You are on page 1of 11

Modelling and Control with Linguistic Equations

Esko K. Juuso Control Engineering Laboratory, Infotech Oulu and Department of Process Engineering University of Oulu, P.O.Box 4300, FIN-90410 Oulu, Finland Phone: +358-8-5532463, Fax: +358-8-5532466 email:
ABSTRACT: Linguistic equations are useful for taking into account non-linearity especially in multivariable application systems. The main benefit is better and more accurate decision-making due to the model-based approach and systematic knowledge management. Insight to the process dynamic operation is the most important issue. Automatic generation of systems, model-based techniques and adaptation techniques are very valuable in developing and tuning the controllers. The linguistic equation approach increases the performance by combining various control objectives in a case-based approach. The LE approach is also very efficient as a modelling technique: models can be generated from data, various types of fuzzy rule-based models can be represented by LE models, and any LE model can be transformed to fuzzy rule-based models. The approach is successfully extended to dynamic modelling. KEYWORDS: linguistic equations, fuzzy control, fuzzy modelling, case-based reasoning, decision making, dynamic simulation, adaptive systems, nonlinear systems, multivariable systems

The process industries face considerable control challenges, especially in the consistent production of high quality products, more efficient use of energy and raw materials, and stable operation on different conditions. The processes are nonlinear, complex, multivariable and highly interactive. Usually, the important quality variables can be estimated only from other measured variables. Constraints, e.g. physical limitations of actuators must be taken into account. Significant interactions between process variables cause interactions between the controllers. Various time-delays depend strongly on operating conditions and can dramatically limit the performance and even destabilise the closed loop system. Uncertainty is an unavoidable part of the process control in real world applications. For the overall production processes, the control systems take care of several subprocesses. For example the overall performance of paper machines can be rated to the operating efficiency, which is the percent of time the sheet is on the reel divided by the time machine is ready to make paper. Operating efficiency varies depending on breaks and the duration of the operations caused by the web break. The web breaks account for 2 7 percent of the total production lost. The reason for breaks can be classified to process related, mechanics related and automation related breaks. Most of the indistinct breaks are due to dynamical changes in the chemical process conditions. The process related breaks are especially important in the wet end of the paper machine. This part of the process is continuous, nonlinear and there are many and long time-varying delays. There are also process feedbacks at several levels, closed control loops, factors that exist and cannot be measured and interactions between physical and chemical factors. In the electronics manufacturing industry, a typical process consists of paste printing, component placing, reflow soldering, solder joint inspection and testing. The process is very complicated because many parameters in different process phases affect the whole process. The pattern of relationships between process parameters and quality of final products is many-to-many relationship. It means that one defect could have several root causes originating from different process phases; and one problem in a process phase might result in many different types of defects. Wellestablished solutions for process control are not available in electronics manufacturing industry. Fast growing production in rapidly changing environment has many challenges for decision-making, testing, inspection and control. Material purchasing is a complicated task in rapidly changing operating conditions. Functional testing requires a lot of expert knowledge since the relationships between defects and process conditions are very complex. An appropriate level for visual inspection automation should be found to improve the process control. By combining and comparing the information collected in the production process it is possible to find reasons for some defects, and use this information in process development. Since the products may also consist of several modules manufactured on different sites, the problems of testing and inspection should be handled also for extended enterprises.

Pricing and production scheduling are universal for every industrial area. The pricing should have an appropriate connection to the real production process in order to achieve efficient production also with some last moment changes in the production plan since the actual production cost depends on process conditions. The production scheduling finally defines how well this can be achieved in real production. On the other hand, production planning and scheduling will not work if the lower level control is not operative on a reasonable level. The scheduling and pricing problems become even more challenging when the production is divided into several sites. Different things are emphasised in the process industry and in the electronic manufacturing industry, but the process control needs to be more and more adaptive in both industrial areas. The production control should utilise the degrees of freedom in the process control to be able to reschedule the production on an appropriate way in changing operating conditions. All the time various possibilities are open to those who can utilise them. Fuzzy set systems can improve these systems in many ways. Linguistic equations provide a new novel technique for combining expertise and data to make the overall production control easier. The technique has been developed in process control and modelling, where it extends the possibilities of fuzzy set systems. Because of very compact implementation, these systems can be extended to both case-based and dynamic applications.


The demands cannot be met by traditional control techniques only, and therefore various methodologies have been developed in order to extend the applicability of the control systems. Since PID algorithms cannot adequately compensate numerous interactive disturbances, these problems are sometimes solved by the introduction of various nonlinear PID modifications where the controller parameters depend on the error or on the parameter scheduling [Juuso 1999]. Most industrial applications of fuzzy control are based on using expert knowledge. Fuzzy modelling can increase the performance of controllers extensively. The adaptive fuzzy control is also mainly related with model-based techniques. The linguistic equation approach extends the performance of the fuzzy control in many ways. The number of the fuzzy control applications has increased very fast, especially during the 90's. A large number of highly successful fuzzy control applications are implemented in process industry. A fuzzy logic control system can be interpreted as a real-time expert system combining symbolic if-then rules with qualitative, fuzzy variables linked to the real world by membership functions. Conventional expert systems can be represented by fuzzy expert systems with much less rules if nonlinearities are taken into account in membership functions. Therefore, complicated expert control strategies can be included to practical control. Traditionally, the rule base of the fuzzy controller is constructed from expert knowledge by trial-and-error, and most applications are still based on this approach. Various methodologies have been developed for automatic generation of fuzzy systems and inverting fuzzy models, i.e. fuzzy systems can be regarded as flexible mathematical approximation tools. However, the fuzzy controllers should be kept transparent to interpretation and analysis on the basis of understanding the process behaviour since this insight is more important than close correlation to experimental data.

Figure 1: Examples of linguistic equation controllers

Many fuzzy controllers are implemented as two dimensional fuzzy logic decision tables where control actions and input conditions are expressed in terms of membership functions. The rule base is essentially linear, and the nonlinearity can be introduced by adjusting membership functions or by modifying rules. The nonlinear control surface shown in Figure 1 is based on trapezoidal membership functions. Modified rule bases are used in multivariable controllers, especially for feedforward purposes. Some experiments with automatic control based on deterministic decision tables were carried out already before fuzzy control. For a cement plant, these decision tables were inspired by instructions found in a textbook for kiln operators, which contained the control rules for manual operation. First experiments with real cement plants were done since 1972 in Denmark [Juuso 1999]. Linguistic equations have been used in replacing and developing control rules for FLCs [Juuso 1996]. This approach can easily be combined with other approaches for FLC design, e.g. to a conventional way of acquiring knowledge from experts by interviews. The traditional fuzzy systems described by if-then rules are represented by matrix equations if nonlinearities are handled by membership functions [Juuso 1992]. The original approach has been extended to fuzzy and real valued linguistic equations with variable specific partition levels [Juuso 1996]. For a symmetrical fuzzy PI type controller, all the rules can be obtained from a single linguistic equation u = e + e , which is a special case of the matrix equation A X + B = 0 , where the interaction matrix A = [1 1 1], the bias term B = 0, and variables X = e e u . In the rule base, the linguistic values {NB, NS, ZO, PS, PB} are replaced by integer numbers -2, -1, 0, 1 and 2. The control surface of the fuzzy PI controller can be obtained by using the integer valued linguistic equation together with trapezoidal membership functions. Alternatively, fuzzy rule bases can be represented by real valued relations [Juuso 1999]. The LE controller is also applicable on any fuzzy partition, and the procedure produces always a rule set which is complete, consistent, and continuous. If a non-complete set is satisfactory, a part of the rules can be deleted already before tuning. The best similarity with the LE controller is achieved if the positions of all the fuzzy sets are coordinated with the real valued linguistic equations, i.e. linguistic values are replaced by real numbers. The real valued LE controller provides a very smoothly changing control surface (Figure 2). Feedforward controllers based on steady state models can have interactions with different strengths. A fairly smooth control surface is also obtained by the fuzzy controller with seven consequent fuzzy numbers described by triangular membership functions. Linguistic equation controllers implemented on the basis of real valued equations need only the membership definitions. Nonlinearity is introduced to the system by membership definitions, which correspond to membership functions used in fuzzy logic controllers. This is the main difference to the fuzzy logic controllers where the nonlinearity is handled through the rule base. Operation of the controller is modified by variables describing operating conditions. Values for the working point variables are obtained by a moving average on a selected time range, and the working point is defined by a working point equation. For large disturbances, this controller implementation can lead to a large overshoot, which is problematic if small control actions are required, e.g. in oscillation situations. This problem is removed by taking the original error into account. The braking action is implemented by a correction, which increases the importance of the change of error e when the controlled variable goes towards the setpoint, i.e. the LE controller tries to slower the speed of error compensation already before the error goes to zero. The braking action is automatically activated when the error becomes large enough, and the system returns to the normal operation when the fast trend is removed. In changing operating conditions, some trends may push the controlled variable causing a slight offset, which is removed very slowly if the control surface is symmetric. Since different asymmetry is needed in different working conditions, the asymmetrical correction is defined by the trend variable. If several trends effect to the offset, an asymmetry equation can be introduced. This action is not used simultaneously with the braking action [Juuso 1999]. This implementation is very compact if compared with model-based predictive controllers, which are an alternative solution for these situations. The overall operation corresponds to a three level cascaded controller: Basic PI type LE controller handles the normal operation with symmetrical membership definitions; Operation condition controller changes the control surface of the basic LE controller by modifying the membership definitions for the change of control variable; Predictive LE controller changes the membership definitions for the derivative of the error. This level contains both the braking and asymmetrical actions.



Fuzzy modelling is an extension of the expert system techniques to uncertain and vague systems. Fuzzy set systems continue the traditions of physical modelling on the basis of understanding the system behaviour. Fuzzy rules and membership functions can represent gradually changing nonlinear mappings together with abrupt changes. Fuzzy models can help in extracting expert knowledge on an appropriate level [Juuso 1996]. Fuzzy models can also be constructed from data, which alleviates the knowledge acquisition problem. Various techniques have been used to fit the data with the best possible accuracy, but in most cases the interpretation of results is not addressed sufficiently. Fuzzy models can also be considered as a class of local modelling approaches, which attemps to solve a complex modelling problem by decomposing into number of simpler subproblems [Babuska et al. 1997]. All these models can approximate both static and dynamic nonlinear systems. There are several alternatives for representation of the system's dynamics. Prior knowledge can used in constructing rule-based fuzzy models: qualitative knowledge can be incorporated in linguistic fuzzy models, or for fuzzy relational models if there are several alternative rules; locally valid linear models can be collected by Takagi-Sugeno fuzzy models. Data-driven fuzzy modelling can be based on various methodologies, e.g. fuzzy clustering, self-organizing maps, neurofuzzy methods and linguistic equations. Different approaches can be combined in the tuning phase, especially the linguistic equation approach is designed for combining different sources of information. Each linguistic equation represents a multivariable interaction: the directions and strengths of interactions are defined by coefficients of the interaction matrix. Only the variables with nonzero coefficient belong to the interaction. Nonlinearities are taken into account by membership definitions which consists of two polynomial functions, one for positive and one for negative side labels. Example of membership definitions are shown in Figure 2. With these definitions the values of input variables are mapped to the range -2...2 which correspond to the labels very low ... very high (or negative big ... positive big); normal (or zero) is always zero. After the matrix calculations, the outputs are mapped from linguistic level to the real scale. Since only five parameters are needed for each variable, the LE systems can be adapted to various working conditions. Usually, fuzziness is taken into account by membership definitions - linguistic equations approach does not necessarily need any uncertainty or fuzziness. However, also the linguistic equations can be used in fuzzy form [Juuso 1996], i.e. all the variables are not included to the model. The fact that experts do not always agree with interactions can also be taken into account by using several interaction matrices with different coefficient values. On the other hand, the directions of interactions can depend on the working area in nonlinear systems. In these cases, different interaction matrices have different degrees of membership. Different combined effects can be taken into account as well. The meaning of the linguistic values depends on the working point. Membership functions for finer partitions are more easily generated by a gradually refining set of levels. Fuzzy partition can be chosen independently for each variable, and some adaptive features can be achieved by this representation [Juuso 1996], e.g. four fuzzy models corresponding to different valve positions can be combined by a single LE model


1 1] [ p1

p2 ] = 0

if the partition level is one for both the valve position $u$ and the current pressure p1 valve positions open, half open, almost closed, and closed are replaced by numbers 1, 0, -1 and -2, respectively, and pressure levels low, medium, high, and very high by numbers -1, 0, 1 and 2, respectively; the resulting new pressure $p_2$ is obtained on the partition level three. In the LE systems, the partition level is needed only for development of fuzzy set systems. Different strengths of the interactions are handled by using other positive and negative integers. Real valued linguistic equations (RLE) provide a basis for a sophisticated nonlinear systems where fuzzy set systems are used as a diagnostical tools. Fuzzification and defuzzification are integrated to the flexible scaling generated for membership definitions [Juuso 1996]. These systems can be easily tuned by neural networks, e.g. self-organizing maps and linear networks have been tested. Linguistic equations provide a method for generalizing results of neural networks. A real valued linguistic equation system can be even considered as a new neural network type. Extension to real numbers was introduced because of difficulties to handle variables with dissimilar fuzzy partitions. Actually, the example described above can be represented by even better by RLEs by using interaction matrix [4 -1 -4] together with same membership definitions for both pressures. The RLEs are also used in a tuning algorithm for reducing the error between model and training data. Fuzzy models can changed into linguistic equation models by replacing linguistic labels with real numbers. The FuzzEqu toolbox [Juuso 1996] includes routines for building a single LE system from large fuzzy systems including

various ruleblocks implemented in FuzzyCon or Matlab FLT. Other fuzzy modelling approaches can be used as channels for combining different sources of information: local linear models are considered as Takagi-Sugeno fuzzy models, and fuzzy clustering and self-organizing maps are used for preprocessing of data. Fuzzy models on any fuzzy partition can be generated from LE models: rules or relations are developed either sequentially or simultaneously [Juuso et al. 1996], and membership functions are generated from the membership definitions on any location. In each equation, the locations of membership functions are defined by the model for one selected variable. Singleton models represent the LE model quite accurately if the locations of the membership functions are based on the shapes of the membership definitions in such a way that the linear surfaces are on appropriate areas. Takagi-Sugeno (TS) fuzzy models have the same requirement, but the locations of membership functions are different. Linguistic fuzzy models are developed from singleton models. Fuzzy relational models are useful for fuzzy LE models. There will be fairly few nonzero elements since nonlinearities are included to the membership functions.

Figure 2: Examples of membership definitions for several cases.


The modelling for different cases can be done with FuzzEqu toolbox created in Matlab environment. In the Linguistic Equations approach, non-linear multivariable models are constructed in three parts: directions of interaction are handled with linear equations, and nonlinearities are taken into account by membership definitions, and time delays depend on operating conditions (process cases).

Automatic generation from data

Membership definitions are generated directly from process data on the basis of areas of operation and variation for each process case. According to the results some relations were found out between membership definitions and web break sensitivity. An example of membership definitions for one variable is presented in Figure 2. In this Figure web break sensitivity increases from left to right and there are five examples in each category. The FuzzEqu toolbox contains routines for modifying membership definitions interactively: definitions can be extended or contracted independently on positive or negative or on both sides; definitions can be moved as well. All proposed changes are checked in the FuzzEqu Toolbox. With these tools the system can be adapted to changing operating conditions.

Interaction matrices can be generated automatically by processing the linguistic levels of the variables as combinations of three variables at the time. All possible combinations were taken into account and the strongest interactions are selected for the model of the example in question. Each interaction will include at least one new variable to the model. Figure 4 shows an example of the interaction matrix. The last column is the bias vector, which is essential in fault diagnosis. Totally the prototype includes 25 sets of equations, one for each process case.

Figure 3: An example of the interaction matrix A for a single case.

Figure 4: Generating linguistic equations from a large fuzzy rule base.

Expert knowledge in Case-Based Modelling

Various fuzzy rule bases can be generated by personnel interviews and data collected from databases, e.g. in the functional testing project, the steps of various measurements are variables and they are grouped as a group of measurements. Different groups consist of 1 - 4 variables and 3 - 50 rules and the total number of rules is more than 300. Each variable has 3 or 5 membership functions. The fuzzy rule base has been tested with real test data by using FuzzyCon, which is a Windows application developed in Control Engineering Laboratory [Juuso et al. 1994] [Juuso and Leivisk, 1995]. The reliability of the rules has been verified and membership functions have been tuned during the testing.

The FuzzEqu-toolbox [Juuso 1996] has been used to form linguistic equations from rules and data. Figure 4 shows an example of linguistic relations for a single fault type. This set of relations consists of 8 relation groups, which contain 37 relations. Total number of variables is here 27. Each group has 2 - 4 variables. Generation of linguistic equations from linguistic relations is pipelined for measurement groups. Each group produces 1 - 3 equations.


Dynamic fuzzy models can be constructed on the basis of state-space models, input-output models or semimechanistic models [Babuska et al. 1997]. In the state-space models, fuzzy antecedent propositions are combined with a deterministic mathematical presentation of the consequent. The most common structure for the input-output models is the NARX /Nonlinear AutoRegressive with eXogenous input) model which establishes a relation between the collection of past input-output data and the predicted output. MIMO systems can be built as a set of coupled MISO models. Delays can be taken into account by moving the values of input variables correspondingly. Dynamic neural networks and linguistic equations can be based on similar structures as the dynamic fuzzy models [Juuso 1998]. The basic form of the linguistic equation (LE) model is a static mapping in the same way as fuzzy set systems and neural networks, and therefore dynamic models will include several inputs and outputs originating from a single variable. The dynamic behaviour is provided by external dynamic models: Rather simple input-output models, e.g. the old value of the simulated variable and the current value of the control variable as inputs and the new value of the simulated variable as an output, can be used since nonlinearities are taken into account by membership definitions. For state-space modelling, each rule correspond to a different working point, and the deterministic equations on the consequent part is transformed into a linguistic equation model. Since the LE model can handle nonlinearities, at least some of the rules can be combined. Local linear models can be combined with this technique which is closely related to Takagi-Sugeno -type fuzzy models. In semi-mechanistic models, approximation rules are combined into linguistic equations. The approximation rules can be based on qualitative knowledge, and the mechanistic models take care of the basic level dynamic simulation. Effective delays usually depend on the working conditions (process case), e.g. the delays are closely related to the production rate in many industrial process. Therefore, an appropriate handling of delays will extend the operating area of the model considerably. In small models, all the interactions are in a single equation. For larger models, the equation system is a set of equations where each equation describes an interaction between two to four variables. A multimodel approach based on fuzzy LE models has been developed for combining specialized submodels [Juuso 1998]. The approach is aimed for systems which cannot be sufficiently described with a single set of membership definitions because of very strong nonlinearities. Additional properties can be achieved since also equations and delays can be different in different submodels. In the multimodel approach, the working area defined by a separate working point model. The submodels are developed by the case-based modelling approach.

Linguistic equation approach combines various intelligent modelling techniques on a unified framework: it was originally developed for handling large knowledge bases in process design [Juuso 1992]; a close connection to fuzzy set systems was important already in the early applications [Juuso 1994]; data-driven modelling properties have brought the approach close ANN techniques. Fuzzy modelling and control was the main application area. Properties of the LE approach are continuously improved and extended on the basis of industrial experience with various application areas, and recently, emphasis is moving from development to direct applications. The LE approach is already used in control [Juuso et al. 1997, Juuso et al. 1998b], in intelligent analyzers [Murtovaara et al. 1998], in fault diagnosis [Juuso et al. 1998a], and in model based control design [Juuso 1998]. Fault diagnosis and intelligent analyzers are combined in model-based diagnostical process analysis (MDPA) [Juuso 1997]: the resulting systems can be used in various ways suitable for software sensors, risk analysis and detection of sensor failures. Sophisticated trend information can be utilized by temporal reasoning on the recent process history. The MDPA methodology has been tested with simulations, expert knowledge and real data. Also the early applications in managerial decision making [Juuso et al. 1993] and in production scheduling [Juuso1993] are recently reactivated.

MANAGERIAL DECISION MAKING The main problem in material purchasing for electronics manufacturing arise from component purchasing times being longer than delivery times to customers. So electronics manufactures cannot wait until orders are received before they begin to plan production facilities and processes, acquire equipment and determine materials requirements. Therefore it seems to be necessary to anticipate the demand of customers in a specific market areas in order to plan the material purchasing process correctly. Electronic devices are composed of hundreds or even thousands of components. Consumption data of various components are represented in discrete time series form (raw data). The anticipating or forecasting model used for this purpose has to process vague and uncertain information. An adaptive and hierarchical Fuzzy Logic Advisory Tool (FLAT) was developed for component purchasing in the electronics manufacturing processes of Nokia Telecommunications, Fixed Access Systems Operations [Frantti and Juuso 1996]. Membership functions are generated through on-line data using a feature extraction method combining qualitative and quantitative data either in numerical or in linguistic form [Mutka et al. 1997]. The model is divided into several hierarchical subsystems, the rule base of each is represented by linguistic relations and the data base by data distributions. The inference and reasoning are based on linguistic equations [Frantti and Juuso 1996]. As the membership functions are dependent on the operating point, the system is adaptive. Consumption prediction of the FCP model followed actual demand more accurately and it kept the storage level lesser level without any shortages. FAULT DIAGNOSIS Fault diagnosis applications require various specialised knowledge and large amounts of data. Although the system development can be started either from knowledge or from data, an appropriate performance is achieved only by combining both data and knowledge.

Web break sensitivity indicator

The web break indicator is aimed to give process operators a continuous indication of web break sensitivity in an easily understandable way. The effect of different process variables on the sensitivity to web breaks has been analysed using actual measurements from a paper machine. The data contains 43 different measurements on the time period of three years. In addition to the actual measurements, the data also contains information of the break occurrence. The development work is concentrated on the consistency of the machine pulp and other variables effecting the operation of the mixing chest. Altogether 32 variables have been included in the analysis for two parallel versions of the indicator. The web break indicators have been implemented as Case-Based-reasoning type applications with Linguistic Equation approach. The indicators compare online data to example cases in the database and give a numerical estimation of the risk level of web breaks as an output. Both versions have been tested with off-line data and also some online experiments have been carried out at a paper mill. According to these results, both versions of the web break indicator seem to operate as planned [Ahola et al. 1999].

Functional testing
Functional testing is the main method to prevent defective products to be delivered to customers. Test personnel has a lot of knowledge on reasons of different kind of failures. An adaptive and intelligent system has been developed using process failure information in the functional testing of plug-in units [Komulainen et al. 1997]. Relations to the production process are even more complicated. Only few human experts with long experience have this kind of knowledge. Most common failures are easy to find based on failure information, but additional measurements are often needed. One of the problems in functional testing is the inexactness of failure information as mentioned above. A failed or damaged component is identified using information, which points the place of the failure. A fuzzy rule base has been generated by personnel interviews and data collected from databases. The steps of various measurements are variables and they are grouped as a group of measurements. The system consists of three levels of hierarchy: channel level, group level and step level. On the channel level, a failed channel or channels are searched. Test results are handled as groups on group level and mainly fuzzy reasoning is done using the fuzzy rule base. When there is a lot of failed test results, conclusions are made using rules, which indicate the right group. Rules have partly been represented by linguistic equations (Figure 4).

PRODUCTION PLANNING SCHEDULING For production planning and scheduling, optimisation is the basis of the decision making: the number of possible alternatives is reduced by comparing alternative fuzzy values of the objective function, or by using linguistic equations together with the objective function without sets of relations. The linguistic equation approach provides a flexible environment for a smooth adaptation of scheduling rules to the changing operation conditions. For the Mid Term Scheduling, an aggregate optimisation problem is handled by an optimisation-based inference engine. In the Short Term Scheduling, there are usually several feasible alternatives, and therefore, some preference criteria can be used as well. The linguistic equation framework is an efficient method in representing the preferences and priorities. The Fuzzy Scheduler expands the ideas of the Knowledge Based Scheduler in a flexible way by using fuzzy preferences and priorities obtained from process requirements and available resources [Juuso and Jagdev 1999]. PROCESS CONTROL The first direct linguistic equation controller was implemented in 1996 for a solar power plant in Spain [Juuso et al. 1997]. The properties of the controller have been extended [Juuso et al. 1998b], and recently the multilevel LE controller has been installed for a lime kiln at UPM Kymmene Pietarsaari mills [Juuso 1998]. Both implementations are very compact and easier to tune than corresponding fuzzy controllers. Membership definitions are required for error, change of error, change of control, original error, trend, correction coefficient and all the working point variables. Modularity is beneficial for the tuning of the controller to various operating conditions, and most important is that the same controller can operate on the whole working area. For the high level control, calculation of state indices can be represented by linguistic equations. Actually, the linguistic value of the control change is at the same time a measure of the state deviation. These linguistic values can be used in selecting priorities for the control actions obtained from different control objectives. This extension is used in the lime kiln control [Juuso 1999]. INTELLIGENT ANALYSERS Linguistic equation models have been developed to forecast Kappa number in continuous cooking. By measuring the liquor concentrations from different circulations of the digester the Kappa number can be estimated in different stages of the cooking process. The control actions can be done before the pulp will reach the blow line of the digester. Measurements of the ABBs Cooking Liquor Analyzer (CLA 2000), temperature and the on-line Kappa measurement have been used in the modelling. According to the tests in different plants, the linguistic equation models work better than the neural network and regression models. The LE model is very suitable for the prediction because it is not so sensitive for changes in process conditions. Linguistic equations offer also tools for adaptive modelling the system in several operating conditions [Murtovaara et al. 1999].

Linguistic equations are useful for taking into account non-linearity especially in multivariable application systems. The main benefit is better and more accurate decision-making due to the model-based approach and systematic knowledge management. Insight to the process dynamic operation is the most important issue. Automatic generation of systems, model-based techniques and adaptation techniques are very valuable in developing and tuning the controllers. The linguistic equation approach increases the performance by combining various control objectives in a case-based approach. The LE approach is also very efficient as a modelling technique: models can be generated from data, various types of fuzzy rule-based models can be represented by LE models, and any LE model can be transformed to fuzzy rule-based models. The linguistic equation approach is succesfully extended to case-based modelling and dynamic modelling. Case-based approaches have used in fault diagnosis. For functional testing, large fuzzy expert systems were systematically constructed from expert knowledge. For most cases, the rule base was transformed into a compact set of linguistic equations. These applications provide a good basis for wider use of intelligent systems together with other techniques in improving process design and control.

Ahola, T., J. Ruuska, E. Juuso and K. Leivisk (1999). Web Break Sensitivity Indicator. Proceedings of TOOLMET99 Symposium, Oulu, April 15-16, 1999 (L. Yliniemi and E. Juuso, eds.), pp. 202-209, Oulu. Oulun yliopistopaino. Babuska, R., M. Setnes, U. Kaymak and H. B. Verbruggen (1997). Fuzzy Modelling: a Universal and transparent Tool. Proceedings of TOOLMET97 Workshop, Oulu, April 17-18, 1997 (L. Yliniemi and E. Juuso, eds.), pp. 1-27, Oulu. Oulun yliopistopaino. Frantti T. and E. K. Juuso (1996). An Adaptive, Hierarchical Fuzzy Logic Advisory Tool (FLAT) for Anticipating the Demand of Transmission Products. 4th International Conference on Soft Computing, Iizuka '96. Vol 1, pp. 410413, 1996, Fukuoka, Japan. Juuso E. K. (1992). Linguistic Equations Framework for Adaptive Expert Systems. In J. Stephenson, editor, Modelling and Simulation, Proceedings of the 1992 European Simulation Multiconference, SCS International. York, UK. San Diego, USA. pp. 99-103. Juuso, E. K. (1993). Linguistic Simulation in Production Control. In UKSS'93 Conference of the United Kingdom Simulation Society, Keswick, UK, 13 - 15 September, 1993 (R. Pooley and R. Zobel, eds.), UKSS, 1993, pp. 34-38. Juuso, E. K. (1994). Linguistic Equation Framework for Adaptive Expert Systems. Proceedingds of Fuzziness in Finland FiF94. Plant Design Report Series No. 21, pp. 50-60, Helsinki, Hakapaino. Juuso E. K. (1996) Linguistic Equations in System Development for Computational Intelligence. Proceedings of the Fourth European Congress on Intelligent Techniques and Soft Computing -EUFIT'96, Aachen, September 2 5, 1996 (H.-J. Zimmermannn, ed.), Aachen, 1996. Verlag Mainz. 1996, Volume 2, pp. 1127-1131. Juuso, E. K. (1997). Intelligent Methods in Diagnostical Process Analysis. Proceedings of the XIV IMEKO World Congress, New measurements Challenges and Visions, Tampere, June 1-6, 1997 (J. Halttunen, ed.), Volume VII, pp. 1-6. Juuso, E. K. (1998). Intelligent Dynamic Simulation of a Lime Kiln with Linguistic Equations. Proceedings of the ESM98, Manchester, UK, June 16-19, 1998, Volume 2, pp. 23-28, Delft, The Netherlands, SCS. Juuso, E. K. (1999). Fuzzy Control in Process Industry: The Linguistic Equation Approach. In: Verbruggen, H. B., H.J. Zimmermann and R. Babuska, editors, Fuzzy Algorithms for Control, International Series in Intelligent Technologies, pp. 243-300. Kluwer, Boston. Juuso, E. K., and H. S. Jagdev (1999). Intelligent Scheduling of Semi-Continuous Systems. Proceedings of TOOLMET99 Symposium, Oulu, April 15-16, 1999 (L. Yliniemi and E. Juuso, eds.), pp. 182-193, Oulu. Oulun yliopistopaino. Juuso, E. K., and K. Leivisk (1995). A Development Environment for Industrial Applications of Fuzzy Control. In Proceedings of the Third European Congress on Intelligent Technoiques and Soft Computing -EUFIT'95, Aachen, August 28 - 31, 1995 (H.-J. Zimmermannn, ed.), volume 2, pp. 796-803, Aachen, 1995. Verlag und Druck Mainz Juuso, E. K., J.C. Bennavail ,and M.G. Singh (1993). Hybrid Knowledge-Based System for Managerial Decision Making in Uncertainty Environment. In: Qualitative Reasoning and Decision Technologies, Proceedings of the IMACS International Workshop on Qualitative Reasoning and Decision Technologies -QUARDET'93, Barcelona, June 16 - 18, 1993 (N. Piera Carret and M. G. Singh, eds.), CIMNE, Barcelona 1993, pp. 234-243. Juuso, E. K., J. Myllyneva, and K.Leivisk (1994). Fuzzy Logic Controller and Adaptive Tuning. Proceedings of the 1994 European Simulation Multiconference, June 1-3, 1994, Barcelona, Spain (A. Guash & R. Huber, eds.), pp. 475-479, San Diego, 1994, SCSI.

Juuso, E. K., T. Ahola, K. Leivisk (1996). Fuzzy logic in Lime Kiln control. Proceedings of TOOLMET96 Workshop, Oulu, April 1-2, 1998 (L. Yliniemi and E. Juuso, eds.), pp. 111-119, UOulu. Juuso, E. K., P. Balsa and K. Leivisk (1997). Linguistic Equation Controller Applied to a Solar Collector Field, Proceedings of the European Control Conference - ECC97, Brussels, July 1-4, 1997, volume 5, TH-E 14, Paper 267, 6 pp. Juuso, E. K., T. Ahola, K. Oinonen (1998a). Case-based Reasoning with Linguistic equations for Web Break Sensitivity Indicator. Preprints of TOOLMET98 Symposium, Oulu, April 16-17, 1998 (L. Yliniemi and E. Juuso, eds.), pp. 145-155, Oulu. Oulun yliopistopaino. Juuso, E. K., P. Balsa and L. Valenzuela (1998b). Multilevel Linguistic Equation Controller Applied to a 1 MW Solar Power Plant, Proceedings of the 1998 American Control Conference - ACC98, Philadelphia, PA, June 24-26, 1998, volume 6, pp. 3891-3895. ACC. Komulainen, K., M. Heikkinen, T. Frantti, E. Juuso, and K. Leivisk (1997). Utilization of Failure Information in Functional Testing. Proceedings of TOOLMET97 Workshop. Oulu, Finland, 109-115. Murtovaara, S., Juuso, E. K., R. Sutinen (1998). Modelling of Pulp Cooking Characteristics by Using Fuzzy Logic and Linguistic Equations. Preprints of TOOLMET98 Symposium, Oulu, April 16-17, 1998 (L. Yliniemi and E. Juuso, eds.), pp. 90-96, Oulu. Oulun yliopistopaino. Murtovaara, S., K. Leivisk, E. Juuso and J. Ikheimonen (1999). Kappa Number Prediction Based on Cooking Liquor Measurements. Proceedings of TOOLMET99 Symposium, Oulu, April 15-16, 1999 (L. Yliniemi and E. Juuso, eds.), pp. 210-215, Oulu. Oulun yliopistopaino. Mutka, P., E. K. Juuso, K. Leivisk, and T. Frantti (1997). Uncertainty Modelling in Chaotic Environment. Proceedings of European Symposium on Intelligent Techniques, Bari, Italy 20-21.3.1997, 40...44.