Fuzzy Logic Control Of A Perlite Plant.pdf

Fuzzy logic control of a perlite plant by Hiroshi Asayama

Phil Burton

Jurgen Gerstacker

Mitsui Mining & Smelting Co.

University of Limerick

Germany

Tetsuya Kohno, Shigeki Matsui and Eoghan O’Lionaird MESCO Inc.

This article covers the application of fuzzy control to a perlite plant (similar to a cement plant). It details the development of the fuzzy control from an initial strategy to the one presently in use. Some simulation results are presented which confirm the practical observations that, for this plant, fuzzy control is more robust than the equivalent PID system. Introduction

P

erlite is a white, powdery, chalk-like substance, produced by burning clay at a temperature ranging from 820°C to 900°C. The consistency of perlite varies from very fine grain chalky powder t o chips of 4-5 mm diameter, depending on the type of clay used and the conditions of production (i.e. temperature of the kilns, water content of the clay, speed of rotation of the kilns, speed of feed of the clay etc). Applications include insulation of walls and roofs by formation of boards, partition boards etc. The perlite plant owned by Mitsui Mining 6 Smelting CO Ltd. was designed for minimal manual requirements and maximal automation from the mechanised loading of the clay right through to the robotised bagging, stacking, and storing of the finished product. However, due to the unwieldy characteristics of the process, automatic control of the plant’s four double skinned rotary kilns has not been possible. The plant produces over 60 different types of perlite at present. Depending on the application and requirements of the customers, the perlite is sold in 100 kg bags, 2 tonne sacks, or loaded directly into trucks from vertical 100 tonne holding bins. Fig. 1 shows an outline of the plant. Clay is mechanically fed t o a crusher and preheater. From there it is fed by conveyor to holding bins to

be distributed by a further conveyor system t o the kilns in the required quantity and blend for the specific operation in hand. The clay is first passed to the outer skin of the rotating double skinned kiln where i t is preheated and dried. It is then loaded into the inner skin where it is burned by a mixture of fuel-oil and air and reduced to perlite and waste gases. The perlite is drawn out of the kilns by a vertical draft, separated in

a cyclone, and periodically weighed by a sample scale ‘weight sampler’ (see Fig. 2) as it is transported t o the baggingipacking area. As for cement, perlite is commonly categorised by weight per unit volume (kdlitre) and there is a loosely defined relationship between the temperature at which the clay is burned and the weight of the finished product: generally the higher the temperature at which the clay is burned, the lighter per unit volume is the produced perlite.

Plant characteristics

T

here are a number of problems with implementing a control strategy for this plant. First, the weight sampling is done periodically: at present once every 9 0 s. This means that if an

hoDDer

conveyoi

I’/

kiln

Fig. 1 Plant outline

COMPUTING 13CONTROL ENGINEERING JOURNAL

DECEMBER 1994

293

\

\

B a

i Fig. 2 Periodic weight sampler

adjustment is made t o the fuel input the results of this adjustment cannot be evaluated for at least 90 s. In addition, the temperature response of the kilns is slow: when the fuel input is changed the temperature of the kiln does not change appreciably for a number of minutes, the rate of response depending on the type of clay being burned and on whether the adjustment in control effort is up or down. Thirdly, although the rate of feed from the holding bins and the conveyor speeds are controlled, the feed t o the kilns is subject to occasional and sometimes severe disturbances. This source of disturbance is for now, at least, unavoidable.

Fuzzy control strategymanagement considerations

T

he rate of feed from the holding bins to the conveyors, the conveyor speeds, and the speed of rotation of the kiln are all controlled by PID loops. Our fuzzy control strategy does not include these loops, although in the future they might be used t o add a further dimension of power t o the fuzzy control strategy used. But initially it was decided to exclude these variables and concentrate on the kiln itself, particularly when previous 294

attempts at automatic control of the kilns using PID control had failed. For manual operation of the kiln the temperature is measured at the outlet of the kiln, at the point where the burnt clay (now red glowing matted lumps] falls into a chute from which it is drawn upwards by a draft (see Fig. 1 1. The temperature transmitter communicates with the control panel via a 4-20 mA loop. The weight of the perlite is measured every 90 s en route to the bagging area by load cells on a periodic sampler (see Fig. 2) and the weight information is also transmitted to the control panel via a 4-20 mA loop. The weight signal is used by the operators t o decide on an appropriate kiln temperature which is controlled by a fuel-oil pneumatic control valve, opening or closing the valve as required. In general a known temperature range produces a given grade of perlite. The manual method of control is to adjust the fuel valve to give the desired perlite density.

Rule elicitation

I

nitially we sought the advice of the plant’s operators and engineers t o assist us in drawing up the rules, and observed all three shifts to compile an average

observation. It took about a week in all but observations were spaced at convenient intervals and did not consist of continuous operator monitoring. Feedback from the plant operators indicated that the fuzzy controller should take into account the temperature at the exit of the kiln and the weight as measured by the weight sampler. Temperature is a secondary variable in the process: the required output from the system is a given weight per unit volume and the output is measured accordingly. However, the operators use temperature as an intermediate measure of the state of the process, and hence it was included in our original design, but its exact significance or relationship to weight has proved difficult t o define or establish. Early trials using both weight and temperature information as inputs to the fuzzy controller indicated that under certain circumstances the combination of a temperature rule with a weight rule could result in an erroneous decision by the fuzzy controller. For example if the temperature is high and the weight is low (which occurs because a higher temperature burns the clay more) then the fuel input should be reduced-this is the correct action. But if the temperature is high and the weight is high then the effect of the input measurements cancel each other (high temperature wants t o set fuel low but high weight wants t o set fuel high] and the output of the fuzzy control may give no change in the fuel setting and the process will continue in this unwanted state. Ultimately, the unnecessarily high temperature results in a large reduction in density and gives rise t o a large unwanted undershoot. This type of behaviour is particularly noticeable for small deviations about the desired norm where, for example, a transient increase in weight due t o an excessive amount of feedstock at a time when the temperature is slightly too high can cause an unacceptable amount of undershoot. To circumvent this, the use of temperature as an input variable t o the fuzzy rules was abandoned and only weight, and difference in weight, between samples were used as inputs t o the fuzzy controller. Weight and change in weight between samples were each defined in terms of five membership functions shown in Figs. 3a and 3b, with an output membership function as shown in Fig. 3e. The rule base with 2 5 rules is given in Fig. 4a. The antecedent (weight)

COMPUTING & CONTROL ENGINEERING JOURNAL

DECEMBER 1994

membership function of Fig. 3a has five fuzzy sets with the ZERO (ZO) membership function having its apex at the required weight value. Each different perlite density requires different process characteristics and the membership functions for each required density are changed by loading new membership function values (see below for details on the software facilities) from the main PC. The rules do not change. The membership functions are also modified t o accommodate differences between winter and summer operation of the plant, which arise from changes in the raw material and the response of the kiln. Most of the experience in operating the plant has been gained with weight and change in weight as the only inputs t o the fuzzy controller as shown in Fig. 5 and the results have been excellent. However recognising that the manual operators appear t o gain additional information from temperature, some experiments have recently been carried out with temperature and change of temperature as additional input variables. For this latter case the two pairs of inputs are treated separately for the input fuzzification and inference as shown in Fig. 6. The outputs of the inference sections are combined at defuzzification with the temperature fuzzy variables being weighted by a factor of 0.1 prior t o defuzzification. The objective of the weighting is t o allow temperature t o have some small influence over the overall performance, whilst at the same time minimising the conflicts which arose when weight and temperature had equal weighting. The results from these experiments are inconclusive.

set point

I

NS

NB

b

PS

PB

-,i:--::::::;:: -1

-1.5

'0

2

1

set paint

o 1

Hardware and software development environment

'

'

860

840

NB

NS

-8

-4

NB

NM

880

I 20

PS

PB

4

8

PM

PB

T

he control was realised using a Fuji Electric Co. Micrex F-250 series fuzzy controller which is based on a 16 bit processor (Fig. 7). It is essentially a programmable logic controller with an add-on fuzzy control facility. The application program is defined using conventional PLC ladder diagrams with the fuzzy inference engine accessed through a subroutine (function) call in the ladder program. A PC is used for program development and the code is downloaded to the controller via a dedicated serial interface bus (T-link). Development software runs under Windows on the PC and comprises a PLC toolkit (shell),

-1

NS

I

'0

PS

1

Fig. 3 Membership functions: (a) Weight; (b)Change in weight; [ c ) Temperature; (d)Change in temperature; and (el Output

COMPUTING & CONTROL ENGINEERING JOURNAL

DECEMBER 1994

295

I

temperature

weight

IL,,1 -

temper- ZO PM PS ZE NS NM

in

NB

IPM IPM

j PM I PS I ZE

b

a

Fig. 4 Rule bases for fuzzy controller: (a] Weight rules; and (b) Temperature rules

designated ES50, and a fuzzy logic development tool given the acronym FRUITAX [Fuzzy Rule Information processing Tool of Advanced control]. FRUITAX facilitates the definition of both IF-THEN statements and membership functions. The membership functions are defined by specifying the three co-ordinates of the triangular membership function. Inference is via the MAX-MIN procedure and defuzzification is via the centroid method. Online changes of rules and membership functions are possible; the application program can also be accessed online, but changes may only be made t o numerical values, more detailed changes must be carried out offline. The software also provides an online display of the degree t o which the rules are firing; this feature has proved very helpful when tuning the membership functions. An incremental type fuzzy controller is used in this application (the software allows both positional and incremental) to provide the required integral action and facilitate

'bumpless transfer' between manual and automatic control. The usual procedure is t o bring the plant up under manual control and then switch to automatic control. Testing of the PLC based fuzzy controller can be done in two ways depending on whether the PLC is connected t o the PC. If the PLC is not connected, performance of the rules may be tested from the keyboard or from a set of data files. When the PLC is connected test signals may be accessed from the 110 modules t o test the rules. An application program in the PLC takes care of N D and DIA conversion, calculating the change in weight between sampling periods, and scaling the values for use with the fuzzy logic procedure. The fuzzy control program runs as a procedure within the main PLC program. The PLC uses 8 bit N D and D/A resolution with 32 bits of digital 110. The PLC uses the fuzzy controller t o calculate a new output to the fuel control valve once every 90 s, corresponding t o the period of the weight sampler.

Results

R

esults t o date indicate that the fuzzy controller using only weight information gives significantly improved performance compared with manual control. During a closely monitored four-week period of operation the fuzzy control system gave an 8% increase in production with a 4% reduction in fuel consumption compared to manual operation over a similar period. In all the plant has been running for two years with the simple fuzzy control based purely on weight information and the performance has been excellent. One of the difficulties created by this success is that the plant operators are reluctant to indulge in more experimentation. Recent modifications t o reintroduce temperature as a minor control variable, as detailed above, were inconclusive. Simple simulations have indicated that this approach should give improved performance, but since it would require several additional weeks of testing and experimentation t o prove success or otherwise it has been decided t o abandon the work for the time being.

Plant simulation

P

weight change in weight

tuzzy

Fig. 5 Fuzzy control using weight only 296

rior t o using fuzzy control, several trials were conducted with a view t o applying PID control t o the perlite kilns. However, this proved impracticable and until late 1992 the plant was operated manually, The attempts at using PID control were not made by the present authors and it is not clear why PID control failed. But since this is a commercial plant there is a very limited time available for experimentation and it is quite possible that there were difficulties in tuning the PID controller and insufficient time t o resolve these problems. Furthermore the plant was COMPUTING 6( CONTROL ENGINEERINGJOURNAL

DECEMBER 1994

-

-

~~

fuzzification and

Fig. 6

c

c

Fuzzy control using temperature and weight feedback

new at the time and there would liave been a lot of pressure t o get the process online under manual control regardless of any automatic control problems. Clearly there is some interest in determining why fuzzy control succeeded when PID control apparently failed and some simple simulations have been carried out with a view t o examining the merits of the two approaches. In order to estimate the plant transfer function a 10% change in the fuel valve setting was applied and the temperature and weight monitored at the output: these are only measurements that can be made on the plant. Surprisingly, following a step change in fuel supply, the output weight changes before the temperature and the explanation for this is not clear. Nevertheless using the information obtained from the step response of the plant and other general observations the approximate model given in Fig. 8 has been derived and used for subsequent simulation work with the full knowledge that i t is not a complete description of the process. The model suggests that the weight of the product and the kiln temperature are independent variables, but this is known to be incorrect because higher temperatures produce lower weight. Fig. 9 a shows the simulated response of the system using PID control for a step input with a change of load after 40 minutes, and Fig. 9b (6.10) shows the response when the plant delay is doubledthe change of load in the second case takes place after 60 minutes. In order t o mimic the real plant the PID settings for Fig. 9a were determined using the Zeigler-Nichols continuous cycling method and these settings

were retained for Fig. 9b. Figs. 9c and d give the system response using the weight only fuzzy controller, as used on the real plant, for the same two sets of conditions. It can be seen from these results that with increased process delay the PID controller Dlant comes close to instability while the fuzzy controller retains reasonably good control.

The standard Zeigler-Nichols procedure for tuning PID controllers gives a phase margin of about 30" which is quite small, and it is not surprising that the system tends to instability with a very large change in the plant delay. The fuzzy controllers used in the simulations. emdoved exactly the membership functions and rules as the ones used on the

Fig. 7 System interfaces

I exp(-80s)

temperature

Fig. 8 Approximate plant model used for simulation purposes

COMPUTING 6 CONTROL ENGINEERING JOURNAL

DECEMBER 1994

297

880 87J

0

865

$

855

2

-effort

\

845 840 835 asn 0

-effort -output

880 8751

-temperature

~711

"U"

20

60 time, min a

40

100

80

120

0

20

40

60

80

100

120

80

100

120

lime, min b

temperature 8551

S! 840

n

840

835

8300 835

change of load

~~~

20

40

60

83

0

10

20

30

40

50

60

70

80

lime, min

time, min

C

d

Fig. 9 Simulation results: (a] PID control using weight: [b) PID control when plant delay doubled; (c) Fuzzy control using weight; and [d) Fuzzy control when plant delay doubled

real plant, and they exhibit relatively little sensitivity t o changes in the plant delay. It is believed that the sensitivity of the PID controller t o changes in the plant characteristics could be the reason why it proved difficult t o get this system working and, similarly, the insensitivity of the fuzzy controller explains its success. Similar responses t o those of Figs. 9c and d are obtained by simulation for combined temperature and weight fuzzy control, as shown in Fig. 6, but with slightly less overshoot when the plant delay is changed.

Discussion

T

his exercise has confirmed that fuzzy control is a viable alternative to conventional forms of automatic control and gives good results in practice. It is tempting t o conclude that fuzzy control worked where conventional control (PID) did not. However, it is not fully clear why the PID control did not work although simulations suggest that PID control is more sensitive t o changes in plant characteristics than fuzzy control. The standard set of Zeigler-Nichols rules which were used in the simulation give a relatively small phase margin (30"]and perhaps a different PID tuning may have given a more robust system. 298

Nevertheless fuzzy control proved easy t o use and did not entail the difficult tuning process that many PID controllers require. Good control was achieved with a plant having significant process delay, and the relative ease with which the control was imdemented is a eood reason for preierring fuzzy con'trol t o more conventional techniques. More work remains t o be done t o tune the membership functions and achieve even better performance, but there is relatively little time available for experimentation. The availability of a fuzzy facility within a PLC, which after all is intended t o be an 'easy to use' controller, indicates that fuzzy control has matured significantly and may reasonably be considered to be an extension of the ladder diagrams widely used in PLC programming. It is interesting that initial feedback from the plant operators indicated that two process variables should be monitored, whereas in practice the single variable weight proved adequate. With hindsight this is not surprising but it does indicate that fuzzy rules need to be carefully formulated and full consideration given t o all potential operating circumstances. There are many examples in real life where operators like t o monitor intermediate variables as well as the final output, perhaps because it gives them a

'feel' for the state of the processbut these intermediate variables may not be necessary and can, in some circumstances, lead to unwanted effects. The ability to easily change membership functions and observe online the degree t o which they are firing is a powerful feature of this installation because it gives a simple method of fine tuning the control mechanism t o allow for changes in system parameters. Clearly there is a need for an adaptive form of fuzzy control which can modify membership functions in line with process trends, but meanwhile this installation has succeeded in bringing automatic control to a process that, in the past, has been difficult t o automate. 0 IEE: 1994 Hiroshi Asayama is with the Perlite Division of Mitsui Mining & Smelting Co. Ltd., 2-1-1 Muromachi, Nihonbashi, Chuo-ku, Tokyo 103, Japan: Phil Button is with the Electronic and Computer EngineeringDepartment, University of Limerick, Limerick, Ireland: Jiirgen Gerstacker may be contacted at Krottensee 51, 91284, Neuhaus, Germany: and Tetsuya Kohno, Shigeki Matsui and Eoghan O'Lionaird are with MESCO Inc., EngineeringDivision of Mitsui Mining and Smelting Co. Ltd., 1352-1 Haraichi, Ageo-shi, Saitama 362, Japan.

COMPUTING 6( CONTROL ENGINEERING JOURNAL

DECEMBER 1994