Sie sind auf Seite 1von 5

Proc. of 1997 American Control Conference, pp. 3301-3305, Albuquerque, NM, June 4-6, 1997.

STABILITY ANALYSIS OF A FUZZY LOGIC CONTROLLER


Gurdal Arslan and Kwang Y. Lee†
Department of Electrical Engineering
The Pennsylvania State University
University Park, PA 16802
†kylece@engr.psu.edu

Abstract this paper for the fuzzy control of a nuclear rector which was
previously treated by Ramaswamy et al. [2] without robustness
The idea of studying the stability of fuzzy logic controllers analysis.
within the Lur’e problem framework is extended to the case where
the plant model contains parameter uncertainty. A fuzzy logic 2. Fuzzy Logic Control of a Nuclear Reactor
controller is designed to control a nuclear reactor. Then, the
stability and the robustness of the overall system is studied by using A fuzzy logic controller is designed to control the power level
the recent developments in the robust control theory. To support the of a nuclear reactor. The overall nonlinear mapping of the fuzzy
theoretical results, a series of simulations is performed for different logic controller is obtained to verify the stability and robustness
plant conditions. While fuzzy logic controller provided good properties of the system before implementation. Fig. 2.1 shows the
performance for the nominal plant, it also maintained the stability general structure of the closed-loop system, where nr is the reactor
under the off-nominal plant conditions. power level, nd is the power demand, zr is the rod speed, and e is
the error signal. The fuzzy logic controller shown has four major
parts, a fuzzifier, an inference mechanism, a rule base and a
1. Introduction defuzzifier.
Fuzzy set theory was introduced by Lotfi A. Zadeh as a method
for modeling human reasoning process [1]. Fuzzy control has been 2.1. Fuzzificaton and Membership Functions
a successful application of this theory and implemented in many The basic role of the membership functions is to convert a crisp
industrial systems. Simplicity and flexibility of fuzzy controllers value to a fuzzy set. Two sets of membership functions have to be
have played an important role in their acceptance. Moreover, it has chosen for both input and output of the inference mechanism.
been claimed that fuzzy controllers are robust with respect to
parameter variations in the plant model. This claim has been mainly fuzzy logic controller
supported by various applications and simulation results [2].
One of the most important issues that arises in the control
theory is the stability of the systems. Furthermore, mathematical
defuzzifier
fuzzifier

models of the systems are typically imprecise mostly because of the nd e inference zr reactor nr

parameter variations in the plant model and the unmodelled high mechanism model

frequency dynamics. Therefore, maintaining the system stability in
the face of uncertainties is a primary concern for the design rule
engineer. Stability and robustness of the conventional systems have base
been extensively studied in the past. But, little progress has been
made towards investigating stability and robustness of fuzzy
controllers [3-9]. Perhaps, the difficulty in studying stability and
robustness of the fuzzy controllers is due to the fact that fuzzy logic Fig. 2.1. General structure of the closed-loop system.
control design is based on heuristics and expert knowledge whereas
conventional control design is based on mathematical model of the
2.1.1. Input membership functions: Three bell-shaped
plant. Though these two approaches are fundamentally different
functions are used to convert the error signal into three linguistic
with respect to their bases, the resulting controllers are almost the
terms: N (Negative), Z (Zero), P (Positive), as shown in Fig. 2.2:
same with respect to their structures, which are nonlinear
controllers. 1 − tanh[500(e(t ) + 0.001)]
μ eNegative [e(t )] = , (2.1)
In this paper, the idea of studying the stability of fuzzy control 2
systems within the Lur’e problem framework [10,11] is extended to
μ eZero [e(t )] = exp[ −200000e2 (t )] , (2.2)
the case where the plant model contains parameter uncertainty. In
particular, the conditions of the circle criterion can be verified by 1 + tanh[500(e(t ) − 0.001)]
μ ePositive [e(t )] = , (2.3)
examining the Nyquist plot of the plant. In addition, recent 2
developments in the parametric robust control theory has made it
easy to determine the important frequency domain characteristics of where e(t ) denotes the error signal obtained by subtracting the
uncertain linear systems. The strict positive realness or the reactor output from the power demand.
boundary of the Nyquist plots of a transfer function family
containing parametric uncertainty can be determined by checking
some prominent members of the family. Such a study is given in
n
∑ ci μ i
1 r

u (t ) = i =n1
*
, (2.4)
∑ μi
r
i =1
0.5

where u* (t ) is the output of the defuzzification operation, ci is the


error
signal
0 e(t)
- 0.01 - 0.005 0 0.005 0.01 center of the fuzzy output set implied in the i-th rule, and μ ir is the
relevance of the i-th rule. By substituting the appropriate variables
Fig. 2.2. Input membership functions. into (2.4), the overall mapping of this fuzzy logic controller is
computed as follows:
2.2.2. Output membership functions: Three triangular-
shaped functions were used to express the rod speed in linguistic 0.05[ tanh[500(e + 0.001)] + tanh[500(e − 0.001)]]
zr (e) =Ψ(e) =
terms, N (Negative), Z (Zero), P (Positive), as shown in Fig. 2.3. 2 − tanh[500(e + 0.001)] + 2exp(−200000e2 ) + tanh[500(e − 0.001)]
(2.5)
1 where e denotes the error signal. The overall nonlinear mapping of
the fuzzy logic controller is shown in Fig. 2.4, which is a smooth
mapping and goes through the origin. The fuzzy controller
0.5
represented by such equation resembles a proportional controller
with a saturation nonlinearity.
rod rod
speed speed
zr ( t ) zr (e)
- 0.1 - 0.05 0 0.05 0.1
0.05

Fig. 2.3. Output membership functions.


error
2.2. Rule Base - 0.01 - 0.005 signal
The rule base is simply composed of a collection of the fuzzy 0.005 0.01
e(t )
if-then rules. In this work, three simple fuzzy if-then rules are used
to control the power level of the reactor:
if the error is Negative, then the rod speed is Negative,
if the error is Zero, then the rod speed is Zero,
- 0.05
if the error is Positive, then the rod speed is Positive.

2.3. Inference Mechanism Fig. 2.4. The overall mapping of the fuzzy logic controller.
The inference mechanism simply determines a mapping from
the input fuzzy sets to the output fuzzy sets based on the fuzzy if- 3. Stability and Robustness Analysis
then rules stored in the rule base. Typically, the decision making In this section, the use of nonlinear system and robust control
process involves several steps explained below. techniques for the robust stability analysis of the fuzzy control
In the first step, the fuzzy inputs to the inference mechanism system is demonstrated. As a result of several intermediate steps,
are compared with the premises of all the fuzzy if-then rules in the the closed loop system is represented in the Lur’e problem
rule base to determine the relevance of the each rule to the current configuration. Then, using the available analytical tools, the robust
fuzzy inputs. In the second step, using the product fuzzy absolute stability of the fuzzy control system with the linear reactor
implication, a fuzzy decision is made by modifying the fuzzy output model is shown.
sets in the consequent parts of the all currently relevant rules. In the
final step, sometimes called aggregation, the modified fuzzy output 3.1. The Linear Reactor Model
sets from all the rules are unified to have a single fuzzy conclusion To analyze the robust stability of a nonlinear system within the
using the sum method. Lur’e problem framework, the system has to be decomposed into a
linear subsystem and a nonlinear feedback term. An equivalent
2.4. Defuzzification and Overall Mapping representation of the reactor model in terms of the deviations from
A fuzzy decision represented by the aggregation of the an equilibrium is given by the following equations:
modified fuzzy output sets is made in the inference mechanism.
This fuzzy decision regarding the current control action has to be dδnr δρ − β n~ δρ 1 G
converted to a crisp value to control the plant. The process of = δnr + r + ∑ β iδcri , (3.1)
dt Λ Λ Λ i =1
finding the crisp value that best represents the fuzzy decision is
dδcri
called defuzzification. In this paper, the most popular = λ iδnr − λ iδcri , i = 1"6, (3.2)
defuzzification method “Center-Of-Area (COA)” is used. If the dt
sum method is used for the aggregation, and the output membership dδT f f f P0 Ω Ω
functions are triangular-shaped with the same bounded widths and = δnr − δT f + δTl , (3.3)
dt μf μf 2μ f
are symmetric around their centers, then
dδTl (1 − f f ) P0 Ω (2 M + Ω ) δzr = Ψ ( −δnr ), (3.12)
= δnr + δT f − δTl , (3.4)
dt μc μc 2μ c
where Ψ is defined in (2.5). One property of the function Ψ is
dδρ r that it represents an odd function. Secondly, Ψ is time-invariant.
= Grδzr , (3.5)
dt Moreover, Ψ is globally Lipschitz in δnr because ∂Ψ / ∂δnr is
α δT continuous and bounded. Lastly, Ψ is similar to a saturation
δρ = δρ r + α f δT f + c l . (3.6)
2 nonlinearity. It globally belongs to the sector [0,14], since it passes
through the origin and lies between the straight lines of slopes 0 and
Clearly, the origin is the equilibrium point of the system (3.1)-(3.6).
14 as can be seen in Fig. 2.4.
As far as the closed-loop system is concerned, the origin is also the
The closed-loop system with the linear reactor model,
unique equilibrium point because the fuzzy controller’s output, the
described by (3.7)-(3.11) and (3.12), has the fuzzy logic controller
rod speed deviation δzr , becomes zero only when the power level
represented by Ψ in the feedback path. Therefore, the fuzzzy logic
deviation δnr is zero. Let x denote the state vector control system exactly fits the Lur’e problem configuration [10].

[δ nr ]
T Since the linear reactor model has a zero at the origin, the linear
δcr1 δcr 2 δcr 3 δcr 4 δcr 5 δcr 6 δT f δTl δρ r , subsystem block is not Hurwitz. It is also known that the nonlinear
u denote the control input δzr , and y denote the output δnr . Then, subsystem, the fuzzy logic controller, belongs to the sector [0,14].
In this case, the circle criterion cannot be used to asses the stability
the first order approximation of this model is given by:
of this system. To make this system suitable to the circle criterion, a
dx loop transformation is performed. Namely, a constant output
= Ax + Bu, and (3.7) feedback −5δnr is applied around the linear subsystem to obtain a
dt
y = Cx, (3.8) new Hurwitz linear subsystem. The effect of this feedback is
removed by subtracting 5δnr from the output of the fuzzy logic
where
controller. It should be noted that these fictitious feedbacks do not
⎡ −β β1 β2 β3 β4 β5 β6 α f n~r α c n~r n~r ⎤ appear in the real implementation of the fuzzy controller and their
⎢ ⎥
⎢ Λ Λ Λ Λ Λ Λ Λ Λ 2Λ Λ⎥ mere purposes are to make the original problem suitable to the Lur’e
⎢ λ1 −λ 1 0 0 0 0 0 0 0 0 ⎥
⎢ λ2 0 −λ 2 0 0 0 0 0 0 0 ⎥
problem.
⎢ ⎥
⎢ λ3 0 0 −λ 3 0 0 0 0 0 0 ⎥ Nominal
⎢ ⎥ transformed
⎢ λ4 0 0 0 −λ 4 0 0 0 0 0 ⎥
A=⎢ λ5 0 0 0 0 −λ 5 0 0 0 0 ⎥, system
⎢ ⎥
⎢ λ6 0 0 0 0 0 −λ 6 0 0 0 ⎥
⎢ f f P0
⎢ 0 0 0 0 0 0
−Ω Ω
0 ⎥
⎥ 0.06 ω = −50
⎢ μf μf 2μ f ⎥
⎢ (1 − f f ) P0 Ω −2 M − Ω ⎥ ω = −100
⎢ 0 0 0 0 0 0 0 ⎥
⎢ μc μc 2μ c ⎥ 0.04
⎣⎢ 0 0 0 0 0 0 0 0 0 0 ⎥⎦ ω = −25
(3.9)
B = [0 0 0 0 0 0 0 0 0 Gr ] ,
T 0.02
(3.10)
C = [1 0 0 0 0 0 0 0 0 0]. (3.11)
0 Extremal
The matrices A and B are the functions of the reactor parameters Im{.} set
some of which are uncertain. These matrices also depend on the
equilibrium value of the reactor power level which changes -0.02
according to the power demand. It will be assumed that the power
demand may have values between the 10% power level and 120% ω = 25
power level. The uncertainty in the linear reactor model (3.7)-(3.11) -0.04
due to the power demand changes are precisely defined by ω = 100
1.2 ≥ n~r ≥ 0.1. The nominal linearized model has an eigenvalue at ω = 50
-0.06
the origin but is otherwise stable. The eigenvalue at the origin -0.04 -0.02 0 0.02 0.04 0.06
Re{.}
enters the model because the rod speed is used as the control input.
If the rod position rather than the rod speed is used as the control Fig. 3.1. Nyquist plots of the extremal and the nominal systems.
input, the real parts of all eigenvalues of the linear reactor model
becomes strictly negative. Fig. 3.1 shows the Nyquist plots of the extremal set and the
nominal transformed system as ω changes from -500 to 500. All
3.2. The Fuzzy Logic Controller the Nyquist plots evolve towards the origin as the magnitude of the
The error between the power demand and the reactor power frequency increases because all the transfer functions involved are
level is equal to the power level deviation with the sign inversion, strictly proper. The family of plots excluding the single plot marked
i.e., e = −δnr , since the equilibrium value of the reactor power level as “Nominal transformed system” belongs to the extremal set
is equal to the power demand. Thus, the overall mapping of the associated with the overbounding system. A circle whose center is
fuzzy logic controller (2.5) in terms of the deviations of the on the real axis and whose circumference passes through the points
variables takes the following form:
r1 = (-0.05+j.0) and r2 = (0.065+j.0) is also shown in Fig. 3.1. This and their minimum values simultaneously. Neither any faster nor
circle corresponds to a nonlinearity that belongs to the sector any slower response is detected in any case.
[ m1, m2 ] where m1 = −1 / r2 = −15.38 and m2 = −1 / r1 = 20 . The results of the test for the nominal case and the two extreme
cases are presented. The three cases are performed are labeled as
Because the Nyquist plots of the extremal set and the nominal follows:
transformed system remain inside this circle, the Nyquist plots of
the overbounding system must remain in this circle implying that nominal case: Gr = 0.01, and α c = 0.00001.
the Nyquist plots of the entire transformed system must also remain stressful case: Gr = 0.02, and α c = 0.00001.
inside this circle. Therefore, the robust absolute stability of the
transformed system for the feedback nonlinearities belonging to the relaxed case: Gr = 0.005, and α c = −0.0004.
sector [-15.38,20] is concluded from the circle criterion [10,11].
In all tests and cases, three system variables are monitored. These
Since the transformed nonlinearity belongs to the sector [-5,9]
three important variables are the reactor power level, the control rod
which is contained in the sector [-15.38,20], the robust absolute
speed, and the exit temperature. The plots of these variables along
stability of the fuzzy control system with the linear reactor model
with their analysis’ are given in the sequel.
described by (3.7)-(3.11) and (3.12) is concluded under all
The simulation results are shown in Fig. 4.1 through Fig. 4.3.
parameter variations. By robust absolute stability, it is meant that
In this test, a step decrease is applied to the power set point from
the fuzzy control system stabilizes all uncertain linear reactor
100% power level to 90% power level at t=5 sec. Fig. 4.1 shows the
models which are linearized around all power levels between the
power level responses for the three different cases. In the nominal
10% power level and the 120% power level. In addition, the sector
case, the power level response settles to the target value in
bound that can be tolerated for the nonlinear term without losing the
approximately 3 seconds. In the stressful case, the power level
stability is found to be much larger than the sector bound on the
response is faster than the one in the nominal case and it settles to
transformed nonlinearity. This indicates that the fuzzy logic
the target value in less than 2 seconds. The fastness of this response
controller actually can accommodate larger uncertainty variations.
is due to the drastic increase in the value of the total rod reactivity.
In the relaxed case, the power level response is rather slow because
4. Simulation Results of the big decreases in the values of the total rod reactivity and the
In this paper, a special attention is paid to the robust stability of coolant temperature reactivity coefficient. The settling time in this
the fuzzy control system. It is shown that the equilibrium point of case is approximately 8 seconds. In all cases, the steady state error
the closed loop system remains asymptotically stable under all is zero. And, no overshoot occurs in any case.
allowable parameter variations and power levels. However, this Fig. 4.2 shows the control rod speed responses for the three
result alone does not justify the practicality of the fuzzy controller cases. In all cases, an abrupt change occurs in the control rod speed
because a truly practical controller should also provide a high when the step decrease is applied to the power set point at t=5 sec.
performance. Providing a high performance under all parameter Then, the control rod speed remains constant in all cases until the
variations, the robust performance, is a difficult task for a fixed power level error in the corresponding cases becomes very small.
controller. When the power level first enters the vicinity of its target value, the
It is known that the power level changes have strong influence control rod speed starts increasing towards its steady state value in
on the nuclear rector. To test the performance of the fuzzy logic all cases. The effects of the parameter variations on the control rod
controller in the presence of power level changes, step change in the speed are consistant with the corresponding effects in the power
power level is applied from 100% to 90%. This is pertaining to the level. The control rod speed is fast in the stressful case and slow in
power level operations in the high or rated power region. First, the the relaxed case.
test is performed when the total rod reactivity Gr and the coolant Fig. 4.3 shows the exit temperature response for the three
cases. The exit temperature has very similar overdamped responses
temperature reactivity coefficient α c have their nominal values to
in all cases. The exit temperature responses settle in approximately
illustrate the nominal performance of the fuzzy logic controller. 25 seconds in all cases. The effects of the parameter variations on
Then, the same test is repeated for the cases where the parameters these responses are also consistent with the corresponding effects in
Gr and α c have variations in their values. It is known that the the previous responses.
worst value of the total rod reactivity Gr in terms of the system
stability is its maximum value. When the total rod reactivity Gr
has its minimum value the mildest system conditions in terms of the
stability are obtained. And, the variations in the coolant
temperature reactivity coefficient α c usually do not have a
significant effect on the nuclear reactor. In the light of this
information, extensive computer simulations are performed to find
the most extreme cases in terms of the time domain specifications.
More precisely, the settling time is used to determine the extreme
cases because virtually no overshoot and no steady state error is
observed in most cases. Based on the simulation results, it is found
that the most stressful and the most relaxed cases in terms of the
system stability correspond to the fastest and slowest system
responses, respectively. The most distinguished system responses in
terms of the settling time are obtained in two cases, where the two
parameters Gr and α c have their maximum values simultaneously
1.02 6. Conclusions
In this paper, a fuzzy controller is designed to control the
1
solid : nominal case power level of a nuclear reactor. The design of the fuzzy controller
reactor power level (nr)

dashed : stressful case is essentially based on heuristics and reactor operation. Through an
0.98 iterative process, the parameters of the fuzzy controller are tuned to
dotted : relaxed case
obtain a good performance for the nominal plant. Then, robust
0.96 stability of the fuzzy controller is shown by using the circle
criterion. Finally, a set of simulations is performed to test the
0.94 performance of the fuzzy controller under parameter variations. The
main conclusion of this work is that fuzzy control is not totally ad
0.92 hoc and that there exist formal techniques for the analysis of a
fuzzy controller. Fuzzy systems constitute a special subset of the
0.9 nonlinear systems. The overall mapping of a fuzzy controller can
always be obtained. Once this mapping is obtained, all existing
0.88
nonlinear techniques can be used to analyze this fuzzy controller.
0 5 10 15 20 25 30 35
time (sec.)

Fig. 4.1. Reactor power level responses. 5. References


[1] L. A. Zadeh, “Fuzzy sets”, Journal of Information and Control,
vol. 8, pp. 338-353, 1965.
0.05 [2] P. Ramaswamy, R. M. Edwards, and K. Y. Lee, “An automatic
tuning method of a fuzzy logic controller for nuclear reactors”,
0.04 solid : nominal case IEEE Transactions on Nuclear Science, vol. 40, no. 4, pp.
0.03 dashed : stressful case 1253-1262, 1993.
dotted : relaxed case [3] W. M. J. Kickert, and E. H. Mamdani, “Analysis of a fuzzy
0.02
rod speed zr

logic controller”, Fuzzy Sets and Systems, vol. 1, no. 1, pp. 29-
0.01
44, 1978.
0 [4] K. S. Ray, and D. D. Majumder, “Application of circle criteria
-0.01 for stability analysis of linear siso and mimo systems
associated with fuzzy logic controller”, IEEE Transactions on
-0.02
System, Man, and Cybernetics, vol. 14, no. 2, pp. 345-349,
-0.03 1984.
-0.04
[5] J. B. Kiszka, M. M. Gupta, and P. K. Nikiforuk, “Energetistic
stability of fuzzy dynamic systems”, IEEE Transactions on
-0.05
System, Man, and Cybernetics, vol. 15, no. 6, pp. 783-792,
0 5 10 15 20 25 30 35 1985.
time (sec.) [6] G. Langari, and M. Tomizuka, “Stability of fuzzy linguistic
Fig. 5.2. Control rod speed responses. control systems”, Proceedings of the 29th IEEE Conference on
Decision and Control, Honolulu, Hawaii, pp. 2185-2190, 1990.
[7] L. X. Wang, Adaptive fuzzy systems and control : design and
317.5 stability Analysis. Prentice-Hall, NJ, 1994.
[8] K. Tanaka, and M. Sugeno, “Stability analysis and design of
317 fuzzy control systems”, Fuzzy Sets and Systems, vol. 45, pp.
solid : nominal case 135-156, 1992.
exit temperature (Tl)

316.5
dashed : stressful case [9] P. Myszkorowski, and R. Longchamp, “On the stability of
dotted : relaxed case fuzzy control systems”, Proceedings of the 32nd IEEE
316 Conference on Decision and Control, San Antonio, Texas, pp.
1751-1752, 1993.
315.5 [10] H. K. Khalil, Nonlinear systems. Macmillan Publishing
Company, New York, 1992.
315 [11] M. Vidyasagar, Nonlinear systems analysis. Prentice-Hall
Publishing Co.,Englewood Cliffs, NJ, 1993.
314.5

314

0 5 10 15 20 25 30 35
time (sec.)

Fig. 4.3. Exit temperature responses.

Das könnte Ihnen auch gefallen