16 Fuzzy Identification of Systems and Its Applications to Modeling and Control TOMOHIRO TAKAGI Abstract—A mathematica tot to build «fuzzy model ofa system where fuzzy implications and reasoning reeset ls presented in ths paper The ‘premise ofan implication isthe description of fuzzy subspace of inputs and fon of sytem tins pets sn Pesce tons ofthe method Yo ind processes are ako ced 2 eter cleaning process and a converter in steel-making process \ T= main purpose of this paper is to present a ‘mathematical tool to build a fuzzy model of a system. There has been a considerable number of studies (1]-[3] on fuzzy control where fuzzy implications are used to express control rules, Most of those implications contain fuzzy variables with unimodal membership functions since those are linguistically understandable and thus called linguistic variables. As for reasoning, the so-called com positional rule of inference or its simplified version is used. However, when we use this type of reasoning together with unimodal fuzzy variables for multivariable control, we have much difficulty since we need many fuzzy variables, ie, many implications; it is usual to use five variables in each dimension of input space. The authors have suggested multidimensional fuzzy rea- soning [6] where we can surprisingly reduce the number of implications. The study in this paper is related to the above idea of reasoning, where a fuzzy implication is improved and reasoning is simplified. Recently some studies [4], [5] have also been reported on fuzzy modeling of a system. Fuzzy modeling based on fuzzy implications and reasoning may be one of the most important fields in fuzzy systems theory. Here we have to deal with a multivariable system in general and so there- fore have to consider multidimensional reasoning method. jenerally speaking, model building by input-ou! data is characterized by two t piss math matical tool to express a system model and the other is the method of identification. A ematical tool itself is required to. have simplicity and generality, The fuzzy implication pre- sented as.a.tgol in the paper is quite simple. It is based on a fuzzy partition of input space. In each fuzzy subspace a linear input-output relation is formed. The output of fuzzy 1. Inrropucrion Manuscript received September 30, 1983; revised Ay er 30, 1983; revised April 6, 1984 ‘The authors are with the Department of Systems Science, Tokyo astute of Technology, 4259 Nagata, Midorko, Yokohaina 227, 0018-9472/85 /0100-0116801.00 ©1985 IEEE IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. SMC-15, NO. 1, JANUARY/FEBRUARY If AND MICHIO SUGENO reasoning is given by the aggregation of the values inferr by some implications that were applied to an input. identification is divided into two parts: tiop and parameter identificat In its nature structure identification is almost indepei dent of a format of system description. We omit this ps in the paper, so by identification we mean paramet identification in fuzzy implications. However, a kind structure problem partly appears. Finally this paper shows two applications to indust processes. One is a water cleaning process where an ope ator's control actions are fuzzily modeled to design a fi controller. The other is a converter in the steel-makin process where the conversion process is fuzzily mode and model-based fuzzy contro! is considered. ‘Most fuzzy controllers have been. designed based_o human operator experience and/or control engin Knowledge. It is, however, often the case that an operato cannot tell linguistically what kind of action he takes in Particular situation. In this respect it is quite useful to gi a way to model his control actions using numerical dat Further, if ‘reason to believe that an operator§ control is optimal, we have to develop model-based contr just as in ordinary control theory. To this aim it is nece sary to consider a means for fuzzy modeling of a syste FORMAT OF FUZZY IMPLICATION AND REASONING ALGORITHM IL. In this paper we denote the membership function of fuzzy set A as A(x), x © X. All the fuzzy sets are ass ated with linear membership functions. Thus, a membe ship function is characterized by two parameters giving tt greatest grade 1 and the least grade 0. The truth value of proposition “x is A and y is B” is expressed by [is A and y is B|= A(x) 4 B(y) (A) Format of Implications. @ of Impl ns. We suggest that a fuzzy implication R is of the form: Ret f(x is Ayers is AQ) then y = g(xTAKAO! AND SUOENO: FUZZY IDENTIFICATION OF SYSTEMS where ’ Variable of the consequence whose value is inferred. xk f the premise that appear also in the consequence. 4, Ag Fuzzy set representing a fuzzy subspace in which the implication R can be applied for reasoning. f function connects the propositions in the premise, 8 Function that implies the value of y when * In the premise if A, is equal to X, for some i where X, the universe of discourse of x, this term is omitted; x, is unconditioned, Example 1 R: If-x; is small and x, is big then y = x, + x) + 2x5. ‘This implication states that if x, is small and x, is big, then the value of y would be equal to the sum of x, 7 and 2x3, where x, is unconditioned in the premise. In the sequel we shall only use “and” connectives in the premise and adopt a linear function in the consequence as 1S seen in the above example. So an implication is written Ri Ix; is A and ---and x, is Ay then y = Po + pix + *"" *PaXe @) Suppose that we have implications R' (i = 1, ‘se above format, When we are given on) of van xf) where x) — xP are singletons, the value of y is inferred in the following steps. U) For each implication R, yi_is calculated by the function g! in the (a= ater 13) The truth value of the proposition y = y' is calculated by the equation [y= y'| = [x9 is Aj and --- and x2 is A,)| A 1R'] = (A(x?) A @) where |+| means the truth valué of proposition + and A stands Tor mip operation, and |x® is A| = AC). i arade of the membership of x°. + AAL(x2)) ATRL (3) © a7 TABLET ass ars Pas yess a [3] The final output_y inferred from_n implications is given as the average of all y' with the weights |y = y'|: Diy =y'lxy! - O) 4 Wein ue 40) Example 2; Suppose that we have the following three implications: Ri: If-x; is small, and x, is small, then y = x; + x2 RP If x; is big, then y = 2X x RP: If x, is big, then 3X x, Table I shows the reasoning process by each implication when we are given x, = 12, x, = 5. The column “Premise” in Table I shows the membership functions of the fuzzy sets “‘small” and “big” in the premises. The column “Con- sequence” shows the value of y’ calculated by the function ¢' of each consequence and “Tv” shows the truth value of ly = yj) For example, we have Iy= y!] = |x? = small) A [x9 = small} = small, (x?) A small, (x?) = 0.25. (Oo) ‘The value inferred by the implications is obtained by referring to Table I 0.25 x17 + 0.2 x 24 + 0.375 X15, 0.25 + 0.2 + 0.375 (9) oh © Peper of Resnng We show two illustrative examples to find the perfor- mance of the presented reasoning algorithm. Example 3: Suppose we have two implications. pouee Z- then y= 0.2x +9 iwus Fig. 1. Results of fury reasoning R tee > - then y = 0.6x +2 Then Fig. 1 shows the relation of x and y, which is marked by the symbol +. The line R' shows the function in the Consequence of R’. The equation in a consequence can be interpreted to represent a law that holds in the fuzzy subspace defined in a premise Let us consider the difference between ordinary piece- wise linear approximatior a mi ‘and_the presented. method. If we take piecewise linear approximation, we first divide input space into crisp subspaces and next build a linear relation in each subspace. For example, in the case shown in Fig. 1, we need another linear relation connecting Ri and R®. It is easily seen that those three straight lines are not smoothly connected. On the other hand, the pre- Sented method enables us to reduce the number of pigoe- wise linear relations and also to connect them smoothly. It is of crucial importance to reduce the number of linear relations in a multidimensional case. Further, with the fuzzy partition of ing ut linguistic conditions to linear relatic small and x, is big.” Thus, for example, we can use the variable_that is observed only by man (see Section IV-B). Example 4: Fig. 2 shows the input-output relation ex- pressed by the implications of Example 2. In this case the Premises are two-dimensional. In the figure the curved Surface shows a highly nonlinear input-output relation whose shape reflects the dominance of each implication in its essentially applicable area and also the conflict of implications in an overlapped area. ut space, we can such as “x, is TIL Aucoriram oF IpenrieicaTion ‘AS has been stated, we consider a fuzzy model consisting of some number of implications that are of the format Ixy is Ay and +++ and x, is Ay then y= py + ppm to type xy characterized by “and” connective and a linear equation For identification we have to determine lowin, three items by using the input-output data of an objective system, D su-+-y4 Variables composing the premises of ian plications. Fig. 2. Dawes Dr. Notice that all the variables in a appear. The items 1) and 2) are related to the partition of space’ of input variables into some fuzzy subspaces. T! item 3) is related to describing an input-output relation i each fuzzy subspace. We can consider the relation among three ite hierarchically from 1) down to 3). The algorithm of th identification of implications is divided into three ste correspor ms. We first give each step, © Choice of Premise Variables: First a combination of prémise variables is chosen out of possible input variable we can consider. Next the optimum premise and conse quence parameters are identified according to the steps 2) and 3), and also the errors between the output values of the ‘model and the output data of the objective system are calculated. We then improve the choice of the premis variables so that the performance index is decreased, w is defined as the root mean square of the output errors. °Q) Premise Parameters Identification: In this step tl ‘optimum premise parameters are searched for the premi variables chosen at step 1). Assuming the values of premise Parameters, we can obtain the optimum consequence Parameters together with the performance index according| to step 3), So the problem of finding the optimum premise parameters is reduced to a nonlinear programming probs Jem minimizing the performance index. © (3) Consequence Parameters Identification: The cons quence parameters that give the least performance ind are searched by the least squares method for the gi jx. Membership functions of the fuzzy sets the premises, abbreviated as premik parameters. ~— _— + Px Parameters i the consequences,-paxaot AND SUGENO: FUZZY IDENTIFICATION OF SYSTEMS snise variables in the step 1) and parameters in step 2). “The outline of the algorithm is shown in Fig. 3. From the next section we shall discuss the method in detail with an iiustrative example at each step. ‘A. Consequence Parameters Identi In this section we show how to determine the optimum dex, provided that. both th variables V aeare-quen. The rs index has above as a root mean square of the output errors, which means the differences between the output data of an origi- nal system and those of a model. ‘Let a system be represented by the following implica- tions: RE fx, is Alye--vand x, is AL then y= Pht Phot tata | sand x, is AT \ +Phe +, x,) is obtained as RY Ix ist, then y = pp + pt-x + xe ‘Then the output y for the input (x, wwe vine oS \ Anl%y) Ao Au Xe) Pa Sante) A Mala) 0) Ye Piedad” (as) P= [phases Pb Phases Phe oy Phas *s PE]. (26) Then the parameter vector P is Seed esse oF x an femueber SeKPS Pe Fo xTKIEP. Tt is noted that the proposed method is consistent with the reasoning method. In other words, this method of identification enables us to obtain just the same parameters ‘as the original system, if we have a sufficient number of noiseless output data for the identification. In this paper the parameter vector P is calculated by a stable-state Kalman filter. The so-called stable-state Kalman filter is an algorithm to calculate the parameters of a linear algebraic equation that gives the least squares of ‘errors. Here we apply it to calculate the parameter vector P in (17). Ronbee Of —— a ~ Laila) a- ROLES (ny [Earn aa) then rs y= EAC ph tate t oo teh om) = Lp Bt pie Bit HP a Bi) a ' teense tn nk feeamaters ore get (19) (yaa set-of input-output date. i Bay 84s 2 (= 1 +> mi) is gga, we can_obiain the consequence Parameters pi, p’ 1 method usi _ st X (m xX n(k +1) matrix), Y (m vector) and P (nk +1) vector) be Busts Basin’ Bus Bm?" Baus Mim Bin? Ye of tape ar = kei an we Matar Nail. be + AA(4)) “(75 + PL + | \ 7 i © (Aix) Ao AAC) \ +P %) (10) Let the ith row vector of matrix X defined in (13) be x, and the ith element of Y be y;. Then P is recursively calculated by (18) and (19) where S, is (nm - (K+ 1)) x (n- (k + 1)) matrix. Pras = Pit Sion Xion Oven — Are Pi) (8) S25 - Seats pg. T+ x Seah (19) P= Py (20) where the initial values of Py and Sy are set as follows. P=0 (2) Sy = a1 (a= big number) (22) where J is the identity matrix. (a3) OF p Seems,20 — semtied etton PETS E Th Fig. 4. Results of identification, Fig. 5 Tnput-output data Example 5: Suppose in a system : ifx, _—\_ then y = 0.6x +2 if x, is _Z. then y = 0.2x +9. Under the condition that the premises of the model are fixed to those of the original system, the consequences are identified from input-output data as follows, where the noises are added to the data. then y = 0.56x + 2.17 if xis if x, is NL 7 + a 0 then y = O.11x + 9.60. Fig. 4 shows the noised input-output data, the original Consequences, and the identified consequences. B. Premise Parameters Identification (2) In this section we show how to identify the f apis ~Premi space of pi that the premise IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL, SMC-I5, NO, 1, JANUARY/FEBRUARY I For example, looking at the input-output data shown Fig. 5, we can see that the input-output characteristi change as the input x increases. So dividing the space of into two fuzzy subspaces such that x is small or x is bi ‘we have a mode! with the following two implications: if x is small then y = a,x + by if x is big then y = a,x + 6, We next have to determine the membership functions “small” and “big” as well as the parameters a,, by, ay by in the consequences. As it is easily seen, to divide the spaces into some f subspaces is to determine the membership functions of fuzzy sets in the premises. The problem is thus to find ‘optimum param of their membership function ‘which the performance index is minimized, We call this tin.” The algorithm is as follows. *@ Assuming the parameters of the fuzzy sets consequences the performance i cussed in the previous section. ‘+ The problem of finding the optimum premise par ters minimizing the performance index is nonlinear programming problem. In this study we use tl well-known complex method for the minimization. fuzzy set is assumed to have Example 6: This example shows the input-output data gathered from a preassumed syst with noises. The standard deviation of the noises is fi percent of that of the outputs. It has to be also noted wwe can identify just the same parameters of the premises the original system if noises do not exist. tis of great importance to point out the above fact. If is not the case, we cannot claim the validity of an identifi cation algorithm together with a fuzzy system descrip language. ‘Suppose the original system exists with the following implications: on then y = 0.6x +2 vw ee then y = 0.28 + 9. 0 ‘The functions in the consequences of the implications the noised input-output data are shown in Fig. 6. The identified premise parameters are as follows. We see that almost the same parameters have been derived.{NKAGH AND SUGENO: FUZZY IDENTIFICATION OF SYSTEMS TTP pret a ows Fig. 6 Consequences and noied data, aw then y = 0.59x + 2.2 $e iw 8 then y = 0.12x + 9.5 7 Too In this section we suggest an algorithm to choose pre: les from the considerable input variables. As sted previously, all the variables of the conse- quences do not always appear in the premises. There are. ‘wo problems concemsd wth the algorithm Que isthe “isthe implies that it itsspave i divided ‘The other is the number of divisions. The whgle problem is a combinatorial one. So in ‘thee eal approach avalible H Here Suppose that we build_a fuzzy. k:input yrs, and single-output system, Step 1? The range of 2 is divided into wo fuzzy sue spaces “big” anc id the ranges of the ‘other variables x,,---, x, are not divided, which means that only x, appears in the premises of the implications. This model consisting of two imp! cations is thus lel_of if, isbig, then if x, is small then - Its called model 1-1, Similarly, a model in which the range of x, is divided and the ranges of the other variables x,,x,,°+", x, are undivided is called model 1-2. In this way we have k-models, each of which is composed of two implications. In if x, is and x, is = ZZ. am Fig. 7. Choice of premise variables. general, the model 1 — / is of the form if x, is big, then +++ if x,issmall, then» For each model the optimum tnd consequence parameters are found bY the algorithm described in the previous, The optimum model withthe etpetoenane poriormance index is adopted out of the k-models. It is called @ sep Starting from a stable state at step 1, say model 15 i, where only the variable x, appears Lind, where pear Dremises, take all the combinations of x,— x, (j= 1,2,"7K) and divide the range of each variable in two fuzzy subspaces. For the combina- tion x,— x, the range of x, is divided into four subspz for example, “big,” “medium big,” “medium small,” and “small” Thus we get ke ‘models each of which is named model 2, Each model consisis of 2X 2 implications. Then find again a model with the least performance index Just as in step 2 that is also called a stable state at this step. spt see aps @2) The number of implications of a stable state exceeds the predetermined number, ‘The choice of the variables in the premises proceeds as is shown in Fig. 7. Example 7: We show an example of identification. The original system is also a fuzzy system with two inputs and single output expressed by the implications as are shown. below. then y= 12x, +0.2x, +1 then y = 2.5x, + 21x, +4 y12 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. SMC-LS, NO, 1, JANUARY/FEBRUARY I > 7s Fig, 10. Identification data with noises, In this system, the range of x, is divided into four f subspaces and that of x, into two fuzzy subspaces. So number of implications is 4 x 2 altogether. Fig. 8 shows a fuzzy partition of the input space in t original system, i., the membership functions of the fu relation expressed in the premises. Fig. 9 shows the input-output relation of the abo system, Now 441, input-output data of this system taken for the identification and noises are added to t outputs, where the standard deviation of the noises is tw percent of that of the outputs. Fig. 10 shows the noi data. Now the system is identified using these data. Stage 1: Let us start from two models, each of whi consists of two implications. They are shown together wit tne ine then y = eu and x, is. ine and x, is. if x, ye and x, is = tas ag sai then y = 0.3x, + 3.0x, +5 Fig. 9 Jk, + 13x, +2 then y = 0.5x, + 24x, +8 then y= 17x, + 13x, +2 KV A then y = 0.5x, + 24x, +8 then y = 0.9x, + 0.9x, +7[TAKAGI AND SUGENO: FUZZY IDENTIFICATION OF SYSTES, 3 ps Fig. 11. Input-output relation ofthe model 1-1. Fig. 12, Input-output relation of the model 1-2 sneit performance indices. Figs. 11 and 12 show the input-output relation of the models. ‘Model 1-1: (the Range of x, is Divided) Implications \ if xis then y = —0.45x, + 1.82x, + 23.2 io 30 / itnis—; then y= 1.71x, + 3.09x, - 49 performance index = 2.55. ‘Model 1-2: (the Range of x, is Divided) Implications if x, is a then y= —O.14x, + 0.65x, + 27.6 if xis 5 = then y = 0.9Lx, + 2.06x, + 2.3 Vi performance index = 3.73. ‘The model 1-1 is found to be a stable state at the stage 1 since its performance index is the minimum of the two. In fact lg, 9 seems more similar to Fig. 11 than to Fig. 12. At the next stage we fix the variable x, in the premises. Stage 2: In this stage the space of inputs is further divided. In the model 2-1 (Fig. 13), which is the extended one of the model 1-1, the range of x, is divided into four subspaces, leaving that of x, undivided. l (small) xis 33 tao \ (medium small) x's range only rey x, is NL (medium big) as \ xis sar124 ig 13. Tnput-output relation ofthe model 2-1. In the model 2-2 (Fig. 14), the ranges of x, and x, are newly divided in two fuzzy subspaces, respectively. x he (small,) rc xis _Zo (bie,) y's range Xp or (small,) _x,’s range xis Zo. (vig) Notice that this partition of the range of x, is different from that of model 1-1. This is because the optimization of fuzzy sets concerned with x, is performed together with those concerned with x3. Model 2.2: Implications tas ens tae vet ne tava vt tan vot ee sf Ey performance index = 1.40, IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. SMC-15, NO. 1, JANUARY/FEBRUARY I Fig. 14, Input-output relation of the model 22. ‘The implications of the two models and their perf ‘mance indices are obtained this time as follows. ‘Model 2-1: Implications if xis. then y = 1.90%, + 3.65; — Se ww if xy is gig then y= 1.46 + 0.69%, + is NN sn y =-0.58x, + 0.78x, + 364 if xis ppg then 9 0 SB + ifx, we then y = 0.88%, + 2.22x; - 1. performance index = then y = 2.31x, ~ 1.36x; ~ 16 then y 1.76x, + 1.99x, + 6.8 then y = —0.92x, + 2.24x, + 26.4 then y = 0.84x, + 0.5Lx, +12{AKAGI AND SUGENO: FUZZY IDENTIFICATION OF SYSTEMS ws ‘The model 2-2 is now found to be a stable state at stage 2. Stage 3: We show the implications and the performance indices of two models (Figs. 15 and 16) extended from the model 2-2. Model 3: Implications Ne te then y = 1.14x, — 0.38x, + 1.8 if x, is ‘NL and x, hn then y = 2.33x, + 1.98%, + 5.6 if x, Lk 1 and x, is Ne then y = 2.20x, + 0.26x, +13 ity Le and x is a then y = ~0.04, + 1.0%, + 100 itm, ns then y = 1.65n, + 1434 +19 ifx, Ne and x, ke then y = 0.62x, + 2.23x, + 79 a Ties is me \ then y= 0.74x, 4-27 if x, is “ 5 and x, is — y= 0.74x, + 2.84x, — 2. it x, pL and x; ae then y= 12x, +0.63x, +48 performance index = 1.08 ‘here the partition of the range of 2, is four and that of x, is two, as is seen. Model 3-2 Implications y itis and x, is = then y = 1.78, + 1.71x, ~ 2.0 y KY and x, is then y = 2.84x, + 4.04x, — 17.5 V ifx, is and x, is then y = a 39x, + 14x, + 0.2 / \ if x, is and x, is. then y = 87x, + 2.07x, + 5.7 3e aT126 Fig. 15. Input-ovtput relation ofthe mode! 31, 3 performance index = 1.33, At stage 3 the model 3-1 is found to be a model with the same structure as the original system, which is of course a stable state. We can say that the presented method enables us to derive almost the same premise parameters as those of the original system, We can also recognize that Fig, 9 is almost the same as Fig. 15. ‘The choice of the premise variables has proceeded in this, example as is shown in Fig. 17. IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. S4C-L5, NO. 1, JANUARY /rEMRUARY sunt eS stable Fig. 17. Choice ofthe premise variables. TV. APPLICATION To Fuzzy MODELING This chapter shows two practical applications of 1 in a water cleaning process. The obtained model may directly used in place of an operator to control the proc The other one is the fuzzy modeling of a converter in| steel-making process. The relation betwee. the input-~ put of the converter is so complex that an approprial algebraic model has not been developed. The obta fuzzy model is applied to the control of the converter, the results are compared with the case when an opera controls it without a’ model. then y = =1.47x, — 0.72x, + 40.2 ’ 1 then y = 0.67x, + 0.40x, + 15.2 then y = 0.15x, + 0.19x, + 26.8 then y = 0.92x, + 1.00x, + 5.9 A. Fuzzy modeling of human operator's control actions Water Cleaning Process: We shall now show an examp! where an operator's control actions are fuzzily modeled. The control process is a water cleaning process for civ water supply as is illustrated in Fig. 18. In the proc turbid river water first comes into a mixing tank whe chemical products called PAC and also chlorine are p\ and mixed in the water. Then the mixed water flows into sedimentation tank where the turbid part of water cohered with the aid of PAC and settled to the bottoIAKAGH AND SUGENO: FUZZY IDENTIFICATION OF SYSTEMS ww 1 do wa L_oneree Fig. 19. Diagram of control process. ‘Alter sedimentation, which takes about 3-5 hours depend- ing on the capacity of the tank, the treated water finally flows into a filtration tank producing clean water. Chlorine is added only for the sterilization of the water. ‘The main control problem of a human operator in this process is to determine the amount of PAC to be added so that the turbidity of the treated water is kept below a ceriain level. The optimal amount, not too little, nor too much, depends on the properties of the turbid water. The amount of PAC must be controlled also from an economi- cal point of view. ‘The process is characterized by a lack of any physical model, significant variation of the turbidity of river water and the fact that turbidity itself is not clearly defined nor accurately measured. So an operator's experience is a key ‘ctor in this control process. However, a number of variables influencing sedimenta~ tion process have been found so far that can be measured. rst list all the variables concerned. Turbidity of the original water (ppm). ‘Turbidity of the treated water (ppm). Amount of PAC (ppm). Temperature of water (°C). PH. Alkalinity. ‘Amount of chlorine (ppm). For example, if TE is lower, then more PAC is necessary. Both PH and AL affect nonlinearly the necessary amount of PAC. The optimal PAC depends on these variables; the relation among them is not clear. There are some other variables influencing the process, eg, plankton in the river Water, which inereases in springtime but cannot be mea- sured at present ‘n most water cleaning processes a statistical model has been built. However the models are not accurate. These cover only steady state, i.e, a small range of TBI. TBL increases for example 100 times more when it rains. So an Z=Tph avrel sLYBA TSR, x 7M, Neral Fig. 20. Partition ofthe ranges of premise variables. ‘operator controls PAC taking into account of TBI, TE, PH, AL, and TB2. Now our process can be illustrated as in Fig 19. Derivation of Control Rules We have a lot of operation data where all the variables are measured every hour for four months. That is, the number of data is 24 hours x 30 days x 4 months = 2880. Table IT shows a part of these. ‘Among the data we have used about 600 for the identifi- cation taken in June and July. June in Japan is a rainy season and July is summer. ‘According to the identification algorithm discussed pre- viously, eight control rules are derived that can be called a fuzzy model of operator’s control as is stated, where a control rule is of the form If (PHis*), (ALis®) and (TEis*) then PAC = py + p; - TBI + pz TB2 + py» PH + py- AL + ps TE18 REE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. SMC-15, NO. 1, ANUARY/FEBRUARY PH, AL, and TE are picked up as premise variables and their ranges dis number of the control rules is thus 2° = 8, Those are shown in Fig. 21. RB wns Se ALis woe and TEis then PAC = 2664 - TB1 — 8093 - TB2 + 11230 - PH ~ 1147- AL ~ 2218 - TE + 8858 R ee ALis \N and TEis aE 30 then PAC = 124 - TBI — 427 - TB2 + 761 - PH + 52- AL ~ 17- TE ~ 7484 Ry amis eat wx and TEis then PAC = 42 - TBI - 54- TB2 ~ 1368 - PH + 10- AL + 158 - TE + 7270 eum Sg ate ge wa then PAC = ided into small and big, as shown in Fig. 20. V rs or N Tar 08 / Tee : a - TB1 ~ 34- TB2 ~ 221- PH — 8- AL + 40 - TE + 2202 Re: if PHis \ oa oe and TE is Ee then PAC = 3- TBI — 6-TB2 + 2110 - PH ~ 13- AL + 2- TE ~ 13918 Re: ims Se ALis \N_ and TEis 7 aT War NY’ then PAC = 22- TBI + 11 - TB2 + 64- PH - 8- AL ~ 9-TE770 RY cms ALis Kn and TEis - then PAC = 159 TBI ~ 14+ TB2 + 2337 - PH — 25 - AL ~ 69 - TE ~ 14819 Rt vin Se Ais Zo and TEis TR hi} wr then PAC = —13 + TB1— 16 - TB2 + 29 P+ 6.4L 441 TE- 37 Fig. 21, Control rules. Results of Fuzzy Control model is represented in (23) that is usually used in a wal ‘The performance of the derived control rules is tested by cleaning process. using testing data, The results are shown in Table IIT as well as operator's control input and the results of a statis: PAC = 9.11 TBI ~ 79.8 PH + 12.7CL + 12556 tical model, where we used 38 testing data. The statistical ¢AKAGI AND SUGENO: FUZZY IDENTIFICATION OF SYSTEM TABLE perasor Seatettet fate ES os 28s3 It is seen that operator's control actions are well mod- led in the form of fuzzy control rules. The average of the absolute differences between the results of the fuzzy model and the operator, and those between the results of a statistical model and the operator are, respectively, fuzzy model 48.5 statistical model 128.0, ‘These results show the excellence of the fuzzy model. B, Fuzzy Modeling of Converter in a Steel-Making Process and its Control * The Problem The steel-making process consists of the following four steps. 1) Iron ore is melted in a blast furnace. The obtained molten iron called hot pig is removed by a torpedo car into a converter after desulfurization, 2) Ina converter, scrap, iron ore, and burnt lime are first added to hot pig, and then decarbonization and dephosphor are performed by oxygen below. After that various alloys are added for adjusting the in- sredients of produced steel 3) Floating slag is taken away and the amount of ingredients is readjusted in a ladle refining process. 4) It is then cast and finally cut into appropriate figures, Each step except the final half of the fourth step depends ‘on a human operator's trained control because it is very 4ifficult to build a process model A steel-making plant produces various kinds of steel according to its ingredients. Especially the manganese ratio in the products is required to be variously adjusted. In this section we deal with the problem of determining the amount of manganese alloy to adjust the manganese ingredient of produced steel in step (2). This process is the most difficult to be controlled among the ingredients ad- justments, ‘We now list all the variables possibly concerned with the process. Mal Original ratio of manganese in input iron, mr 129 Fig. 22. Conversion process. fae ccontersee) [naa to (3) operator Fig 23, Controt of manganese ratio. Mn2_ Final ratio of manganese in produced steel. Mn2* Required ratio of manganese in produced steel. MA Ratio of manganese alloy put into input iron. HP Hot pig ratio of input iron, [0] ; Ratio of oxygen in input iron after oxygen blow. [Si] Ratio of silicon of hot pig, SG__ State of floating slag. Here input iron consists of hot pig and scrap. Among those, SG is physically measured only after the conversion process is finished. But it can be evaluated by operator's observation before that. Fig. 22 shows the conversion process. Manganese alloy is added into hot pig after the oxygen blow process to produce steel such that the percentage of ‘manganese ingredient, Mn2, becomes a required value Mn2* and its ideal amount can be calculated by (24), based on physical analysis. (Mn2* ~ Mni), (24) However, the actual ratio of manganese in the products is usually less than the physically estimated value from Mnl and MA because of absorption by slag or other reasons due to, for example, various other ingredients in steel, So a human operator predicts Mn2 by referring to {0}, [Si], HP, and observing the state of slag SG etc., and controls MA by correcting the guided value by (24) so that Mn2* is attained. Fig. 23 shows the present situation of manganese control. ‘The past trials to derive a mathematical model of the converter with respect to manganese control have not been successful where the inputs are the observable variables and the output is the manganese ratio of the produced steel For the purpose of control, we try to build a fuzzy model of the converter by finding some clues from an experienced ‘operator's way of control. This approach has the following, advantages. 1) His manner tells us how he recognizes the characteris tics of the conversion process based on his experience. For example we can know important variables that should be put into premises of fuzzy implications.130 EEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. SMC-15, NO. 1, JANUARY/FEBRUARY 1985 TABLE IV TABLE T T ™ chotce of prenise | pertornance | correlation io sectsbie i mae eesti 135.00 ——— 30a (eee) T5201 o.taprs 33.00 1 1 9-00 ae 1 ar | o.se10s {6-00 i | i.e 2 ss | 0.98640 19-00 I 1 3.00 i { 100 (stage 2) mmf kar Loses 38:00 i t tesa mem | anf ovens 3.00 { t Bio mensch | 4.8 | ovseane 1100 i L 100 IfMA = \ SG= wa so t 107 1fMA = << C SG= 3 16 I'MA = Z a SG oF i 4 Indicates 2 stable acace at each stage then AMn = —0.12HP + 7.77MA + 22.18 then AMn ~ 0.21HP + 9.03MA — 14.07 then AMn = 0.87HP + 4,92MA — 17.56 then AMn = —0.27HP + 7.34MA + 49.70 Fig. 24. Fuzzy model of converter 2) We can even use input variables that only he can ‘measure, for example, by just watching. Those variables are easily used as premise variables of our model. Since a ‘model is of the form ‘if. then +--,” the obtained model may be refined by his knowledge. 3) We can derive fuzzy control rules from the model, rather than from his control actions which may not be the best from a quantitative point of view. Needless to say, fuzzy control rules are easily understood qualitatively by him and we can adjust control rules also by his way of control. This is a very important point if an operator remains as a key essence in process control. Modeling: We have taken 61 operation data from among the ones obtained in one month, and have used them for the identification of the conversion process. Further, we prepared 20 testing data different from the above identifi- cation data to check the validity of the obtained model. Table TV shows some of the data. The input and output variables of the converter model are as follows [input] Hp = —hotpig_ _ __hotpig (percent) input iron ~ hot pig + scrap MA = Manganese alloy input iron %10~* (Percent) SG = indication about softness of slag output] ‘AMn = Mn2 ~ Mani = increment of manganese ratio. In this study, other variables [0] and [Si] are found not to seriously affect the process and so are deleted. For SG we conventionally use the measured values after the con- version is finished, which can be replaced by operator's observation. An experienced operator can measure SG rather qualitatively such as soft, medium, or hard, For this, reason we put SG only into the premises for conditioning the input-output relation and do not use it in the conse- quences, According to the proposed identification algorithm, each range of MA and SG is divided. The range of HP has remained undivided. Finally we have obtained four impli- cations as are shown in Fig. 24. The premise variables, the performance indices of the models, the correlation coefficients of the original output and models’ output through the identification process are shown in Table V. As is seen in the model, the space of each premise variable is divided only into two fuzzy subspaces. This is mainly because of the shortage of data. Notice that thereTABLE VI are five parameters in one implication: 4 x 5 = 20 al- together. On the other hand, the number of data is only 61. Results of Fuzzy Model: Table VI shows the results of the fuzzy model, those of a statistical model and converter utputs, when the models are applied for the testing data. The statistical model is represented in (25), whose parameters were obtained by linear regression using the identification data. AMn = —0.24HP + 8.64MA + 45.60. (25) ‘The performance indicies of the results of the fuzzy model and the statistical model for 20 testing data are as follows: fuzzy model 7.15 statistical model 7.77. The results are better than those obtained by a statistical model. It should be noted that the fuzzy partition of the state of slag, SG, derived from the data shows a good agreement with that by an operator: he usually recognizes SG accord- ing to a similar partition and uses this information in his control. Control of Converter: We now try to control the con- verter by using its model. Given a desired output AMn*, we can calculate a necessary input MA from a model. Here for simplicity we use this MA instead of designing a fuzzy controller. The problem is how to compare the results of the ‘model-based control with those of an operator, because we cannot make an experiment at present. Let us take as an index of control performance in* — AMn| AC = average of 4Ma” — Mal ‘AMa* where ‘AMn* desired output AMn — actual output, 13 As for an operator's control, this index is obtained from input-output data since, given AMn*, he controls a con- verter. Let us denote it ACise- In case of a model-based control, we assume that the output AMn of a converter is a desired output. Then we get the optimal input MA? from a model, input MA? to a process and see its output AM. This output AMIn can be estimated by taking into account of the accuracy of the model without experiments. So we can set AMn= AMn + where ¢ is the error of the model. Now we have |AMa ~ Wn} ACgouei = average of ST = average of error of the model. We obtain the following results: ACyosa = 4:7 percent AC ope = 6.7 percent. From the results we can expect that the control based on the obtained fuzzy model gives us better results than the present control by the operator. ‘Apart from the above method to calculate the input MA, we can directly derive fuzzy control rules in this case from the data (AMn, HP, SG) + MA. Those are shown in Fig. 25. V. Conetusion We have suggested a mathematical tool to fuzzily de- scribe a system. It has a quite simple form, but it can represent highly nonlinear relations as has been shown in examples. An algorithm of identification has also been shown and two applications to industrial processes have been discussed. The applications as well as illustrative examples show that the proposed method is general and thus very useful. We can put the results of fuzzy measure- ments by man such as “temperature is high” into the Premises of implications. Linear relations in the conse- quences enable us to easily deal with this mathematical tool, as we well know in linear systems theory. However, to claim the validity of the method, more studies have to be performed. The system theoretic ap- proach is especially important. For example we have minimal realization problems, decomposition problems, de- sign problems of controller, etc. To solve these problems, it is required for us to deal with fuzzy system representation just as we do with a linear system. In this paper, modeling of a human operator's control actions, rather than that of a process, has been mainly discussed. It is, however, possible to apply the method presented for the identification of a dynamical system. Modeling of a multilayer incineration furnace, a dynamic and distributed parameter system, is now under study.132 a a o [EEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. SMC-15, NO. I, JANUARY/FEBRUARY 1985, SGis NX then MA = 1.12HP + 0.104Mn* ~ 95.71 cy or x then MA = —0.21HP + 0.12AMn* + 16.53 o SGis SGis SN then MA = —0.01HP + 0.10AMn* + 0.91 SGis If HP is If HPis NL ItHPis _Z. ar Zo then MA = 0.19HP + 0.114Mn* ~ 17.92. It HP is <6 + Fig. 25. Control rules REFERENCES ‘Man-Machine Studies, vol. 12, pp. 3-10, 1980. [4] RM. Tong, “The construction and evaluation of fuzzy models,” in E. H. Mamdani, “Application of fuzzy algorithms for control of Advances to Fuzzy Set Theory and Applications, M. M. Gupta Ed. simple dynamic plant,” Proc. IEEE, vol. 121, no. 12, pp. 1585-1588, ‘New York: Plenum, 1979, 1976. [5] M. B. Gorzakzany, J. B. Kiszka, and M. S. Stachowice, “Some R. M Tong, M. B. Beck, and A. Latten, “Fuzzy control of the problems of studying adequacy of fuzzy models," in Fuzzy Set and activated sludge wastewater treatment process,” Automatica, vol. 16, Possibility Theory, R. R. Yager, Ed. New York: Pergamon, 1982. p. 659-701, 1980. [6] M. Sugeno and T. Takagi, “Multi-dimensional fuzzy reasoning,” P.M. Larsen, “Industrial application of fuzzy logic control,” Int J. Fussy Sets and Systems, vol. 9, no. 2, 1983