Professional Documents
Culture Documents
www.elsevier.com/locate/powtec
Abstract
It is now well accepted that process automation, closed-loop control and optimisation technologies enable higher feed rates, more
consistent product qualities, lower utility usage and higher yields in a diverse range of process industries. In the minerals processing industry,
as a general rule, it is advisable to consider the grinding circuit first, as good operation here is necessary for stable operation in the rest of the
plant. Operating conditions in a grinding circuit are frequently such that production is significantly reduced due to the effects of process
disturbances, most notably those associated with the feed ore. This paper demonstrates the design of a robust control scheme for a semi-
autogenous grinding (SAG) mill. Feedback (FB), feedback plus feedforward (FF) and reduced order (i.e. less complex) robust controllers
were all developed and testedeach providing tight control over the SAG mill power draw by adjusting the feed rate to the mill. This
example clearly shows that using the available Matlab-based controller design package, it is possible to develop robust (i.e. in terms of strict
guarantees on both controller performance and stability, despite model uncertainties) control systems for the minerals industry as readily as
for other industries where advanced control is more routinely used. D 2002 Published by Elsevier Science B.V.
was used in this model, the largest size (designated size dsi X
i1
f2;i f11;i aij rj sj ri si i 2; . . . ; n 1 2
1) and smallest size (designated size n in the following dt j1
equations) being particles with nominal diameters of 152
and 0.001 mm, respectively. Such a population balance
approach gives rise to the following set of differential dsn X
n1
equations for the SAG mill (all variables are defined later f2;n f11;n aij rj sj 3
dt j1
in the Nomenclature):
ds1 An additional differential equation is used to describe the
f2;1 f11;1 r1 s1 1
dt variation with time of the mass of steel balls within the SAG
mill, with Eq. (4) implying that the consumption rate is 3. SAG mill control
proportional to the mass of balls within mill:
dB In general terms, the benefits sought from improved
Bi BrB 4 primary grinding circuit control include increasing through-
dt
put at the desired grind size, reduction of grinding ball
The remaining modelling equations comprise mass balance consumption and minimisation of the SAG mill power
relationships (which may be regarded as pseudo steady-state requirement. In the absence of an adequate control system,
because of their fast, relative to the SAG mill, dynamics) any disturbance to the process will be reflected in changes to
together with a number of empirical relationships describing the solids hold-up within the SAG mill, changes in the
the mill discharge flow rate and the efficiencies (i.e. particle circulating loads, and eventually changes in the size distri-
classification as a function of size) for the vibrating screens bution of the product sent on to the flotation circuit.
and cyclones. These equations are well described in the In the plant under consideration, the major constraint on
literature [1]; the only ones of relevance here being those maximising the feed rate to the plant is the limitation on the
relating to the power requirement of the SAG mill.
The empirical relationship giving the gross power drawn
by the SAG mill is
P 25:9WE WL 2064:5 5
X
n
WL si B 7
i1
SAG mill draw power. If the feed size distribution were to (here 61.5 mm), P is the measured SAG mill power draw
increase and/or the feed were to became harder, then more (kW), Psp is the power draw set-point (kW), f1 is the fresh
solid material would be held up in the SAG mill, which ore feed rate, (t/h), while eP and e80 are errors (or deviations)
would draw more power. As a result, this study concentrated from their reference values.
on designing a controller whose objective was to keep the Any process control book, such as Ref. [4], describes
SAG mill power at a set value (somewhat below the straightforward procedures for designing conventional feed-
maximum permissible) by manipulating the ore fresh feed back and feedforward controllers. Similarly, any modern
rate. distributed control scheme (DCS) has configurable versions
The simplest controller would employ feedback (FB) of such controllers that merely require the appropriate
only control by adjusting the SAG mill feed rate to maintain tuning parameters to be specified. However, the FB FF
the SAG mill power draw at its set-point value. As with any controllers generally available within a DCS are non-
feedback controller, such a system can only take action after robust in that they can provide no guarantees with regard
a disturbance has affected the process and caused a meas- to loop performance and/or loop stability. If performance
urable deviation from the set-point. A simple remedy here is and/or stability is inadequate, then the usual option is simply
to include a feedforward (FF) control loop that also adjusts to resort to empirical detuning of the control scheme. Since
the SAG mill power draw, but this time on the basis of a the 1980s, however, modern control theory has made rapid
measurement of the disturbance(s) affecting the system. advances in terms of the development of techniques for
A feedback feedforward (FB FF) control system designing robust control schemes that can make powerful
should reduce the effects of disturbances propagated down- guarantees with respect to control loop performance and
stream through the process (here to the flotation circuit). To stability [5,6]. In Appendix A, one such robust design
reject disturbances, however, such a control scheme must be technique will be presented and used to design an FB FF
able to detect them with reasonable accuracy. In this case, controller that meets all user-specified performance and
the main disturbances are associated with fluctuations in the stability requirements.
feed ore physical properties (e.g. the hardness and size Fig. 3 compares the regulatory response of two robust
distribution). Hardness and size distribution are known to SAG mill controllers to a number of step changes in the
be correlatedin that a coarse feed is generally the result of disturbance variable F80. At t=5 h, F80 was stepped up from
a hard ore, and vice versa. Thus, the only disturbance 61.5 to 63.3 mm (i.e. the feed particle size distribution was
considered in this study was the feed size distribution (as increased), while at t=36 h, F80 was stepped down from 63.3
measured by the F80 parameter) for which a number of to 57.7 mm (i.e. the feed particle size distribution was
measurement devices are already available on the market. decreased).
The FB FF controller architecture is shown in Fig. 2, The PI response in Fig. 3 is for the case where a
where F80 is the disturbance measurement (80% passing proportional integral controller (as widely used in industry)
size in mm), F80,nom is the nominal value of the disturbance was employed. However, here, the PI controller was not
requirement imposes magnitude constraints on the sensitiv- inherent trade-off involved in the solution of the robustness
ity function problem, as since S (s)+Ty (s)=I, then S (s) and Ty (s) cannot
1 both be made small simultaneously.
Ss A:1
1 gp sk1 s
A.1. Hlcontroller design for SAG mill
to be small for all frequencies where the disturbances and/or
set-point signals are large. The first step in the design of a robust controller for the
The last design requirement, robustness, asks for both SAG mill involves obtaining linear dynamic models
stability and performance to be maintained not only for a (expressed as transfer functions) for the impact of the
nominal plant model, but also for a set of neighbouring plant manipulated variable (i.e. feed flowrate) and the disturbance
models that result from the unavoidable presence of mod- (i.e. the F80 parameter) on the variable to be controlled (i.e.
elling errors (generally referred to as model uncertainties). the SAG mill power draw).
This robust requirement imposes other magnitude con- ys gy sus gd sds A:9
straints on various system transfer functions. One com-
monly used description of model uncertainties is the so- These transfer functions were obtained by putting step
called unstructured multiplicative uncertainty changes (in feed flowrate or F80) into the nonlinear dynamic
SAG mill model and fitting the responses obtained to low-
gp s 1 wy sDsgp s A:2
order linear models. In both cases, a simple first-order model
where gp (s) represents the nominal plant model, gp (s) is the (with no appreciable dead-time) gave an acceptable fit.
uncertain plant model, wy (s) is a stable minimum-phase 2:93
system and D(s) is any stable perturbation (or model gp s A:10
0:21s 1
uncertainty) whose transfer function satisfies
25:21
rmax D jxV1; bx A:3 gd s A:11
0:14s 1
Stability for this uncertain system requires that the comple- As outlined above, the robust design problem involves the
mentary output sensitivity functions synthesis of a controller k1 (s) that both stabilises the nominal
plant and shapes a number of magnitude functions such that
gp sk1 s
Ty s A:4 all performance specifications are met. This shaping proc-
1 gp sk1 s ess is a challenge as (i) the stability and performance
requirements are, to some extent, contradictory, and (ii) the
gp sk2 s gd s shaping process is carried out in the frequency domain.
Td s A:5
1 gp sk1 s Effective techniques are, however, available using opti-
misation tools and Hl methods to design robust controllers
be small for all frequencies where the uncertainties are large. that minimise the weighted mixed sensitivity criterion [7]:
In an analogous way, performance for this uncertain
system requires that the sensitivity transfer function S (s) min c A:12
k1 s;k2 s
of the closed-loop system be sufficiently small (in a
weighted Hl-sense) for the worst case plant: rmax y jx rmax ed jx rmax ey jx
while the robust performance objective reduces to [5] us wu sk2 sds k1 sSsrs A:16
rmax wy jxTy jx < 1; bx A:8 rmax d jxV1; rmax r jxV1 A:17
Condition (A.7) establishes the robust stability of the feed- Note that an actuator weighting wu (s) is generally included
back control loop while condition (A.8) demonstrates the in Eq. (A.12) (along with weighting functions for the error
270 O. Galan et al. / Powder Technology 124 (2002) 264271
we,y and control loop output wy (s), as this can often play a Table 1
Parameters for robust SAG mill controller
significant role in shaping the closed-loop performance and
robustness behaviour. i ai(1) ai(2) bi
The shaping process referred to previously involves 0 1 0 1
determining the three weighting functions in Eq. (A.11) so 1 5.54103 2.96103 997.79
2 2.16105 1.14105 3.62104
as to ensure both controller stability and performance.
3 2.19106 1.15106 3.16105
Although guidelines exist, this procedure generally requires 4 6.39106 3.16106 6.18105
a number of iterations until a solution is found satisfying all 5 1.39106 1.58104 2.78103
requirements. 6 3.13103 19.56 3.14
For the SAG mill case, the following weighting functions
were found to be satisfactory (Fig. 6).
prising a 6th-order numerator and denominator (the
4:4s 1
3 controller parameters are given in Table 1). However, such
we;y s 10 A:18
440:8s 1 controllers are simple to program within a modern DCS, and
as such are becoming increasingly commonplace.
4:4s 1
we;d s 10 A:19 One of the nice features of the robust design procedure
440:8s 1 is that it is a straightforward matter to consider using lower
3:6s 1 order controllers [8]. In some cases, this will be infeasible as
wy s 0:05 A:20 reducing the controller complexity generally reduces its
0:11s 1
flexibility in terms of meeting all stability and performance
28:6s 1 requirements. However, in the SAG mill case, it proved
wu s 103 A:21
0:04s 1 possible to design a simple FB FF controller where each
component had the form of a low-order lead lag transfer
Using the process transfer functions together with the
function:
weighting functions above, the following robust controller
was determined for the SAG mill. Note that this controller s1;i s 1
kreduced;i s Ki A:24
has been split into two to provide feedback (i.e. k1 (s)) and s2;i s 1s3;i s 1
feedforward (i.e. k2 (s)) components:
2 3 This robust controller would be particularly easy to imple-
k1 s ment within a DCS. The controller parameters for this
Ks 4 5 A:22 reduced order controller are given in Table 2.
k2 s
X
N
j Appendix B. IMC PI controller design
ai sN i
i0
kj s j 1; 2 A:23 A typical PI controller has the following transfer func-
X N
N i tion:
bi s
i0
Kc
us uss Kc es B:1
This robust version bears little resemblance to a conven- sI s
tional [4] FB FF controller, with each component com-
For a first-order plant given by:
K
gp s B:2
ss 1
The PI tuning parameters are as follows [5]:
s
Kc B:3
kK
Kc 1
B:4
sI kK
Table 2
Parameters for (reduced order) robust SAG mill controller
i Ki s1,i s2,i s3,i
1 982.70 2.58 0.0011 434.78
Fig. 6. Hl weighting functions. 2 8.15 166.66 0.0011 434.78
O. Galan et al. / Powder Technology 124 (2002) 264271 271
For instance, when K=2.93, s=0.21 and k=0.1, it results in [2] W.J. Whitten, A matrix theory of comminution machines, Chem. Eng.
Sci. 29 (1974) 589 599.
Kc=0.7161 and Kc/sI =3.4130/sI =3.4130. The steady-state
[3] S. Morrel, Power draw of wet tumbling mills and its relationship to
value for the input is uss=250 t/h. change dynamicsParts 1 and 2, Trans. Inst. Min. Metall. 105 (1996)
C54 C62, January April.
[4] G. Stephanopoulos, Chemical Process Control, Prentice-Hall, 1984.
[5] M. Morari, E. Zafiriou, Robust Process Control, Prentice-Hall, 1989.
[6] J.C. Doyle, B.A. Francis, A.R. Tannenbaum, Feedback Control Theory,
References Macmillan, 1992.
[7] S. Skogestad, I. Postlethwaite, Multivariable Feedback Control, Wiley,
[1] T.J. Napier-Munn, S. Morrell, R.D. Morrison, T. Kojovic, Mineral 1996.
Comminution CircuitsTheir Operation and Optimisation, Julius [8] K. Zhou, J. Doyle, K. Glover, Robust and Optimal Control, Prentice
Kruttschnitt Mineral Research Centre, Australia, 1996. Hall, 1996.