Sunteți pe pagina 1din 6

Numerical Stability Analysis of a Fuzzy Controller

Neil Petroff* Paul A. C. Mason** Kevin Walchko* Kim D. Reisinger **

Department of Mechanical Engineering University of Florida

Abstract This work describes a numerical stability technique that can be utilized for fuzzy and other intelligent control structures. It allows for the examination of local and global stability by approximating a time- and statevarying form of the fuzzy control effort as a series of linear control efforts. This form can then be equated to linear control theory for which the stability range is defined. This technique, which is an extension of state-space and classical stability theory, has been developed for single-input-single-output (SISO) and multi-inputmulti-output (MIMO) systems. Two simulation examples are used to illustrate the merits of this technique. For both examples, the method is implemented with a stable and an unstable controller. While this work presents the application to linear, timeinvariant (LTIV) plants, it is directly extended to complex, time-varying systems. 1. Introduction Advancements in technology and system integration have led to the development of complex, multifunctional, dynamic systems. These include smart appliances, medical devices, robots, and autonomous vehicles. Such systems are typically modeled using a linear plant to which nonlinearities and constraints are added. Furthermore, additional plant assumptions must be made to provide an analytical description of the complex system. The ability to control such complicated systems with traditional methodologies is difficult. Advanced control systems are required to achieve the desired performance of these systems. 1.1 Benefits of fuzzy logic for complex systems In engineering and mathematics, it has been common practice to represent information in a continuous rather than a discrete manner (Kosko, 1997). Despite this, fuzzy logic, which utilizes linguistic representations for information, is just beginning to make inroads in the United States, mainly in government applications to control large processes. In *Research Assistant **Assistant Professor

addition, there has been a recent push to examine fuzzy logic as an alternative to traditional controllers in industrial and scientific applications. Fuzzy logic is one of the most promising intelligent control schemes, making it ideal for application to complex, nonlinear systems. Fuzzy logic allows operator experience and system knowledge to be encoded without resorting to approximate analytical descriptions. It describes system dynamics through membership functions and rules rather than mathematical models. For this reason, it lends itself to the control of complex, time-varying, systems with random disturbances or noisy inputs. Fuzzy logic provides a robust technique to achieve desired performance in complex systems. However, issues associated with stability must still be addressed. 1.2 Stability When designing a control system, the most important criterion is stability. Several methods exist for analyzing the stability of dynamic systems. For LTIV systems, the stability theory can be reduced to algebraic analysis for location of the poles or eigenvalues of a model (Palm, 1983). Since only one equilibrium point exists in a LTIV system, local stability ensures global stability. Popular techniques for defining the stability of a LTIV system include Routh-Hurwitz, root locus, Bode, and Nyquist. For nonlinear and time-varying systems, the stability of all the equilibrium points or cycles must be classified to define global stability (Slotine and Li, 1991; Friedland, 1996). In some cases, nonlinear, timevarying systems can be viewed as a set of local, linear, stability and control problems. Often, the stability analysis of a nonlinear system is based on a linearization about an operating point. Another commonly used stability criterion for nonlinear systems is the Liapunov stability criterion (Slotine and Li, 1991; Ogata, 1997). This states that if the system energy is s decreasing, where the energy is a positive, scalar function, it is guaranteed that the trajectories will remain bounded in a given domain.

Intelligent control techniques such as fuzzy logic, neural networks, and genetic algorithms, do not require a mathematical model of the system to define a stable controller. Hessburg and Tomizuka (1995) make various assumptions about the system inputs and plant behavior such that the closed-loop system is boundedinput-bounded-output (BIBO) stable. In other methods, the stability of the system is defined based on a timeinvariant approximation of the functional form of the controller and plant. If the functional form is too complex, a linearization may be required to simplify the analysis. In Garcia-Cerezo et. al (1993), the fuzzy logic controller is approximated by a linear combination of weights and linear feedback controllers. Next, they perform a Liapunov stability analysis using the approximated controller output. 2. Fuzzy Logic Theory The power of fuzzy logic is its tolerance for ambiguity. This is accomplished by describing sets with linguistic rather than quantitative variables. An element in the set is said to have membership in the set. The value of this membership is normalized from {0, 1}. An element with a membership value of one is typically called a prototype. For example, a curve with a circumference-to-diameter ratio = would have a membership value of one in the class of circle. As the ratio deviates from , the curve can still be considered a circle, only to a lesser extent, yielding a membership value less than one. The practical implementation of fuzzy logic involves fuzzification of crisp inputs through linguistic membership functions. The rule base is evaluated in parallel to determine each rule effect on the system. s The rule outputs are aggregated to create a fuzzy output set. Finally, a defuzzification process generates a crisp output. Inputs and outputs are fuzzified through membership functions that linguistically describe the range of the value. Several methods can be used to develop these membership functions. They are intuition, neural networks, genetic algorithms, inductive reasoning, inference, rank ordering, angular fuzzy sets, soft partitioning, meta rules, and fuzzy statistics (Ross, 1995). A generic set of membership functions for an input variable is shown in Figure 1. The input variable is described by three linguistic variables which are pos, neg and zero. The membership function for neg and pos are trapezoidal in shape, while it is triangular for zero. A given input value can belong to one or several membership functions, and the degree to which it

belongs to each membership function is given by the fuzzy value between zero and one.
1 0.9 0.8 0.7 Fuzzy Value 0.6 0.5 0.4 0.3 0.2 0.1 0 -1 -0.5 0 Input 0.5 1 neg zero zero pos

Figure 1. Membership Functions to Fuzzify Input by Linguistic Variables neg, zero, and pos. Next, if-then rules are evaluated in parallel for the fuzzified inputs. Each rule evaluation forms a fuzzy, output set. The individual output sets are aggregated to form a final, fuzzy output set. This set is defuzzified through one of several methods such as center-of-area, or gravity (centroid), bisector, mean-of-maximum (mom), largest-of-maximum (lom), and smallest-ofmaximum (som) (Gulley and Roger Jange, 1995). Center-of-gravity (COG) is the most common defuzzification method (Kosko, 1997). It is given by
COG =

j=m

w ja w
j

j=1 j=m

j=1

b x gV c a b x gV
j j j j

(1)
,

where m is the number of rules, wj is the weight given to rule j, aj is the membership value of the variable in the fuzzy set for rule j, Vj is the volume of the output fuzzy set for rule j, and cj is centroid of that volume. Often, all the rules are weighted equally, allowing wj to be removed from (1). Defuzzified locations of the discrete output for a generic, fuzzy output set are shown in Figure 2. This discrete value is the control effort.
1 0.9 centroid Fuzzy Membership 0.8 0.7 0.6 0.5 mom 0.4 0.3 lom 0.2 0.1 0 -10 -5 0 Deffuzified Output 5 10 som bisector

Figure 2. Output Locations for Various Defuzzification

Methods.

Plot was developed with the use of Matlab

3. Stability Formulation for Fuzzy Controller This paper presents a numerical method for analyzing the stability of fuzzy controlled systems by assuming a control structure which is a function of time and of the states. By assuming this form, the fuzzy controller is analogous to linear control structures for which the stability range is defined based on full-state feedback and pole placement methods. Next, the control effort is analyzed to verify that it is within the stable range based on a linear approximation of the system and of the linear control structure. We emphasize that the methodologies that follow are also applicable to MIMO, nonlinear, and time-varying systems. For simplicity and illustration, however, only single and multi-variable examples are presented. Utilizing full state-feedback control, the state-space representation of the closed-loop system matrix is (2) where A is the system matrix, B is the input matrix, and K is the feedback matrix. For asymptotic stability, all of ~ the eigenvalues of A must be in the left-half plane. With some manipulation, (2) can be expressed as ~ ~* ~ ~ * Adiag + Adiag / 2 = P T A P + P T A P i / 2 < 0 , (3)

1 0 0 ~ A= 0 0 1 1 + k1 2 + k 2 3 + k 3

Applying the Routh-Hurwitz stability analysis, the stability bounds on the elements of k are ( 1 k1 ) k1 > 1 k3 > 3 k2 > 2 ( 3 k3 ) For a linear, SISO system, the stability is a function ~ of the elements in k, such that A is positive definite. From linear systems theory, a matrix P is positive ~ definite if xT Px > 0 (Chen, 1984). Now if P = A and P is positive definite then

( )

x T Px = x T ( A + bk ) x = x T Ax x T bkx > 0 (4)


where b is a nx1 vector and k is a 1xn vector. Consequently, a linear, time-invariant system is stable when

~ A = A + BK

x T bkx < x T Ax

(5)

)
i

([ ] [ ])

where P is a similarity transformation, A is the complex conjugate, the subscript diag denotes the diagonal of the matrix, and the subscript i denotes the ith eigenvalue. Equation (3) defines the stability of the system and the bounds for the elements in the K matrix. This relationship is valid for any system.
*

The stability analysis of most controllers requires that the structure of the controller be known or assumed. This is also true for intelligent controllers such as fuzzy logic or neural networks. In this work, the structure of the fuzzy controller, for a state-space representation, is assumed to be in the form (6) u fuzzy = K fuzzy (membership functions) x This is used to determine mathematically equivalent models of fuzzy control efforts

u fuzzy K x , t x

b g

(7)

3.1 SISO Systems For single-input systems, identification of the eigenvalues is not necessary to determine the stability range. Routh-Hurwitz stability criterion can be utilized to define the stability range for the elements in K without solving for the eigenvalues. To illustrate this, the following example is given. Assume the following controllable canonical statespace representation 0 0 0 1 0 0 1 x + 0u & x= c 1 2 3 1 with feedback gain

After equating the input/output (I/O) mapping of the fuzzy controller with a full-state feedback controller, feedback linearization theory can be applied to the fuzzy controller to investigate its stability. In other words, we require

x T BK fuzzy x < x T Ax ,

(8)

for the closed-loop system to be stable. This is only true if the approximation in (7) is valid. Consider the fuzzy controller for a single input variable depicted in Figure 3. This nonlinear fuzzy logic controller can be approximated by a series of straight lines. Therefore, the I/O mapping can be described by m linearizations

k = [ k1 k2 k3 ]
The closed-loop system model with positive feedback is

yi + 1 = f x1 f xi + Ki = f x

b g b g Ffx I b x G J H K
x = xi x = xi

i+ 1

xi

g
(9)

Similar to the SISO formulation, the MIMO system stability is a function of the elements in the K matrix, such that

xT Px = xT ( A + BK) x = xT Ax xT BKx > 0 (11)


Therefore, a linear, time-invariant system is stable when

u = yi + 1 yi ~=x x x i+ 1 i ~ u= Kx
i

x T BKx < x T Ax

(12)

where m indicates the number of line segments necessary to describe the nonlinear curve.
0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 u

In the case of a single-input, multi-variable feedback system, perturbation theory can be used to show that the surface over a small interval is described/approximated by its gradient. Consequently, the gradient is an approximation of the feedback gain vector

Ki

(13)

Consider the output surface for the multivariable case shown in Figure 4. For this particular case, with a single input and two feedback variables, x and w, the control effort can be described by the following

yi + 1 = f ( w , x) f ( w , x) xo , wo
0 0.2 0.4 x 0.6 0.8 1

Figure 3. Single-Variable Feedback Fuzzy I/O Map.

Ff I G x J H K
f x

( xi + 1 xi ) + k2 =
x = xo

x = xo

F f I Gw J H K
w = wo

( wi + 1 wi )
(14)

w = wo

Therefore, the fuzzy controller is approximated as a time- and state-varying full-state feedback controller (10) K fuzzy ( membership functions ) x = Ki x For this control system to be stable, Ki must satisfy the stability criteria for all x. Based on the assumption of the structure of the fuzzy controller, the stability bounds of the linear system can be applied to the fuzzy controller. The only difference being that K varies with time and the current state of the system. 3.2 MIMO Systems In a linear, MIMO system K is a matrix rather than a scalar or a vector. This K matrix defines the stability of the compensated system (Kailath, 1980). The control effort associated with full-state feedback is a direct function of the states and, therefore, multiple inputs can be examined independently. Furthermore, the stability analysis of a MIMO fuzzy controller, which is approximated by a linear, time-varying, full-state feedback controller, can be simplified to the analysis of several single-input control systems. In the case of MIMO systems, the I/O mapping for each input is described by a n-dimensional hypercube.

k1 =

~= x x w

L OP MN Q

f w

x x u1 = k11 k12 ~ = ~
This is easily extended to the MIMO case with n inputs and m feedback variables

k Lu OP Lk MM MM MP = MN u PQ MM kM N
11 1 21 n n1

k12 L k22 M kn2

OP L k ~ P x = Kx = x M MP P L k Q
k1m
2m nm

(15)

where K is a nxm matrix. If a state is not used in the control, then the element in K associated with that state is set to zero. This constrained feedback gain must still satisfy the stability criterion. If all equilibrium points are assumed stable, the system can be considered globally stable.

on the same plant. The I/O map is shown in Figure 6.


4
1

3
0.8

2 Control Effort
1 0 .5 w 0 0 x 0 .5 0.6 u 0.4 0.2 0 1

1 0 -1 -2 -3 -4 -5

Figure 4. Multi-Variable Feedback Fuzzy I/O Map. 4. Results 4.1 Single-Variable Feedback (SISO System) To verify the stability analysis, a fuzzy logic controller was designed and the output surface plot analyzed to determine the stability of the system. For the single-variable feedback case, a first-order transfer function, G ( s ) = 5 , was used. Routh-Hurwitz
s + 5

0 Error

Figure 6. I/O Map for Unstable, Fuzzy Controller. It is evident from Figure 6 that only small portions of the output space had slopes < -1. This highlights another benefit of this method by indicating local instabilities in the system operating range. Figure 7 s depicts the unstable fuzzy controller output and an unstable proportional controller (Kp=-1).
1 0.5 0 -0.5 Output -1 -1.5 -2 -2.5 -3 -3.5 0 Proportional Fuzzy Desired

analysis shows a single gain must be > -1 to ensure stability. The controller was used in the feedforward portion of the closed-loop system. The slope of every approximate line segment for the controller I/O map was positive, similar to Figure 3. Therefore, the system should be stable. Figure 5 shows the system is, indeed, stable for a unit-step input. The transient performance characteristics for the fuzzycontrolled system and for a proportionally-controlled system, with KP=1, were similar. The latter, however, had a smaller steady-state error.
1

2 Time (sec)

Figure 7. Unstable System Response to a Unit-Step Input Using Fuzzy and Proportional (KP = -1) Controllers. The proportionally-controlled simulation included a saturation block after the controller to better compare with the fuzzy-controlled system. Without saturation, the output of the proportionally-controlled system will continue unbounded. The fuzzy-controlled system, on the other hand, is bounded by the range of the controller output variable. s 4.2 Two-Variable Feedback (SISO System)

0.8

Proportional Fuzzy Desired

Output

0.6

0.4

0.2

2 Time (sec)

Figure 5. System Response to a Unit-Step Input Using a Fuzzy Controller and a Proportional Controller (KP = 1). Next, a known unstable controller was implemented

For the two-variable feedback system, a velocity 5 transfer function, G ( s ) = , was used.
s
2

s + 5

Position and velocity were the feedback variables. A full state feedback stability analysis shows the following requirements for K1, relating to position, and K2,

relating to velocity, to maintain system stability K1 < 1 and K 2 < 0 .2

(16)

A fuzzy controller was designed to control this system. The output is shown in Figure 8.
1.5

to be validated for stability since the approximated gains must adhere to the same stability criteria determined from full-state feedback analysis. The legitimacy of this approach was shown by application to a single-variable and to a two-variable feedback, SISO system. However, the method can easily be extended to MIMO systems. In addition, local stability can be examined with this technique. 6. References Chen, C. Linear System Theory and Design. Holt, Rinehart and Winston, 1984.

Output

0.5

-0.5

Position V e locity Desired 0 2 4 Time (sec) 6 8 10

-1

Friedland, B., Advanced Control System Design, Prentice-Hall 1996. Garcia-Cerezo, A., Ollero, A., Aracil, J. Dynamic Analysis of Fuzzy Logic Control Structures, Fuzzy Control Systems, Edited by Kandel & Langholz, CRC Press, 1994, pp. 142-159. Gulley, Ned, Roger Jang, J. S. Fuzzy Logic Toolbox for use with Matlab . The MathWorks, Inc., 1995. Hessburg, T., Tomizuka, M. Model Reference Adaptive Fuzzy Logic Control for Vehicle Guidance, Proc. of the American Control Conference, Vol. 3, June 1995, pp. 2287-2291. Kailath, T., Linear Systems. Prentice Hall, 1980. Kosko, Bart. Fuzzy Engineering. Prentice Hall, 1997. Maciejowski, J.M. Multivariable Feedback Design. Addison-Wesley Publishing Company, Inc., 1991. Ogata, Katsohiko, Modern Control Engineering, 3rd Ed. Prentice-Hall, 1997.

Figure 8. Second-Order System Response to a Unit Input Using a Fuzzy Controller. The gradient of the output surface over each variable satisfied (16), and the system is stable. Next, an unstable system was developed by trial and error. In this case, only gain criteria for the position feedback was violated. The resulting output is shown in Figure 9.
3

1 Output

-1

-2

Position Velocity Desired 5 Time (sec) 10 15

Palm III, William J., Modeling, Analysis and Control of Dynamic Systems. John Wiley & Sons, 1983. Ross, T. J. Fuzzy Logic with Engineering Applications. McGraw-Hill, 1995. Slotine, J.-J., and Li, W., Applied Nonlinear Control, Prentice-Hall, Inc 1991.

-3 0

Figure 9. Unstable, Second-Order System Response to a Unit Input Using a Fuzzy Controller. 5. Conclusion This paper presents a fuzzy logic stability analysis based on an extension of state-space linearization techniques and classical control concepts. The stability formalization approximates a fuzzy gain based on infinitesimally small portions of the input/output map. This allows various control schemes

S-ar putea să vă placă și