Fuzzy Model Reference Learning Control with Modied Adaptation Mechanism

Otto Cerman and Petr Huek s

Department of Control Engineering FEE, Czech Technical University in Prague Technick 2, 166 27 Prague 6, Czech Republic a Email: [email protected], [email protected]
AbstractIn this paper dierent approach to adaptation mechanism for fuzzy model reference learning control (FMRLC) method will be introduced. In comparison to original method the proposed procedure shows larger robustness to responses to initial conditions and various reference signals. The main idea consists in adaptation of the inverse model an important part of FMRLC. Moreover, in order to eliminate abrupt changes in the process of adaptation, the shape of the surface of the fuzzy controller is kept similar during whole control process. The advantages of the proposed modication are presented on control of magnetic suspension system. Index Termsfuzzy control, adaptive fuzzy control, intelligent control, fuzzy model reference learning control

I. Introduction In many situations complex systems can be controlled successfully using fuzzy controllers. Fuzzy controllers based on theory of fuzzy logic and fuzzy sets represent successful tool for converting a linguistic control strategy obtained from knowledge of an expert into a mechanism capable to control real plants. Unfortunately, a lot of time is needed for nding the rules because they are often searched by a trial-error method. In addition they are found out with diculty for complex systems. Moreover, in many practical applications the operating conditions vary in time that requires an automatic adaptation of the controller rules. If the fuzzy controller uses to some extent experience from the past it becomes selforganizing or self-learning controller. Self-organizing, self-structuring or self-learning controllers belong to the group of adaptive fuzzy controllers. They can be seen as heuristic controllers in which control rules are generated and automatically improved. Their basic functions are: 1. to generate suitable control signal according to evaluation of controlled system behaviour, 2. to modify controller in a suitable way to reach the required behaviour even during changes of operating conditions of controlled system. The rst self-organizing fuzzy controller was developed and described in [1]. The idea has been further elaborated by many authors, especially in [2]. Unfortunately the approach did not become successful because the design becomes very dicult for tracking reference signals dierent

from step signal. The FMRLC algorithm eliminating that drawback was for the rst time introduced in the work [3]. The main idea is based on the conventional model reference adaptive control (MRAC) [4]. An approach to adaptation ensuring convergence of the controller rule base in FMRLC is analyzed in [5]. Other self-learning or adaptive fuzzy control methods are described e.g. in [6], [7] or in [8]. A method dealing with a special case of MIMO system can be found in [9] and [10]. Self-learning approach is used in wide range of applications, e.g. in biomedical engineering [11], [12], in power engineering [13]], in automotive industry [14] or for actuator mechanisms [15]. Other areas of utilization selflearning approach in control theory one can nd e.g. in [16], [17] or [18]. Possibilities of implementation of selforganizing controllers in PLC are analyzed in [19]. Among the self-learning concepts the FMRLC seems to have the greatest potential that is supported by some successful applications [20], [21], [16], [22], [23], [24], [25], [26]. In this paper an innovated adaptation method based on adaptation of the inverse model is presented. II. FMRLC In this section a short description of the principle of the FMRLC will be given. The principal scheme of the FMRLC is shown in Figure 1. It has four main parts: the plant, the fuzzy controller to be tuned, the reference model, and the learning mechanism (an adaptation mechanism). Based on the dierence between the output of the reference model ym (kT ) and the output of the plant y(kT ) the learning mechanism evaluates current performance and automatically synthesizes or adjusts the fuzzy controller so that the dierence goes to zero. The controller itself is usually implemented as a fuzzy PD controller. The membership functions covering the input space of the controller (the antecedents of the rules) are xed whereas the location of the corresponding xedshaped output membership functions (the consequents of the rules) is adjusted by the learning mechanism. The learning mechanism consists of two parts: a fuzzy inverse model and a knowledge-base modier. The fuzzy inverse model transforms the deviation from the desired

Fig. 1.

Diagram of the FMRLC (gure taken from [27])

behavior ye (kT ) to changes in the process inputs p(kT ) that are necessary to force ye (kT ) to zero. Knowledge-base modier brings this change into eect by shifting by p(kT ) the output membership function of the rule that acted with the highest strength at the time instant kT Td where Td is the delay of the plant. In both the controller and the inverse model minimum inference engine and center of average defuzzier are used. More details can be found in [27]. III. FMRLC with New Adaptation Mechanism Although there are many successful applications of original FMRLC the method has several disadvantages. Firstly, the rule base of the fuzzy controller is strongly inuenced by the initial condition and reference signal trajectory that is not acceptable in practical applications. Moreover, original FMRLC does not work well in case that the inverse model is not correct. The purpose of the paper is to modify the original adaptation procedure such that the nal fuzzy controller preserves the control surface for various reference signals and dierent initial conditions. In addition to that the proposed method adapts the inverse model. The FMRLC usually starts with a rule base with the centers of all output membership functions located at zero. The proposed modication relies on one of two following alternatives for initial fuzzy rule base. The rst one uses fuzzy inverse model of the plant as a starting rule base for the fuzzy controller. The second one consists in design of a fuzzy controller whose structure is able to successfully control the plant. In many practical applications fuzzy PD controller with uniformly distributed output membership functions proved to be sucient as an initialization of the rule base. Note that the proposed method uses singletons as output membership functions. In order to keep the shape of the controller surface similar during whole process it is checked after each step whether the shift of an output singleton caused by the

actual penalty p(kT ) does not violate the sequence of the output membership functions. If adaptation requires shift of an output singleton to a position that would change the sequence of the output singletons the penalty will be limited to the maximum admissible value. Beyond keeping similar controller surface the proposed adaptation mechanism makes in contrast to original procedure possible to keep reasonably small number of output membership functions that makes the rule base of the fuzzy controller more transparent and interpretable. Moreover, for a small number of output membership functions a small value of the gain gp is sucient for quick adaptation. Let us note that small value of gp decreases risk of instability of the adaptation mechanism. Moreover, the proposed method contains adaptation mechanism for the rules in the inverse model. The question is how to recognize that the output of a rule in the inverse model (the penalty p(kT )) is not correct and it should be adapted. The proposed algorithm proceeds in such a way that if the consequence of a rule is a worse reference signal tracking from the control error e(kT ) and its dierence ce(kT ) point of view the output of the rule is modied. When a rule in the controller is activated again (after time interval Tm ) then values of relative change of the error em (kT ) = e(kT ) e(kT Tm ) and relative change of the error dierence cem (kT ) = ce(kT ) ce(kT Tm ) are tested. In case that the actual values of em (kT ) and cem (kT ) are both positive or both negative the consequent of the rule in the inverse model is changed. For this purpose a new fuzzy system for modication of the inverse model is established that maps em (kT ) and cem (kT ) to the necessary changes in the inverse model output p(kT ). Note that similarly to the fuzzy controller and the inverse model this fuzzy system contains scaling gains gme , gmce and gmu . The input membership functions cover the whole input space of the fuzzy system. A part of the rules in the fuzzy system for the modication of the inverse model is depicted in Table I. The fuzzy system has 3 input triangular membership functions uniformly spaced on the interval [1; 1] on both inputs. The membership functions are labelled by linguistic values Negative, Zero and Positive. On the output there are 9 singletons uniformly spaced on the interval [1; 1] labelled by linguistic values Negative Big, Negative Medium, Negative Small, Zero Negative, Zero, Zero Positive, Positive Small, Positive Medium and Positive Big. The fuzzy system for modication of the inverse model has the rules of the following form: If em is E j and cem is C k Then um is U l where em and cem denote the linguistic variables associated to controller inputs em (kT ) and cem (kT ), respectively, um denotes the linguistic variable associated to the fuzzy system output um (kT ) and E j , C l and U l are fuzzy sets. To get the nal value of pm (kT ) for modication of

TABLE I Rules in the fuzzy system for the modification of the inverse model mf e(kT ) Negative Negative Negative .. Positive Positive Positive mf cem (kT ) Negative Zero Positive .. Negative Zero Positive output mf Negative Big Negative Medium Negative Small .. Positive Small Positive Medium Positive Big

V. Comparison of both FMRLC methods For simulation we use the values in Table II. Common parameters of both methods are summarized in Table III. Both methods use 5 triangular membership functions uniformly spaced on the interval [1, 1] on each input in the controller and 5 singletons initially distributed uniformly on the interval [30, 30] on the output corresponding to the allowable range of control signal. For the inverse model and the fuzzy system for modication of the inverse model we use 7 and 3 triangular membership functions uniformly spaced on the interval [1, 1] on each input, respectively, and 7 and 9 singletons distributed uniformly on the interval [5, 5] and [1.5, 1.5] on the outputs, respectively. Transfer function of the reference model was chosen with regard to dynamics of the controlled system 100 as Ym (s) = s2 +20s+100 . R(s) A. Comparison of the controllers before and after adaptation At rst let us compare both methods in term of the controller surfaces before and after adaptation. The starting controller surfaces before adaptation takes eect are the same, see Fig. 3. The surface of the inverse model is depicted in Fig. 4 a). In Fig. 5 one can see that regulation without adaptation is very poor.
Centers of the output mf []

the rule in the inverse model the output um (kT ) from the fuzzy system is multiplied by the value sign(p(kT Tm )). This fuzzy system brings the change into eect by shifting the output membership function of the rule in the inverse model by pm (kT ). IV. Magnetic suspension ball system Original FMRLC and FMRLC with new adaptation mechanism will be compared on magnetic suspension ball system, see Figure 2. The nonlinear third order system is described by the following state space equations: dx1 (t) = x2 (t) dt dx2 (t) x2 (t) = g 3 dt M x1 (t) R 1 dx3 (t) = x3 (t) + v(t) dt L L y(t) = x1 (t)

20 0 20 2 0 ce [m/s] 2 0.1 0.05 0 0.05 e [m] 0.1

(1)

where y(t) is the ball position [m] (output of the system), M is the mass of the ball [kg], g is the gravitational acceleration [ms2 ], R is the winding resistance [], L is the winding inductance [H], v(t) is the input voltage [V] and i(t) is the winding current [A].

Fig. 3.

Initializing controller surface

The results of the original FMRLC and the proposed FMRLC method after adaptation are shown in Fig. 6, the improvement is apparent. The corresponding controller surface of the original FMRLC is depicted in Fig. 4 c).
TABLE II Parameters in the model setup M 0.05 kg g 9.81 m/s2 R 60 L 0.6 H vmin -30 V vmax 30 V

TABLE III Values of design parameters of both FMRLC methods Fig. 2. Magnetic ball suspension system ge 0.1 gc 1.5 gu 30 gy e 0.1 gyc 1.5 gp 5 gme 0.05 gmce 1 gmu 1.5

Centers of the output mf []

Centers of the output mf []

a)

b)

0.18
5 0 5 2 0 0.1 0 2 0.1

5 0 5 2 0 0.1 0 2 0.1

r ym y original FMRLC method y the proposed FMRLC method

0.16 Distance y [m]

0.14

ce [m/s]

e [m] Centers of the output mf []

ce [m/s]

e [m]

Centers of the output mf []

c)

d)

0.12

20 0 20 2 0 0.1 0 2 0.1

20 0 20 2 0 0.1 0 2 0.1

0.1

0.08 0 1 2 3 4 Time [s] 5 6 7 8

ce [m/s]

e [m]

ce [m/s]

e [m]

Fig. 4. Controller surfaces: a) inverse model for both methods, b) inverse model after adaptation c) controller surface after adaptation for original FMRLC d) controller surface after adaptation for the proposed FMRLC method

Fig. 6.

Comparison of original and modied FMRLC

0.22 0.2 0.18 Distance y [m] 0.16 0.14 0.12 0.1 0.08 0 5 10 Time [s] 15 r ym y 20

Fig. 7. Inverse model: a) rule base of the initializing inverse model b) rule base of the inverse model after adaptation with highlighted rules which were adapted; numbers in cells are centers of output membership functions

Fig. 5.

Control without adaptation

preserves the sequence of the output singletons. One can see that this surface considerably diers from the original controller surface depicted in Fig. 3. A controller surface of the proposed procedure after adaptation is shown in Fig. 4 d) which is more similar to the original one. The surface of the inverse model after adaptation is depicted in Fig. 4 b), the corresponding rules are shown in Fig. 7. In Fig. 7 a) initializing inverse model is depicted. The shadowed cells in Fig. 7 b) show the rules which were adapted. The output membership functions before and after adaptation by the proposed FMRLC method shown in Fig. 8 a) and b), respectively, conrm that the nal controller rule base of the FMRLC with new adaptation mechanism B. Inuence of initial condition on the controller surface Next inuence of the initial conditions on controller surfaces of both methods will be discussed. Let us compare both methods for two initial conditions: 1) y(0) = 0.25 m, 2) y(0) = 0.06 m, A.
dy(t) dt (0) dy(t) dt (0)

= 0.2 ms1 and i(0) = 0 A = 0.25 ms1 and i(0) = 0.25

In Fig. 9 we can see that both FMRLC methods work well for various initial conditions. The corresponding controller surfaces after adaptation are shown in Fig. 10. One can see that the controller surfaces for the proposed FMRLC method are almost identical whereas those for original FMRLC dier considerably.

Fuzzified value []

E1

E2

E3

E4

E5

Centers of the output mf []

a)

Centers of the output mf []

a)

b)

20 0 20 2 0 0.1

20 0 20 2 0

0.5

ce [m/s]

0.1

e [m]

ce [m/s]

0.1

0.1

e [m] d)

30

20

10

0 b)

10

20

30
c) Centers of the output mf [] Centers of the output mf []

Fuzzified value []

E1

E2

E3

E4

E5

20 0 20 0.1 0.05 0 0.05 2 0 0.1

20 0 0 0.1 0.05 0 0.05 0.1 2 ce [m/s] 2

0.5

20

30

20

10

10

20

30

e [m]

ce [m/s]

e [m]

Fig. 8. Distribution of the output singletons of the proposed method FMRLC: a) before adaptation b) after adaptation

Fig. 10. Inuence of initial conditions on controller surface: a) initial condition (1) using original FMRLC b) initial condition (2) using original FMRLC c) initial condition (1) using the proposed FMRLC method d) initial condition (2) using the proposed FMRLC method

a) 0.25 Distance y [m] 0.2 0.15 0.1 0.05 0 0.09 Distance y [m] 0.08 0.07 0.06 0
r y
m

r ym y original FMRLC method y the proposed FMRLC method

0.15 0.14

0.5

Distance y [m]

1 Time [s] b)

1.5

0.13 0.12 0.11 0.1 0.09 0.08 0.07

y original FMRLC method y the proposed FMRLC method

0.2

0.4 Time [s]

0.6

0.8

0.06 0.05 0 2 4

r ym y original FMRLC method y the proposed FMRLC method

Fig. 9. Inuence of initial condtitions: a) initial condition (1) using original FMRLC and the proposed FMRLC method b) initial condition (2) using original FMRLC and the proposed FMRLC method

6 Time [s]

10

12

Fig. 11. Reference signal tracking using nonadaptive fuzzy controller

C. Nonadaptive fuzzy controller with rule base originated from FMRLC methods Let us compare ability of fuzzy controllers obtained by adaptation during tracking step reference signal (reference signal is depicted in Fig. 5) to track dierent reference signal without further adaptation. The controllers obtained using original and modied FMRLC with control surfaces shown in Fig. 4 c) and d), respectively, are compared in Fig. 11. The comparison conrms better behaviour of the proposed FMRLC method. VI. Conclusion In the paper modied adaptation for FMRLC method was presented. The presented procedure has several advantages comparing to the original one. Firstly, the controller rules obtained using the proposed approach of adaptation are less inuenced by the initial conditions and various reference signals. Next, modied FMRLC keeps the number of output membership functions small in contrast to original FMRLC. The same adaptation rate can be reached by smaller adaptation gain that decreases risk of instability. Moreover, the proposed FMRLC method modies the inverse model to improve its accuracy. The simulation results show that reference signal tracking using the proposed method is improved apparently. The presented method overcomes the original one also in the case when the fuzzy controller is applied without further adaptation.

Acknowledgment This work has been supported by the project of the seventh framework program no. 228346 LearnForm sponsored by the European Union. References
[1] E. Mamdani. and T. Procyk, A linguistic self-organizing proccess controller, Automatica, vol. 15, pp. 1530, 1979, Pergamon Press. [2] S. Shao, Fuzzy self-organizing controller and its application for dynamic processes, Fuzzy Sets and Systems, vol. 26, pp. 151 164, 1988. [3] J. Layne and K. Passino, Fuzzy model reference learning control, in IEEE Conference on Control Applications, 1992, pp. 686691. [4] K. strm and B. Wittenmark, Adaptive Control, Second EdiA o tion. Addison-Wesley, 2008. [5] O. Cerman and P. Huek, Fuzzy model reference learning cons trol with convergent rule base, in IFAC Workshop on Intelligent Manufacturing Systems, 2010, pp. 8691. [6] D. Driankov, H. Hellendoorn, and M. Reinfrank, An Introduction to Fuzzy Control. Springer, 1996. [7] J. Blanco, Hybrid self-learning fuzzy PD+I control of unknown linear and nonlinear systems, in Proceedings of the Fifth Mexican International Conference in Computer Science. IEEE Computer Society, 2004, pp. 233240. [8] Z. Kovaic and S. Bogdan, Fuzzy Controller Design - Theory c and Applications. CRC-Taylor and Francis, 2006. [9] Q. Lu and M. Mahfouf, Multivariable self-organizing fuzzy logic control (SOFLC) using a switching mode linguistic compensator, in IEEE Conference on Intelligent Systems, 2006, pp. 243249. [10] Y.-T. Kim and Z. Bien, Robust self-learning fuzzy controller design for a class of nonlinear MIMO systems, Fuzzy Sets and Systems, vol. 111(2), pp. 117135, 2000. [11] J. Shieh, M. Fu, S. Huang, and M. Kao, Comparison of the applicability of rule-based and self-organizing fuzzy logic controllers for sedation control of intracranial pressure pattern in a neurosurgical intensive care unit, IEEE Transaction on Biomedical Engineering, vol. 53(8), pp. 17001705, 2006. [12] Y. Hao, L. Feng, R. MacArthur, J. C. D. Barth-Jones, and Y. H. L. Crane, A self-learning fuzzy discrete event system for HIV/AIDS treatment regimen selection, IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics, vol. 37(4), pp. 966979, 2007. [13] N. Khaehintung, C. Kangsajian, P. Sirisuk, and A. Kunakorn, Grid-connected photovoltaic system with maximum power point tracking using self-organizing fuzzy logic controller, in International Conference on Power Electronics and Drives Systems, 2005, pp. 517521. [14] C.-M. Lin and C.-F. Hsu, Self-learning fuzzy sliding-mode control for antilock braking systems, IEEE Transactions on Control Systems Technology, vol. 11(2), pp. 273 278, 2003. [15] Wai, Rong-Jong, Lin, Chih-Min, Hsu, and Chun-Fei, Selforganizing fuzzy control for motor-toggle servomechanism via sliding-mode technique, Fuzzy Sets and Systems, vol. 131, no. 2, pp. 235249, 2002. [16] D. Ling, Fuzzy model reference learning control for a nonlinear model of hydrogenerator unit, in IEEE International Workshop on Intelligent Systems and Applications, 2009, pp. 14. [17] T. Orlowska-Kowalska, M. Dybkowski, and K. Szabat, Adaptive sliding-mode neuro-fuzzy control of the two-mass induction motor drive without mechanical sensors, IEEE Transactions on Industrial Electronics, vol. 57(2), pp. 553 564, 2010. [18] P. Patre, S. Bhasin, Z. Wilcox, and W. Dixon, Composite adaptation for neural network-based controllers, IEEE Transactions on Automatic Control, vol. 55(4), pp. 944 950, 2010. [19] S. Bogdan, Z. Kovacic, and D. Krapinec, Sensitivity-based selflearning fuzzy logic controller as a PLC super block, in Proceedings of Mediterranean Conference on Control and Automation, 2007.

[20] Y. Yuan, Y. Feng, and W. Gu, Fuzzy model reference learning control for aircraft pitch autopilot design, in Proceedings of 8th International Conference on Control, Automation, Robotics and Vision, vol. 3, 2004, pp. 1923 1927. [21] J. Layne and K. Passino, Fuzzy model reference learning control for cargo ship steering, IEEE Control Systems Magazine, vol. 13(6), pp. 2334, 1993. [22] M. Sheppard and M. Tarbouchi, Design and implementation of a stable fuzzy model reference learning controller applied to a rigid-link manipulator, in IEEE International Conference on Industrial Technology, 2004, pp. 11861191. [23] H. Rehman and W. Mahmood, A fuzzy model reference learning controller based direct eld oriented control of induction machine, in IEEE International Conference on Engineering of Intelligent Systems, 2006, pp. 488492. [24] S. N. Ndubisi, A fuzzy model reference learning controller for synchronous generator terminal voltage control, European Journal of Scientic Research, vol. 24, no. 3, pp. 378386, 2008. [25] A. Naceri, Y. Ramdani, and H. Bounouna, A fuzzy model reference learning controller of asmes to improve transient power system stability, Mediterranean Journal of Measurement and Control, vol. 3, no. 3, pp. 126133, 2007. [26] S. R. Munasinghe, M.-S. Kim, and J.-J. Lee, Adaptive neurofuzzy controller to regulate UTSG water level in nuclear power plants, IEEE Transactions on Nuclear Science, vol. 52, no. 1, pp. 421429, 2005. [27] M. Passino and S. Yurkovich, Fuzzy Control. Addision-Wesley, 1998.