IFACMMM 2009. Viña del Mar, Chile, 14 -16 October 2009.

Experiences and Lessons with Advanced Control Systems

for the SAG Mill Control in Minera Los Pelambres

Daniel A. Silva*, Luis A. Tapia**

*Senior Control Eng., Automation Department, Minera Los Pelambres, Choapa, Chile

(4001 Apoquindo Ave., 18th

Floor, Santiago, tel.: +56-2-7983609; e-mail: dsilva@pelambres.cl).

**Concentrator Operations Superintendent, Minera Esperanza,Antofagasta, Chile

(2670 Vitacura Ave., 6th

Floor, Santiago, e-mail: ltapia@mineraesperanza.cl)

Abstract: After some years of applying diverse advanced control systems by using self-developed

automation schemes, vendor expert systems, university-developed systems or a combination of such, a

number of lessons regarding the need for such systems, their areas of applicability as well as their key

performance indicators emerge as general guidelines to be used in future developments in Minera Los

Pelambres. Their main characteristics, capabilities and limitations as well as our experience with them

are also presented. As a conclusion, the main factors that determine success (or failure) of advanced

control strategies are discussed.

The different technological approaches implemented are discussed further and a case study for

optimizing SAG mill operation is also presented.

This paper discusses the needs of advance control technologies and strategies in the mining industry,

including the impact of these technologies on best practices for their implementation, their operation and

long term sustainability.

Keywords: Expert systems, fuzzy logic, MPC, variability.

1. INTRODUCTION

In the year 2002, only after two years of start-up of Minera

Los Pelambres (MLP), several advanced control systems

were first implemented with a varying degree of success.

Then, with the arrival of new technologies and the gained

process knowledge, new control approaches were (and still

are) implemented and tested. As more control solutions

continue to arrive it becomes important to share insights and

experiences to help clarify and improve SAG mill control.

Undoubtedly, using advanced control systems is an

advantage over traditional control, but the general perception

in the mainstream mineral processing industry that using

experts systems is the best alternative (see Bartsch et al.,

2008), might prove to be wrong.

In order to better evaluate the advantages of any one system

and help select the adequate one, several questions must be

addressed first. Some of them are:

- What is required or what objectives are being looked after?

A number of objectives depending on the particular

conditions of each SAG mill and ore characteristics arise.

This could include maximizing tonnage throughput,

decreasing variability, help protect SAG liners thereby

maximizing mill availability or limiting power consumption

due to power constraints or a combination of anyone of them.

- What is the implementation effort and the corresponding

support effort? A successful system depends heavily on the

implementation time (which should be minimal), degree of

complexity of such implementation (regarding programming,

configuration, fine-tuning, etc.) and the long-term

“maintainability” of the resulting control strategy.

- What measured variables are available or needed and are

they reliable over time? Here it is very important to feed the

system with good-quality variables or to implement for the

missing ones. Also the instrumentation technology has to be

reliable, easy to check or calibrate.

- What will be the measuring scale of success and the key

performance indicators (KPI’s) to look for? Even with a

statistical approach, performance evaluation is difficult owing

to many factors such as continuous changes in the ore feed or

the inherent variability of baseline performance capability or

the inability to isolate the effect of one change from other

simultaneous changes which may be occurring in the process.

These indicators could be an economic factor (such as

increase in throughput) or a safety factor (such as liner wear-

out time) or a combination.

Some of the answers to these questions are discussed further

as the different technologies were tested and their results

were recorded and analyzed. Also, for sake of this discussion

a brief overview of such technologies for the SAG mill

control and their main characteristics is presented.

IFACMMM 2009. Viña del Mar, Chile, 14 -16 October 2009.

1.1 SAG mill DCS PID-based regulatory control

The first and default implementation consisted of basic PID

loops that controlled SAG Weight, Tonnage and Noise

control. Limits for the operation of these loops were set by

the operators. As an example, the Weight-Tonnage control

scheme is presented in Fig. 1.

Fig. 1. Basic Weight-Tonnage control scheme.

Their advantage is their simplicity, ease of implementation

and fine-tuning as they are embedded in any PLC or DCS

based control system. They also work well for most controls.

Also the set-points and outputs excursion limits are easily

defined on the pre-made PID blocks available.

When well tuned they perform well for their given role. But

their limitations are evident as:

- In the SAG control there are multiple interactions between

the various variables and they cannot incorporate easily these

other effects or variables.

- They have no predictive capacity, so response could be slow

if the process has slow dynamics. Also, SAG loops have long

delays because of slow feeders response (and non-linear)

which are not well handled by slow tuned PID controllers.

- There is no pre-emptive response to the interaction among

other controlled or manipulated variables and no knowledge

of time-based dynamics, which means poor disturbance

rejection. This usually requires operator intervention.

- Usually, the relationship between the feed and weight is

both variable and non-linear in nature.

Despite these facts, these basic strategies have to be

implemented as they are needed for start-ups or other

abnormal situation or as a backup when advanced control

systems are not available or they are out of service.

1.2 SAG Mill Expert System- based Control

Usually, a control software that includes a combination of

expert rules, fuzzy logic, crisp rules and sometimes neural

nets (so the model can learn the process dynamics and adapt

over time to different feed ore sizes and mineral types) is

called an expert system. Its control scheme is presented in

Fig.2 and the control strategy deployment is shown in Fig.3.

Fig. 2. Basic Expert Weight-Tonnage-Noise control scheme.

Fig. 3. Basic Expert Control strategy.

In order to respond as early as possible to changing

conditions in the SAG, the rate-of-change (ROC) is an

important characteristic of SAG mill controlled variables.

This requires as a pre-requisite that at least, owing to the

noisy nature of the measurements, filters and moving

averages be implemented within the DCS. The quality of

such signal treatment is very important for the overall

performance of any control strategy.

The strength of this approach is particularly useful when a

single manipulated variable (e.g., SAG mill feed) needs to be

manipulated by more controlled variables. It is also relatively

easy to combine with logic functions to make special changes

to controller mode or variables to respond to a particular

event, so full and rich applications can be accomplished.

There is some strength in the ease of implementation for

instrumental problems detection because statistic tools,

history and process knowledge can easily be incorporated to

decide when a sensor has a failure and alternative actions can

be taken.

IFACMMM 2009. Viña del Mar, Chile, 14 -16 October 2009.

Also, it can take in the experience of the operators in the way

a process flow diagram would do. The resulting strategy

(although not the programming) can be easily explained to

non-expert people.

One of the advantages often overlooked is that it standardizes

into one set of rules and instructions the diverse and often

non-unanimous operating strategies and practices. That

alone represents an improvement.

Some of the disadvantages are:

- Programming. It requires considerable effort and skill and

the maintenance of the implemented strategy is limited to the

expert engineers of the application due to its complexity.

Also, it is not usually well understood by process people that

are not familiar with computer programming.

- Tuning of the fuzzy sets. It is not easy and a very time

consuming task and it normally needs to be tuned by an

experienced control engineer.

- Process variability. Mill feed is usually highly variable in

terms of feed size, grade and hardness. These means

developing different set of rules for the different ore types

adding to the complexity of the program.

- Non predictive. It is difficult to program predictive control

so response could and actually is, slow. Again, because of its

very nature of measuring the ROC and not keeping

knowledge of past control movements of the controlled and

manipulated variables, it is more of a reactive controller and

so, slow to respond.

- Complexity. When used in combination with neural

networks it demands extensive use of computing power and

due to its intrinsic complexity it becomes a black box with

hidden dynamics. This calls for extensive training and testing

to insure proper control actions.

- Limited response range. Our experience with neural

networks and optimizing algorithms showed that they do not

handle well large or sudden changes in the ore type or mill

feeds.

1.3 SAG mill Mutivariable Predictive Control (MPC)

MPC as shown in Figure 4 is an adaptive model-based

predictive controller that solves many of the limitations of

previous control algorithms.

Developing such models includes:

- Making step changes on key variables to measure their

effect. This can be avoided by “guessing” responses obtained

through data-mining of previous SAG historical behaviour or

using prior process knowledge.

- Obtain a matrix defining manipulated and controlled

variables. Although a MIMO (multiple-inputs, multiple-

outputs) or MISO (multiple-inputs, single output) models can

be obtained for SAG mill control, it is easier to handle the

last ones to avoid hidden gains or unexpected control

dynamics.

- Testing and fine tuning of the models. This can be

accomplished either off-line (by simulators) or by on-line

testing.

If the models are correctly developed, MPC offers stable

solutions (defined and repeatable), they are well understood

(i.e., in terms of gains, time constants, variable interactions,

etc.) and most important, because the system knows the

dynamics and interactions of variables, it can work in

prediction as well as feedback mode.

Because variability and excursions are reduced, tighter set-

points can be reached and better averages can be reached,

thus allowing for more stable operations conducive to

consistent and higher throughputs.

However, their success will be heavily based on:

- Reliable and on-time validated instrumentation data.

Actions will be taken based on these measurements so they

must be highly accurate and available (e.g., %Fines). These

signals must contain significant information that accurately

represents the process (amplitude and variability) and

adequate filtering must be used.

- Short implementation time. Configuration and tuning of the

models must be easy and straightforward to shorten the

available SAG trial and testing time. Pilot developments or

low-level language programming are to be avoided.

- In-depth understanding of the process and interactions.

This will help improve the model development and will

shorten implementation time.

Fig. 4. SAG Mill MPC strategy diagram.

2. IMPLEMENTATION OF CONTROL STRATEGIES AT

MLP

Some of the highlights on the theory and the deployment of

three control strategies at MLP are discussed further as

examples on the actual use of both expert systems and MPC.

IFACMMM 2009. Viña del Mar, Chile, 14 -16 October 2009.

2.1 Expert System Control Review

Its implementation included the use of a built-in supervisory

control based on expert rules combined with fuzzy logic and

statistics.

Using operator experience the fuzzy sets were defined first

(what is HIGH WEIGHT or FAST SPEED ROC, etc.) and

then tuned.

As shown in Figure 3 several validations, calculation steps

and determination of the actual operational conditions

precede the actual control. Depending on the obtained

conditions they are executed in the following priority:

- Emergency rules. They override any other condition and

have a priority of execution. If two or more simultaneous

emergency conditions are determined, the one that comes

first in the queue will be executed. Once it is solved, the next

one will be attended and solved and so on. This calls for a

careful arrangement of the rules within the program.

- Process rules. If no emergency conditions are present they

are executed next and their purpose is to keep within

acceptable limits the key variables. Again, in case of conflict,

there is precedence and priority within them. For example,

HIGH WEIGHT will be more important than HIGH

SOUND. Another example is given in the following table.

Table 1. Example of SAG Weight Fuzzy Control

- Optimization rules. If none of the preceding are present then

these are executed. They are similar to process rules, but their

actions are much softer as the process is already on or close

to target. Then it is possible to take key variables to optimal

desired values.

Of course, any of these actions must then be de-fuzzified and

transformed into crisp output variables (namely, process set-

points).

2.2 Expert System “Virtual Geologist”

A special fuzzy-based optimization strategy was devised -

called “Virtual Geologist”- which defines a “Grindability

Index” for the incoming ore based on the performance of a

selected group of variables. This can be seen graphically in

Figure 5.

It acts much in the same way as a group of process experts -

based on their particular field of expertise - would classify

the incoming ore to the SAG mill. It basically, recommends

changes to the target set-points that are used to take the

process closer to optimal operation points, thus helping

maximize key performance indicators.

Fig. 5. SAG Virtual Geologist Screen for Grindability Index.

The advantages of this approach are that is not needed to

have a pre-defined optimization function or model and that it

self-regulates towards maximum throughput.

2.3 MPC Review

After the implementation of an Expert System and due to ore

changes, a new step was given trying to find a better control

technology. The chosen and implemented strategy was MPC

and the one selected was derived from a very robust adaptive

model-based predictive control algorithm summarized in the

following paragraphs.

2.3.1 Basic Equations

The selected one is an adaptive model-based predictive

controller that has its origins in the work of Zervos and

Dumont (1988). Use of a state-space model derived from the

Laguerre orthogonal basis functions allows for adaptive

control without the need to know process order or the time

delay in advance. Laguerre basis functions are defined by:

( )( )

( )( )( )ptt

dt

d

i

ptptL

i

i

i

i 2exp!1

exp2

1

1

1

−−

=−

−

−

(1)

iL is the ith

Laguerre function and p is a scaling parameter.

Each Laguerre basis function is a polynomial multiplied by a

decaying exponential; so these basis functions make an

excellent choice for modelling transient behaviour as they are

similar to transient signals. For use in modelling processes,

these functions are written in the form of a dynamic system.

IFACMMM 2009. Viña del Mar, Chile, 14 -16 October 2009.

With appropriate discretization (assuming straight lines

between sampling points), the orthogonality of the basis

functions is preserved in a discrete time model of the form:

kk

kkk

y

u

cx

bAxx

=

+=+1

(2a)

(2b)

u is the input variable, which is adjusted to control the

process variable y . x is the nth

-order Laguerre state. The

subscript k gives the time index in terms of number of

controller update steps. The matrices A and b , of

appropriate dimension, have fixed values derived from the

Laguerre functions. Values of the matrix c are adjusted to

match process model predictions to actual plant behaviour.

The dimension n gives the number of the Laguerre basis

function and the algorithm uses n=15, which has proven to be

enough to model almost any industrial plant behaviour.

Model adaptation is executed in real time following the

recursive algorithm of Salgado et al. (1988).

Using the model equations, the value of the input variable

that brings the process to set point (SP) d steps in the future

can be calculated. This equation is the basic control law:

( )( )bIAc

xIAc

++

−−−=

−

+

1d

k

d

kdkk

ySPu (3)

where I is the identity matrix. As with most predictive

control strategies, implementation involves calculating and

executing the new control output, waiting one controller

update period, and then repeating the calculation.

Equations (2) and (3) give only a basic adaptive model-based

predictive controller. Extensions are needed so that this

controller can be applied to control problems that include

common industrial process complications.

The effects of feed forward variables may be modelled by

building state equations, analogous to (2a), for the feed

forward variables, and including the feed forward states in

the process output equation (2b). It is then straightforward to

re-derive the control law (3) based on the extended process

model equations.

Use of the Laguerre model (2) implicitly assumes self-

regulating process behaviour. If the process exhibits

integrating behaviour, such as SAG mills can exhibit at high

loads, the model must be altered so that the Laguerre state

gives the change in the slope of the process response rather

than simply the process response and a disturbance

component must be added. The feedback control law is then

derived based on the Laguerre modified model.

Finally, MIMO processes may also be modelled by a

Laguerre state-space system. However, in the multivariable

case it is preferable to derive the control law by minimizing a

cost function such as in the generalized predictive control of

Clarke et al. (1987), rather than using the d-steps ahead

approach.

2.3.2 SAG Mill MPC Implementation

As shown in Figure 4, three independent but inter-related

MPC loops were utilized with their selected feed-forward

(FF) and feed-back variables. Quick tests were run to

determine things as dead-times, time-constants and process

gain making the total implementation time from model

development to on-line operation less than three weeks work.

The interface to the DCS system was done using OPC (OLE

for Process Control) technology. Its actual architecture shown

in Figure 6 permits the open integration and simultaneous

interaction of DCS, Expert System and MPC (and others).

OPC Calc

Server

Bailey DCS

OCS

MPC Application PC Expert System

Fig. 6. The integration of systems involved in the SAG mill

control solution.

The MPC module consisted of two parts: one is the MPC

algorithm itself with its control loops and modelling and

trending functions and the other is the OPC calculator module

(a third-party software) which runs “libraries” of equations at

fixed intervals of time and also validates data, sets limits and

performs watchdog functions (indicating that all control loops

and hardware are healthy and communicating). By using the

OPC module for this additional signal processing and

calculation, changes to the program could be made more

quickly, easily and with less risk to plant operation than if

this work was done within the DCS. Also, it helps to

minimize the complexity of the DCS program.

In order for OPC calculator, MPC and Expert modules to

pass information between each other, “alias tags” were

created in the OPC Server serving as Read/Write blocks

shared between the applications.

3. RESULTS

Table 2 summarizes the statistics for the cases with and

without MPC control. During the period tested, mill feed rate

increased with increases in mill power draw, power draw

increased with increases in mill speed, torque, and the total

mill charge volume all decreased with increases in mill speed

(all of which contributed to grinding efficiency at or near the

"sweet spot"). Mill speed was kept within a narrow range and

mill feed rate is controlled to smoothed variation of weight in

the normal manner.

IFACMMM 2009. Viña del Mar, Chile, 14 -16 October 2009.

Table 2. Preliminary Statistical Plant Results

It is also interesting to note that not only the key mean values

increased (or decreased) favourably, but their standard

deviations also decreased. This accounts for the better control

and the consequent reduction in variability.

The power draw proved being a variable to be excursion-

limited or supervised rather than a controlled one. It is an

indirect outcome of manipulating feed, weight, water, speed

and adequate ball charge. As such it can be used as a KPI to

indicate for the effectiveness of the control and/or operational

actions.

Finally, although not measured, it was observed a better

overall performance on the SAG liners life duration, less

overload events and faster start-ups.

4. CONCLUSIONS

It has been proved that reliable and cost-efficient OPC-based

control is feasible if process time constants are long enough

compared to communication delays, thus allowing the model

to absorb them. This was the case for the SAG mill.

Also, based on the results obtained, a combination of expert

systems based on robust MPC proved to be an optimal

solution to handle supervisory as well as advanced predictive

control.

The implemented MPC, with more than fifteen years of

industrial success in other-than-mining areas, verified the

effectiveness of adaptive model-based predictive control

algorithms for SAG mill control. And although, the purpose

of its implementation was not optimizing, but to stabilize the

key process variables by predicting the process behaviour

under the influence of primary and secondary variables, it

brought as a consequence that the process and its operation

were stabilized also. This in turn, indirectly optimizes

operation. In other words, as a result of applying efficient

stabilizing strategies, the process was indirectly optimized

and not because optimization functions or on-line models

were applied.

Also, it has been demonstrated that because of the adequate

and intelligent process-variables control and the resulting

reduction in variability, it brings about an opportunity for

more optimization, such as the case for the virtual geologist

shown.

As such, any advanced control strategy can be best seen as if

its main objective is to reduce variability, both in operational

practices (by means of standardizing) as well as in the key

process variables. Thus, it is not properly an optimizing, but a

stabilizing control. Now, if such control is well developed

and implemented, it will bring as a result, a close-to-optimal-

point control.

Considering the above, it is the author’s proposition (as well

as Jonas, 2008) that regarding SAG mill control the best

results are obtained by the use of Multivariable Predictive

Control only if an easy, affordable, fast-to-implement (no

pilot developments), standardized and underlying-robust

control is available. Moreover, the mining industry can

improve economical results by adopting more and better

almost-off-the-shelf advanced control solutions such as the

one shown and discarding other technologies that deliver

fewer benefits and are less reliable.

Again, the work done suggests that variability is one of the

main factors to look for when a decision on advanced control

technology is to be made. Actual process variability must be

measured and predictions on its possible reduction must be

transformed into performance KPI’s and then they must be

compared against actual results and implementation costs.

Ultimately, as far as the mining industry is concerned; the

chosen solution must prove to be of some economic benefit,

standardized and reliable to be sustainable.

ACKNOWLEDGEMENTS

The authors wish to thank Minera Los Pelambres for

permission to publish this paper. They further wish to

acknowledge the knowledgeable support and commitment of

PhD. Mike Forbes of Andritz Automation and its product

BrainWave® Control. Thanks also to all Concentrator

Operations group at Minera Los Pelambres for their patience

and contributions to the success of this project and for letting

us perform the required but unavoidable tests.

REFERENCES

Bartsch, E., Comeau, G., Hardie, C. (2008). Evolution of

SAG mill process control at the Xstrata Nickel Raglan

operation. Proceedings of the 40th

Annual Meeting of The

Canadian Mineral Processors Conference-2008, 445-

464. Ontario, Canada.

Clarke, D.W., Mohtadi, C., and Tuffs, P.S. (1987).

Generalized predictive control - Part I. The basic

algorithm, Automatica, Vol.23, No.2, 137-148.

Jonas, Robert K. (2008). Advanced control for mineral

processing: better than expert systems. Proceedings of

the 40th

Annual Meeting of The Canadian Mineral

Processors Conference-2008, 421-443. Ontario, Canada.

Salgado. M.E., Goodwin, G.C., and Middleton, R.H. (1988).

Modified least squares algorithm incorporating

exponential resetting and forgetting, Int. J. Control,

Vol.47, No 2, 477-491.

Zervos, C.C., and Dumont, G.A. (1988). Deterministic

adaptive control based on Laguerre series representation.

Int. J. Control, Vol.48, No.6, 2333-2359.