeting of the Canadian Mineral Processors.
Ottawa, (Jan. 18th, 1994).
.
Samdani, G.S., Software Takes on Air Monitoring. Chemical Engng., 30-33 (Dec. 1994).
3. Rao, M. & J. Corbin., Intelligent Operation Support System for Batch Chemical Pulping Process.
Proceedings of lFAC Symposium on Process Dynamics and Control. College Park (MD). 405--410
(1992).
. Kim, H.C., Shen, X. & Rao, M., Artificial Neural Network Approach to Inferential Control of
Volatility in Refinery Plants. Proceedings of 2nd IFAC Workshop on Algorithms and Architectures
for Real-Time Control. Seoul (Korea), 90-93 (1992).
.
Lynch, A.J., Mineral crushing and grinding circuits. Elsevier Scientific Publications Company.
Amsterdam, 342 (1992).
6. Rao, T.C., The characteristics of hydrocylones and their application as control units in
comminution circuits. Ph.D. Thesis, University of Queensland, Australia (1966).
7. Wills, B.A., Mineral Processing Technology. 5th Edition. Pergamon Press. Oxford, 855 (1992).
8. Gonzalez, G., Cipriano, A., Feddersen, A., Zamora, C. & Valenzuela, J., On-line dynamic particle
size estimator for grinding plant control. Proceedings of IFAC Symposium on Identification and
System Parameter Estimation. Beijing (PRC). (1988).
.
Gonzalez, G., Mendez, H. & de Mayo, F., A Dynamic Compensation for Static Particle Size
Distribution Estimators. ISA Transactions., 25(1), 47-51 (1986).
10. Gonzalez, G., Gonzalez, M.A. & Cartes, J.C. Control problems due to the replacement of sensors
by softsensors. ISA paper #92-0396 (1992).
11. Dubr, Y., Lanthier, R. & Hodouin, D., Computer aided dynamic analysis and control design for
grinding circuits. CIM Bulletin. 80(905), 65-70 (1987).
12. Flarnent, F., del Villar, R. & Lanthier, R., Computer aided design of a control strategy for an
industrial grinding circuit. Computer Applications in the Mineral Industry. Second Canadian
Conference, Ed. A.L. Mular. Vancouver. (Sept. 1991).
13. Austin, L.G., Klimpel, R.R. & Luckie, P.T., Process Engineering of Size Reduction: Ball Milling.
ISBN 0-89520-421-5. SME/AIME. New York, 561 (1976).
14. Plitt, L.R., The mathematical modelling of the hydrocyclone classifier, CIM Bulletin, 69(766),
114-123 (1976).
15. Hodouin, D. & Flament, F., The SPOC manual, chapter 3.1, BILMAT computer program.
CANMETkEMR SP85-1/3.1E. Ed.D.Laguitton. Canadian Government Publishing Services. Ottawa
(1985)
16. Flament, F., Hodouin, D. & Spring, R., The SPOC manual, chapter 7.3, RTD and MIXERS
computer programs. CANMETkEMR SP85-1/7.3E. Ed.D.Laguitton. Canadian Government
Publishing Services. Ottawa, (1985).
17. Gupta, V.K., Hodouin, D. & Spring, R. The SPOC manual, chapter 7.2, FINBS computer program.
CANMETkEMR SP85-1/7.2E. Ed.D.Laguitton, Canadian Government Publishing Services. Ottawa,
(1985).

18. Laguitton, D., The SPOC manual, chapter 5B, Unit models (Group B). CANMETkEMR SP85-1/1E.
Ed.D.Laguitton. Canadian Government Publishing Services. Ottawa, (1985).
ME 9:l-F
72 R.G. Del Villaret al.

19. Haykin, S., Neural Networks - - A Comprehensive Foundation. Maxwell Macmillan Canada.
Toronto, 696 (1994).
20. Caudill, M. & Butler, C., Naturally Intelligent Systems. The MIT Press. Cambridge, 304 (1990).
21. Van Breusegem, V., Thibault, J. & Ch6ruy, A., Adaptive Neural Models for On-line Prediction in
Fermentation. Can. J. Chem. Eng., 69, 481--487 (April 1991).
22. Young, P., Recursive Estimation and Time-Series Analysis - - An Introduction. Springer-Verlag.
New York, 300 (1984).
23. Roffel, B. & Chin, P., Computer Control in the Process Industries. Lewis Publishers Inc. 257
(1987).
24. Ljung, L., System Identification: Theory for the User. Prentice-Hall Inc., Englewood Cliffs 519
(1987).
25. Thibault, J., Van Breusegem, V. & Ch6ruy, A., On-line Prediction of Fermentation Variables Using
Neural Networks. Biotechnol. Bioeng., 36, 1041-1048 (1990).