Sei sulla pagina 1di 5
Self-organising fuzzy logic controller B.S. Zhang J.M, Edmunds Indexing terms: Algorithm, Fuzzy logic controller Abstract: A frequently used learning algorithm, for self-organising fuzzy logic control has been analysed and found to be inadequate for some applications. The key problem is how to gain the credit value for control rule modification. A new ype of Tearning algorithm is proposed. Some supervision rules are given {o secure rule modifi- cations. Anew approach is derived to calculate the credit value for the rule generation and modifi cation. Simulation studies for controlling different process models and different model parameters sare demonstrated, 1 Introduction Fuzzy controllers have been successfully used for many years [1-5] However, there are some drawbacks to this design approach; first, a reliable linguistic model of the operator's control straiegy may not always be obtainable, and secondly some significant process changes may be outside the operator's experience. The applications of fuzzy controllers are thus limited. ‘An attractive approach to solving these problems is provided by the self-organising fuzzy logic controller, first proposed by Procyk and Mamdani [6], which has a learning algorithm and is capable of generating and modifying control rules based on an evaluation of the system's performance. The generation and modification of control rules is achieved by assigning a credit or reward value to the individual control action(s) that make a major contribution to the present performance, value is obtained from_a fuzzy algorithm ss the desired performance linguistically and form as the control algorithm of the generic fuzay logic controller. The learning algoithm proposed by Procyk and Mamdani [6] has since been adopted by others (7-11) However, the control rules may often be improperly ng this learning algorithm for the sell- “organising fuzzy logic eontroller when either the set-point ‘or the disturbance changes, This problem has been recog- nised [8, 9, 11]. A complementary learning algorithm is, 2 Problem descriptions ric fuzzy logic controller is implemented [12] as, X, xX) O Re wo Paper BRGGID (C9) fst rosie 18th July 1991 a in revised fo March 1992 The authors ate with the Control Systems Centee, UMIST, Manchester Meo 1QD, United Kingdom 19th 460 where Ry is the controller input/output fuzzy relation, X, and X, ‘are the fazztied process output error and its derivative, respectively, U, ts the fozzy output ofthe con- troller, x denotes the Cartesian product and © denotes the maximum-minimum compositional rule of inference When X, and X, are measured as fuzzy singletons (crisp values) as x,,&,, and Rp, together with the defuzai- fication procedue, isin the form of a look-up decision matrix R, then the defuzzitied controler output ean be expressed as 1M, = AX, &) 2) where 15, 5) 8 the element of R at the position (x, 8) x, and i, are quantised into a finite number of values these quantised values may not be tegularly spaced) so R has a finite number of elements. A selorganising fuzzy logic controller is achieved by evaluating the performance enfiancement matrix P [6] 5uy-a = oss Si) “oe where p(x,, %,) is the element of P at (x,, %), and d Alenotes the time delay from the control action to the process output response For the current state (x,, %,) dt, is a credit value to be added to the control action at the instant ¢ —d. The decision matrix is then conceptually modified as Ws Sd)! = oe Baa) + Og a This means that if the process state is (x,, %) at that instant ¢, a better state would have been attained if the previous control action at instant ¢ —d had been modi- fied by the amount 5u,..= p(x,, %,), and the modified control action should be remembered in the control rules for similar future occurrences. The symbol used, as in PASCAL, to emphasise that the righthand side is assigned (0 the left-hand side when updated. By using the above strategies the controller can start working, even from an empty rule base. Control rules are generated and modified by the learning algorithm evalu- ‘ating the performance enhancement matrix. And after several runs (or several changes of reference signal) the controller can give a good control performance. According to the performance enhancement matrix, this method aims at achieving a ‘good trajectory’ to ‘within a certain band where the rule base is not modified The good trajectory band is similar to the switching curve in time optimal control oF sliding-mode control, When the state of the process is outside this band, a nonzero credit value will be obtained according to the performance enhancement matrix. Consequently some control rules will be modified. Potentially this implies a drawback of this approach. According to above method, when the state is outside the good trajectory band the rule base will still be modified even if the state moves satisfactorily towards the good IEE PROCEEDINGS-D, Vol. 150, No.5, SEPTEMBER 1992 trajectory. Consequently, the control. performance may eventually deteriorate. ‘The problem that the learning algorithm of the self: organising fuzzy controller proposed by Procyk and Mamdani [6] is sensitive to external signals (set-point changes or disturbance) has been recognised [8, 9, 11] jome passive methods have been proposed. The method indicated by Daley and Gill [11] is to widen the band to make the modification procedure less active. Thus. the good trajectory band may no longer be good. The method used in Reference 9 is merely stopping rule modification when a good performance is thought to have been achieved. The controller then becomes a fixed fuzzy controller until the performance deteriorates to the point where rule modification is thought to be necessary. To certain extent, this may slow down the performance deterioration, However, it does not solve the problem because the correct and incorrect modification of the rules cannot be separated in this manner. Similar prob- lems have been encountered in self-tuning controllers. 3. Supervision of rule generation and modification To avoid the disadvantage of the self-organising fuzzy logic controller mentioned in the preceding section, we first need to establish whether or not the control rules should be modified. We can then obtain some super vision rules to govern the modification of the rules. In general, if the state is initially in the good trajectory band, the controller should be able to keep it within the band and move to equilibrium, If the state in the band is, driven to cither side beyond the band, the control rules should be modified. When set-point changes or disturbance occurs, the process siate will go outside the good trajectory band for certain time, the controller may force the state to 20 into the good trajectory band and then bring the state to. the equilibrium. If the controller cannot force the state to go into the band in a satisfactory manner, Then the control rules also need to be modified. If the controller drives the state so that it moves to the band sufficiently fast, the control rules do not need to be modified. To, simplify the analysis, we take the good trajectory as a switching curve (surface) and analyse in a phase plane (see Fig.1), weg) 0 | \ ois) Fig. 1 Phase plane analysis of system dynamic performance Suppose the state is at A(x, x3) where x3 = &,, the next position is at B(x, + Ax,, %; + Ax). The movement of the state from position A (0 position B gives informa- Ton about the system dynamies. We can use the value AX.JAx, to describe the information and the manner of the state when moving towards the switching curve. A general qualitative fact about this ratio is that for a given (si. ¥2h the absolute value of the ratio increases as the corresponding control action increases, hence the speed IEE PROCEEDINGS-D. Vol. 139, No.5, SEPTEMBER 1992 of the state as it moves to equilibrium. More specifically, lel. 0 = W(x,, x3) =0 be the switching curve, when ox, < 0... the State is in the area moving towards the siwiieh- ing curve when there is no control action, if Ax,/Ax, > lj. X,) > 0 for some specified g,(x4, x3) thet & means That the state moves towards the switching curve in a satisfactory manner because the point moves towards the switching curve fast enough. No matter how distant thi point is from the switching eufve usually the rule base is modified according to this distance), the rule base should fot be modified. Similarly, when vx) > 0, ie, the state is The area moving away from the switching curve when there is no control action, if Axy/AX, < q3(¥). ¥3) <0 for some specified 4,(x,, x3) the state will move towards the switching curve satisfactorily. In this ease the control rule should not be modified as well, We can therefore con- struct following pair of rules for not modifying the rule Thase. If rx, <0 and Ax,/Ax, > q, > 0 then make no. modification. If ex, >0 and Axs/Ax, 0 then m= ~g, + Axa/Axy Hm<0 then make no modification The modification is needed if m > 0 and the amount of modification depends on the value of m, which gives the information on the deviation of the direction of the movement of the process state from the specification, The credit value is generated by Btu, g = SAL (tgs m= 580 ()) 0 where J, is a coefficient whose value controls the learning speed and the coarseness of rule modification, sgn(- ) is a signum function and sat (a, x) stands for a saturation function axa a=} x Iisa 6 -a x<-a Since m is determined by the deviation of the direction of the state motion instead of the position of the state, the rales can also be modified according to m when m <0, When the system is near its equilibrium, the value Ax,/Ax, may be unreliable. In this area, m can be caleu- lated by m =v + eAv, where 6 > 0. aot For actual implementations, the measurements are quantised into 2N +1 levels, ranging from —N to N. ‘The rule matrix can then be expressed as that with the elements r(x), i), where ij, € (1, 2,-1.4 2N + 1}. The rule modification can be expressed a Hix, 4) + i, HR 4) + 8): Sat rae Sg) + 8h Bo) 4 8) wy5H-g) (7) where df and 6) = —1,0, 1; (4) + dé and j..) + 6) € {12.005 2N + 1} wy Weighting factor L f (14 15D + 15D. This means that the learned signal is used to modify the control rules spreading around the measured values. This procedure is equivalent to fuzzifying the measurements ‘and producing fuzzy control rules. The fuzzy sets have a membership function like yp(2) = 1/(1 +|2— 2;)). The fuzzified control action is similar to the result calculated by the product-rule for fuzzy implication [12]. It should be noted that although it also works when crisp sets are used, the learning is quicker and the control is smoother when fuzy sets are used 6) 5 Controller parameter selection In the design of a fuzzy controller the measurements are frst of all normalised into a certain range [12] and are often discretised into a finite number of leveis. When the measurements have been normalised the rest of the design of the controller could be a standard procedure. The selection of the scaling factors has been studied [6, 7, 11], Here we briefly describe the method we used. Let e be the measured error between the set-point and the process output, u, be the process input, and if the process input is the output of the controller, then ke xs he o 4, = KAU) sla) where k,, ke, ky are sealing factors and Fix.) d(x) is the weighted output from the control rules. If the normalised range is [—L, LJ, the guideline for determining the scaling factors is as follows ky < LMle laws KS fle lane (19) iy < Me man! If the maximum values of the measurements are not Known a priori, then scaling factors are chosen so that ke, keé € [—L L] is usually satisfied, the values outside the range are limited [12] The value of L is not meaning- ful, it only affects the value of the quantum for the quan- tisation (discretisation). Therefore we chose L = 1, The ‘quantisation is achieved as sat (N, 1+.N + qntzr (Nx/L)) ssat (N, 1-4 N+ qntze (Nxz/L)) ay where qntz() denotes a quantisation function with unity quantum so that iox,) and jlx,) € (1, 2..-, 2N i 1}. The value of the control actions stored in the rules fe not quanti ‘Suppose that the switching curve is chosen as v = q,%, 4.x; =0, where q, > 0, the design parameters q, and q3 462 should satisfy 42<~4, <0< 0) (12) To make the control trajectory closer to time optimal control or to consider the nonlinearity of the controlled process, q, and gz should vary according to the position Of the states, ie. 9} = gilt Xah da = daly, x2) However, even if q, and q, are kept as constants, the control system can still work well. And qy(x,, x2) and qa(xy, x3) can be combined to form a trajectory matrix Q with elements a(x, %2) since they are in nonoverlapping subspaces. To incorporate the fuzziness of the controlled process, the switching curve can be replaced by the good trajectory band featured by the performance enhancement matrix, ie. v= ply x3) if necessary. 6 Convergence and stability To ensure the convergence of the rule modification and the stability of the control system, itis esssential to select the desired good trajectory and the trajectory matrix Q aller the scaling factors of the controller's input/output Signals have been chosen. First, the good trajectory should be attainable. Taking the input/output saturation into account some trajectories may not be attainable. If the trajectory is not attainable, the control rules will always be modified. However, when the good trajectory is unattainable, it is associated with regular switching rather than mode-sliding. The system can stil be stable with a certain level of overshooting for the step response Secondly, the matrix @ cannot be chosen arbitrarily. The choice of the value of the matrix Q should ensure that the speed of the process state moving towards the good tra jectory is also attainable. Fortunately, the suitable region of these parameters is very wide, and the choice of these parameters is very fexible. 7 Examples ‘To demonstrate the applicability of the self-organising fuzzy logic controller proposed above, we give a simula tion study of some of the examples. The first. process model is a double integrator. The second model is an integrator with a time lag, The third is a time lag. The fourth is a time lag with a pure time delay. The param- eters of the process models vary within bounds, The four models and the parameter values are chosen as shown in Table 1. ‘Table 1: Controlled process mod Type Model Parameter model * k=10~100 Model 2 A «=1.0~100,7=20~50 lodel? Sass) k=1.0~100,7=20~50 Teri Model 4 0,7=20,T,-08 teat For all simulations, the design parameters of the con- troller such as scaling factors, switching curve, qi. 42» bw fle, are same and remain unchanged. The controller ‘output for the third and fourth model is the rate of change of the process input instead of the process input Tor the first (wo models. For each simulation, the control- ler starts from an initially empty control rule base. IEE PROCEEDINGS-D, Vol 139, No.5, SEPTEMBER 1992 For the purpose of demonstrating the adaptability and robustness of the proposed controller, linear controller is also used for comparison. The parameters of the linear controller are designed for the process parameters at T =2,k =10 to give reasonable performances The design parameters. for the controller are as follows. The maximum of the process input is supposed 10 be tag = 1.5; the scaling factors are chosen as: k, 1k, = irk, = 0.1. The maximum value ofthe elements in the'R matrix is raag = 15. The parameter in the switching curve is q,= 1. "The parameters in the selCorganising algorithm’ are: q, =0.2(1 + |xyIM001 + 1x3), a2 “5, dig = 0.5, ly = 0.1, Each normalised measurement is quantised into [5 levels (N = 7), The sampling interval isT.=01 First we demonstrate the learning ability of the con- troller when it start working from an empiy control rule base t0 control the first thres models. Figs. 2 and 3 show al _ ‘0 80100180" 700-780-308 380 Zoo times Fig. 2. Syuare wave responses of selorganiing fussy control system, Contaler tats from empty contol rebate 10,7 =20 Proceso af ode 2 roses cutat of mode ey 4 “30100 180200 250 30 me LTS erst— eae aeons cor a se =A process outpet of model 1 Senet E=eeoos: 300360 400 the process output responses to a square wave signal for low process gain (k= 1) and for high gain (k= 10), respectively. For both situations the system works well and eventually gives similar transient responses. We can. also see that the performance convergence speed for the Jow gain process is not as fast as for the high gain process. This is because for the low gain process the con- troller needs to produce a large signal (analogous to large controller gain) but the learning pace is limited (by dup, and |). Therefore it takes longer to accumulate the required control signals as well as to produce satisfactory responses, If we increase the values of d144 and ly the con- vergence for controlling the low gain “process will be improved and a performance similar to the high gain process can be obtained. To look at the performance after the initial learning period from time 0 to time 300, we examine the responses in the interval 300-400. Figs. 4-7 show the square wave responses of the sell-organising fuzzy control systems and sd gain linear control systems for the first three IEE PROCEEDINGS.D, Vol. 139, No.5, SEPTEMBER 1992 models, respectively. Fig. 4 shows the responses for the first model with differnt process gains. Fig. 5 shows those for the second model with different time constants, Figs. 6 and 7 show the responses for the second model with different time constants, and for the third model with dif- ferent time constants, respectively. From these compari sons we can sce that the variation of the responses of the self-organising control for different models and param- clers is smaller than that of the responses of the linear control for the same changes in the models and param- eters of the controlled process. This indicates the adapt- oF = j mt 300" 3040-380 3i0-aE0 3630303800 times io. Compare of srpesng fry contro and eer conrel forwoces nok} tom spt by erp coil t= 10 foe pay fase ua PD somk = 100 f YS ~ 300310” 380-330" 30, 380 360310380380 HOO Fig. 5 Comparison ofself-organising fizzy control and linear control for process model 2 proces cul by e-rgaiingcomrol T= 20 Froese by corpnung control T = 50 Froce ouput by PD- control P= 20 tty-— \ ¥ol —_, 4 SS 300310 320-340 340-380 360° 370 380380 Go time © Fig. 8 Comparison of seorgonisng fay control and linear control Jor processmael? ra20 oxen ott by tebrging con. = 19 <== Fos ont selergnig conta t= 100 locos cut by PDsconte t= 18 Fico cua y contrat = 100 300310 330” 340-340-380 360 340380340 ZOO Fig. 7 Comparison of selorganisng fuzzy control and linear control for process model 3 keto proces ouput by self organising convo T 2 proces utp by sllrganting contol T= 80 Proce onal by PL conv = 20, rose output by PI conto. T= 80 463 ability and robustness of the proposed self-organising controller. In the next example we compare the performance of the previous selforganising control method with that of the proposed one. The process model is the fourth model in Table 1. Because of time delay in the model, instability or limit eycles may ovcur as the loop gain increases. To control such a process, the previous self-orgainsing control method may sulfer from the eyclic phenomena of control response as indicated in Reference Il. Fig. 8 Hep E 080 160° 180-200-250 300 380 400 time, & Fig. 8 Control performance hy previous seltorganising controler for process model # wpross inp shows the process output and control responses by the previous selF-organising method and a performance enhancement matrix similar to that used in Reference 11 We could see that at the beginning the performance was improved as self-learning, but eventually deteriorated due to improper learning and modification of the control rules, Fig 9 shows the process output and control 2 0-80 100 180 260 280 300 380 400 Fig. 9 Contra perfrmance by proposed stéamganising controller for process model 4 {pecs apa responses by the method proposed herei that it performs satisfactorily. It is obvious 8 Conclusions As part of the design of the self-organising fuzzy logi controller, the method proposed by Procyk and Mamdani [6] has been analysed. A complementary algo- rithm for self-learning has been proposed and demon- strated. First, a set of supervision rules makes it possible to avoid certain improper modifications. This may be taken as a safety net, Secondly, the credit value for rule modification is determined by the direction of the move~ ment of the process state instead of the position of the process state. This seems more reasonable. The learning strategy of the controller is closer to making the system implicitly follow a dynamic model, so it is more like a model reference control strategy. It differs from the usual model reference control method in that the control actions for different process states are stored in a network-like decision matrix which may possess high nonlinearity. Hence it may particularly be suitable for controlling processes which contain uncertainty or non- linearity. Another feature of the proposed sell-organising fuzzy control method is that the dynamic performance of the system is determined by the parameters of the trajec- tory matrix Q and switching curves which can be designed qualitatively by using mode-sliding control theory, 9 References 1 MAMDANI, EH. and ASSILIAN, S.:"An experinent i linguistic sythesis with a fay loge controller, In. J. Man-Machine Stud pp Ld RT, WIM, and VAN NAUTA LEMKE HR. “Applic on of fzzy logic controller ina water plant’ Automatica, 1976, 12 5p 301-308 KING, PJ, and MAMDANI, EL: ‘The application of fazey conteol systems to industrial proceses, Automatica, 1977, 13, pp. Sean LARSEN, P.M: ‘Industrial applications of fuzzy logic conto nt J Man-hiachine Stud, 1979, 1 pp. 67-70, 5 TONG, RIM."'A conteol engncering review offurzy systems’ Autor rata, 1977, 13, pp. 889-509 @ PROCYK, 1, and MAMDANI, E1.:°A linguistic selCorganising ~ process controler’, Automatica, 1979, 1, pp. 13-30 7 YAMAZAKI, T, and MAMDANI, EH. "On the performance of fule-based sei-organising controlle’. EEE Conference om Applica ons of Adapiive and Multivariable Control, Mull, UK, 1983, pp sss & SUTTON, R, ROBERTS, G.. and FOWLER, P.: Scope and lin tations of fuzzy sell-onganising controle for warship stabilisation” Proceedings of International Conference on Simulation and Contvol of Marine Craft, Fueter, UK, 1990 SD FARBROTHER, H.N, STACEY, BA, and SUTTON, R ‘Fuzzy seltorgaising control of a remotely operated submersible. TEE ernatianal Conference on ‘Control, Edinburgh, UK, 1991, 9p. S04 10 LINKENS, DA. and HASNAIN, S.B- ‘SelForganising fuzz loge Control and application to muscle relaxant anaesthesia TEE Proc 1, 1991, 138, 3} pp. 274-284 11 DALEY, 5, and! GILL, K.P. °A design study of « se-organsing fazay logic controler Proc Pain ec Engrs, 1986, 200, pp 39-9 12. ZHANG, BS. and EDMUNDS, JM. 'On fozzy log controllers TEE International Conference on Contcol, Edinburgh, UK, 19. pp. 961-965 IEE PROCEEDINGS-D, Vol. 19, No.5, SEPTEMBER 1992

Potrebbero piacerti anche