Manuscript submitted to Control Systems, 98
APPLICATION OF FUZZY LOGIC IN LIME KILN CONTROL Esko Juuso, Timo Ahola, Petri Oksanen and Kauko Leiviskä Control Engineering Laboratory Infotech Oulu and Department of Process Engineering University of Oulu Abstract: This paper shows some examples of fuzzy modelling and control of an industrial lime kiln. The first example concerns with the use of linguistic equations in replacing and developing the control rules of fuzzy controller for the lime kiln flue gas fan. Rule-based deduction is replaced by matrix operations by using integers instead of linguistic levels. The original rule base consists of 22 rules that have been formulated originally based on operator interviews and later tuned manually by trial and error. The second example describes the modelling of burning end of the kiln that was found necessary for designing the fuzzy controller for the burning temperature. The final example shows how the Linguistic Equation Controller controls the burning temperature of the lime kiln. Keywords: Lime kiln, control, modelling, fuzzy logic.
INTRODUCTION Lime kilns are used in paper industry to convert lime mud to lime in the chemical recovery. Because of the large quantities of lime mud handled, it is desirable to employ an effective control scheme for the economical production of good quality lime. The control of the lime kiln is multivariable by nature. The retention time in a rotary kiln is long and varying and there is a large quantity of lime mud and lime in process within the system at any given time. Classical model based supervisory control of process is difficult, and the experience of veteran operators is often the best solution. Control changes should be made in small increments, allowing sufficient time to observe the effect after each change has been made. Overreaction leads to overcompensation, which generally causes oscillations and upsets which may take hours to bring back to normal operation. Control Engineering Laboratory, University of Oulu has been active in lime kiln modelling and control since the very early modelling studies in 70's [1,2] that led to a commercial computerised control system for the lime kiln [3]. Lime kiln models were also included in a steady-state flowsheet simulation package [4]. Later, sensor based applications [5] and an expert system application for lime kiln control were reported [6]. In the beginning of 90's, good results were also gained by using fuzzy logic control in an actual rotary lime kiln [7]. Fuzzy control of rotary kilns is not a new area; first applications at both cement and lime kilns started already at the end 70's [8]. References [9-13] show also other applications. This paper introduces several applications of modelling and control of a lime kiln. Applications are based on Linguistic Equations Approach developed in Control Engineering Laboratory [14] and they use data from an industrial scale lime kiln at Wisaforest Pietarsaari mills. The first application models the early fuzzy control system for the flue gas fan [7, 15]. FLC was originally implemented in Alcont-II system. The second example describes the burning end model of the kiln [16]. Linguistic Equation Controller was applied first for controlling the solar power plant [17]. The final example in this paper shows, how it controls the burning temperature of the lime kiln.
LINGUISTIC EQUATION APPROACH Linguistic equations provide a flexible environment for modelling and control of both data intensive and knowledge intensive applications. The model or the controller is represented by linguistic relations, which can be changed into matrix equations. The system is adaptive since the meaning of the linguistic values depends on the working point of the process. This presentation is easily generalised for finer fuzzy partitions and transferred between the programming systems. Membership functions are tuned by simulation experiments, by expert knowledge or by data from real systems. Linguistic equations form a framework for the development of hybrid knowledge-based systems that are needed actually in all practical applications. According to the original framework [14], a set of linguistic rules or relations can be changed into a compact equation. m
SA X j =1
ij
j
=0
(1)
where Xj is the linguistic level for the variable j, j=1…m, i.e. the linguistic values very_low, low, normal, high and very_high are replaced by numbers –2,-1,0,1 and 2. Fuzzy control applications take usually advantage of rule bases as shown in Figure 1. These can be converted to equation form by replacing linguistic levels for error and error derivative by integer values as shown in the Figure. Multipliers Aij {-1,0,1} describe the direction of interaction between variables. Linguistic models can be presented as matrix equation AX = 0
(2)
where A is n x m-matrix.
Figure 1. Rule base for fuzzy PI-controller.
CONTROL OF THE FLUE GAS FAN According to Figure 2 the fuzzy controller for the flue gas fan has seven inputs and one output. The rule base consists of 22 rules that have been formulated originally based on operator interviews and later tuned manually by trial and error [7].
Change in feed end temperature
Feed end temperature Excess oxygen content
Change in excess oxygen Burning temperature
Fuzzy Decisions
Change in fan speed
Change in burning temperature
Temperature after cyclone
Figure 2. Variable definitions for the fuzzy controller of the flue gas fan [18].
The rule base for the above controller has been replaced by five linguistic equations shown in Figure 3. Each equation corresponds to a group of rules describing the effect of two input variables to the control variable (change in the fan speed). In Figure 3 variables are: (1) feed end temperature, (2) deviation in feed end temperature, (3) burning end temperature, (4) deviation in burning end temperature, (5) excess oxygen, (6) deviation in excess oxygen, (7) temperature after cyclone and (8) change in the fan speed).
Figure 3. Fuzzy rule base for the flue gas fan controller in the equation form. Note: black colour shows the variables included in each equation [18].
MODELLING OF THE BURNING TEMPERATURE The modelling of the burning temperature was a part of a larger study, which is aimed to further improve the fuzzy control of the lime kiln. The results were used in designing the burning temperature control for the kiln as shown in the next chapter. After preliminary testing the modelling of the burning end temperature was found out to be necessary. The model was used for off-line testing of control alternatives. Two models were tested. In the first model, the fuel flow was used as the input and the output was the burning temperature. In the second model, the lime flow rate was used as an addi-
tional input variable. Modelling utilised both Linguistic Equations and ANFIS –Adaptive-Network-based Fuzzy Inference Systems. Only the results from Linguistic Equations Approach are shown here. The comparison of these two methods is given elsewhere [16]. The data for the first model is presented in Figure 4. Fuel rate (WI-51) was used as input and burning temperature (TI-162) was used as output. A data set of 563 measurements was selected. The correlation between the fuel rate and the burning temperature in this data set was 0.93.
Figure 4. Training data.
This simple model can be represented by a single linguistic equation with interaction matrix A = [1 -1]. The Linguistic Equations approach generates non-linear membership functions shown in Figure 5. Tuning of the linguistic equation system is fast, and can be easily extended even to more complex systems. The second model uses the fuel and lime rates as inputs and the output is the burning end temperature. For two-input/one-output model a data set of 306 data pairs was selected. In this set the correlation between the fuel rate and burning temperature was 0.90 and between lime rate and burning temperature -0.53. Between fuel and lime rates, correlation was -0.72. In training data (Figure 6) lime feed changes rapidly from 35 t/h to 41 t/h. Otherwise there are no changes. This causes difficulties for modelling. In this case the model can be represented by a single linguistic equation with the interaction matrix A = [-1 3 -3]. Expert knowledge was used to compensate the weaknesses in input data. The feasible ranges generated from the data were modified to make the rapid change lower in lime feed. The effect of lime feed was also changed negative and the rates for fuel feed and burning temperature were increased to reduce the effect of lime feed to the final model. Figure 7 shows the comparison of linguistic equation model for burning temperature with the actual measurements. Figure 8 shows the corresponding model surfaces.
Figure 5. Membership functions from linguistic equations approach [16].
42
55 50 Fuel feed
Lime feed
40 38 36
45 40 35
34
30 0
100 200 300 Experimental Case
400
0
100 200 300 Experimental Case
400
0
100 200 300 Experimental Case
400
Burning end temperature
640 620 600 580 560
Figure 6. The training data set [16].
CONTROL OF THE BURNING TEMPERATURE Dynamic simulators based on linguistic equations are continuously used in development of multilayer linguistic equation controllers, which consists of basic controller, working point controller and a module for unsymmetry handling and braking. The burning end controller is a multilevel linguistic equation controller. This new type of controller was first implemented on a solar collector field in a solar power station at Plataforma Solar de Almeria [16]. In the control design, hybrid techniques combining different modelling methods in a smooth and consistent way are essential for successful comparison of alternative control methods. Switching between different submodels in multiple model approaches should be as smooth as possible. Adaptation to various non-linear multivariable phenomena requires a highly robust technique for modelling and simulation [19].
Figure 7. Comparison of linguistic equations output and training data after training [16].
Figure 8. Model surface of the linguistic equation models for two-input/one-output model [16].
A multilevel linguistic equation controller has been tested with the simulation model of the lime kiln (Figure 9). For constant production rate, set point tracking is fast with very smooth
fuel feed changes. The controller operates also very well for smooth production rate changes. Compensation of disturbances caused by fast production rate changes take more time since the lime kiln process is slow. For safety reasons, the maximum cumulative change of control on predefined time intervals is limited, and this saturation can be seen in the cases of fast production rate changes. The multilevel LE controller is now on on-line testing in an industrial lime kiln, and the experiences are very similar to the simulation results. The application is running in G2-environment connected to the lime kiln automation system. Smooth production rate changes are found to be preferable also in the real process. Tuning of the LE controller with the robust dynamic simulator is needed because the process is slow, e.g. the simulation shown in Figure 9 corresponds to almost three days operation.
Figure 9. Linguistic equation controller testing with the multimodel simulator [19].
CONCLUSIONS Fuzzy logic control has been successfully implemented in the control of an industrial scale lime kiln. This experience is now transferred to a wider control system. Linguistic equation framework is a novel approach, which provides a flexible and compact method for developing and testing complex systems. This method has been shown to operate on the application previously implemented on conventional fuzzy techniques. Extension of the application to the control of the burning end of the lime kiln is reported in this paper. REFERENCES [1] P. Uronen & H. Aurasmaa & K. Leiviskä, Static and dynamic modelling of a lime-circulation loop, Paperi ja Puu 58(11) (1976) 775-780.
[2] P. Uronen & K. Leiviskä & H. Aurasmaa, Simulation study of the lime kiln, Proceedings of Simulation '75, An International Symposium and Short Course, (Zürich, June 23-26 1975), p. 222-227. [3] M. Elsilä & K. Leiviskä & K. Nettamo & T. Pulkkinen, Computer control of causticization and lime kiln area is possible, Pulp & Paper 53(12) (1979) 152-155, 159. [4] E. Jutila & K. Leiviskä & M. Visuri & P. Uronen, Optimizing energy in chemical recovery, PMS/CR. Tappi 65(1) (1982) 43-46. [5] P. Uronen & K. Leiviskä, New topics in lime kiln control, Pulp & Paper Canada 90(9) (1989) 113-117. [6] K. Leiviskä & R. Huttunen & P. Uronen, Expert Systems in the lime kiln control, Proceedings of XIX Conference The Use of Computers in Chemical Engineering (CHEMDATA 88), (Gothenburg, June 13-15 1988), p. 379-383. [7] K. Haataja & J. Ruotsalainen, Fuzzy Logic in Lime Kiln Control, Proceedings of MEPP’94, (Mariehamn, June 13-17, 1994). [8] J.-J. Ostergaard, Fuzzy control of cement kilns, a retrospective summary, EUFIT '93 - First European Congress on Fuzzy and Intelligent Technologies, (Aachen, 7.-10.9.1993). [9] M. Ducay, Lime kiln upgrade at Scott Maritimes slices fuel costs, boosts efficiency, Pulp & Paper 62(11) (1988) 121-122. [10] F. Erens, Process control of a cement kiln with fuzzy logic, EUFIT '93 - First European Congress on Fuzzy and Intelligent Technologies, (Aachen, 7.-10.9.1993). [11] K. Gadeberg & L.P. Holmblad, Automatic kiln start-up by fuzzy control, World Cement 18(6) (1987) 229232. [12] J.P. Kemmerer, Automating the kiln process - updates on fuzzy logic and shell scanners, World Cement 22(6) (1991) 4-9. [13] S.E.Sheridan & P. Skjoth, Automatic kiln control at Oregon Portland Cement Company's Durkee Plant, utilizing fuzzy logic, Ciments, Betons, Platres, Chaux 743-4(1983), 227-231. [14] E.K. Juuso & K. Leiviskä, Adaptive Expert Systems for Metallurgical Processes. In Proceedings of the IFAC Workshop on Expert Systems Mineral and Metal Processing (Espoo, Finland, August 26-28), 119-124, Pergamon, Oxford, UK, 1992. [15] E.K. Juuso, T. Ahola & K. Leiviskä, Fuzzy logic in lime kiln control. Proceedings of TOOLMET '96 Tool Environments and Development Methods for Intelligent Systems, University of Oulu, Finland, pp. 111119. [16] E.K. Juuso, T. Ahola & K. Leiviskä, Fuzzy Modelling of a Rotary Lime Kiln Using Linguistic Equations and Neuro-fuzzy Methods. SICICA’97, 579-584. [17] E.K. Juuso, P. Balsa & K. Leiviskä, Linguistic Equation Controller Applied to a Solar Collector Field, Proceedings of ECC’97, Paper 267, 6 pp. [18] K. Leiviskä, T. Rauma, T. Ahola, E. Juuso, J. Myllyneva & P. Alahuhta, Fuzzy modelling, tuning and control. Report B No. 2, January 1996. University of Oulu, Control Engineering Laboratory. 44 p. (In Finnish). [19] E. K. Juuso, Robust Dynamic Simulation with Linguistic Equations in Intelligent Control Design. In Proceedings of the Eurosim'98 Simulation Congress (Espoo, Finland, April 14-15 1998)}}, volume 2, pp. 324-331. Contact point: Kauko Leiviskä, Control Engineering Laboratory, University of Oulu, P.O. Box 444, FIN-90571 Oulu, Finland, Telephone: +358-81-5532460, Telefax: +358-81-5532466, E-mail:
[email protected].
