You are on page 1of 6

I .

A Comparison of Robustness:
Fuzzy Logic, PID, & Sliding Mode Control
Charles P. Coleman and Datta Godbole
Department of Electrical Engineering and Computer Sciences
University of California at Berkeley
Berkeley, CA 94720, USA

Abstract- This work is performed in order to D . A Call for Peace or at Least a Cease Fire! and
foster and promote unbiased and accurate compari- Negotiations
son of fuzzy logic control and classical control design
methodologies. We are motivated to execute this Fortunately, some have called for control engineers
study by the demonstrated robustness to plant per- to consider inclusion of fuzzy logic control into their
turbations of the fuzzy logic controller given in [l].
Robust fuzzy logic, PID, and sliding mode con- control enginering toolboxes, and for the utilization
trollers are designed to control the speed of a nominal of both fuzzy logic control and classical control when
third order linear timdnvariant model of a motor. and where appropriate ([12] [13]). This appears to
The step response performance of each controller, ap-
plied to the nominal and two perturbed motor plants, be a reasonable request. At the very least, the fuzzy
is presented. logic tool can be compared t o classical control tools.
We conclude fuzzy logic control can be a useful tool Should the fuzzy logic control tool prove useful and
for the control engineer. We encourage more bench-
mark comparisons of fuzzy logic control with classical acceptable to the control engineer, it should become
control techniques for the benefit of the practicing welcome addition to her control toolbox.
control engineer.
E. Robust Controllers
1. INTRODUCTION The control engineer is concerned about such prop-
A . New Control Tools, New Enemies? erties as the robustness of a controller to plant per-
turbations and uncertainty. We take the perspective
The successful use Of logic controllers has greatly of the control engineer who has several control tools
expanded in the last twenty years ([2]-[51)* This ex- available for the synthesis of robust controllers, and
Pansion has Prompted much comparison to classical who iS interested in determining the capabilities of
techniques* The ensuing diScussions have each control tool to synthesize a robust controller.
not always been amicable. Nor have they necessar- To assess the usefulness of each tool in designing
ily lead to manifest results and conclusions for the a robust controller, we chooeKa benchmark prob-
practicing control engineer. lem, and then engage in a comparison of the control
B. Debate and Dilemma designs which result from the application of each
control design method.
At times the debate over the use of fuzzy logic con-
trol techniques versus classical control techniques 11. CHOICE OF CONTROL TOOLS
h a become quite heated with ardent detractors on Fuzzy logic control, PID control, and sliding mode
both sides ([6]-[ll]). control tools are chosen for our investigation. Fuzzy
engineers have been placed in a logic control is chosen because of its empirically demon-
by this debate, and have been left with- strated robustness properties shown in [l]. PID con-
Out an Or comparison which trol iS chosen because it is one of the most corn-
to formulate a useful opinion. monly used controllers in industry, and because it
C. WelCBeing of the Control Engineer has plant perturbation robustness properties which
can be mathematically analyzed. We choose a slid-
In %he process of this raging and seemingly grow-
ing mode controller because it is a robust non-linear
ing battle, it appears that the best interest of the
control method whose robustness properties can also
control engineer has been occluded. The practicing
be mathematically analyzed.
control engineer usually has little interest in ideolog-
We do not claim to investigate and present every
ical debates. It is her interest to produce a working
robust control technique. Indeed, the popular and
controller, with efficacy, making use of any and all
successful H , and adaptive control techniques are
techniques a t her disposal.
noticeably absent from this study. We felt it was
necessary to restrict ourselves to fuzzy logic control
and two other control methods, in order to keep our

$4.0001994 IEEE
0-7803-1896-X/94 1654
study brief, intelligible, and meaningful. loop unit step responses of the nominal plant and
the two perturbed plants. To be acceptable, the
111. CONTROLPROBLEM closed loop step responses must simultaneously have
A . Motor Speed Control short rise time and no overshoot.
Motivated by the work in [l]and [14], and encour- F. Design Approach
aged by the robustness results presented for fuzzy
logic control, we consider the robust speed control We design the fuzzy, PID, and sliding mode con-
of a rotating motor, modeled by a third order linear trollers based on the nominal plant given in e q u a
time-invariant transfer function. The general con- tion (1). Each designed controller is then applied
figuration for the motor speed control problem is to all three transfer functions, and numerical sim-
given in Figure 1. ulations are performed to analyze the controller’s
robustness.
The design of the fuzzy logic controller is given
Controller in Section IV. The design of the PID controller is
given in Section V. The design of the sliding mode

I . I
controller is given in Section VI. Plots of each con-
troller’s closed loop step response and control effort
are shown in Appendix A.
Figure 1: Motor Speed Control Problem IV. ANALYSIS
A N D RECONSTRUCTION
OF A
FUZZY ROBUSTCONTROLLER
We use the nominal plant, perturbed motor plants, We implement the robust fuzzy controller given in
and the performance criteria given in [l] to design [l]and [14]. The inputs to the fuzzy controller are e
and compare robust fuzzy, PID, and sliding mode and b c . The output of the fuzzy controller is U. The
controllers. universes of discourse of e , bC,and U are partitioned
B. Restrictions on Controller Inputs and Outputs into seven fuzzy sets:
As shown in Figure 1, the reference step input speed 0 NB - negative big
w, and the output motor speed wc are available for 0 N M - negative medium
comparison. Each controller only has access to the 0 NS - negative small
motor speed error e = w, -w, and the motor angular 0 ZE - zero
acceleration wc as inputs. Each controller produces 0 PS - positive small
only one output U. 0 PM - positive medium
0 PB - positive big
C. Nominal Plant
Each fuzzy set is represented by a Gaussian mem-
The nominal motor plant is modeled by the follow-
bership function. The rule base of the fuzzy logic
ing transfer function:
controller contains forty-nine rules which are tabu-
lated in Figure 2. The output of each rule is de-
5 termined by min-inference. The crisp output U of
Gi(s) =
s(s + l)(s + 2) the fuzzy logic controller is generated by centroid
defuzzificat ion.
D. Perturbed Plants
The two perturbed motor plants are given by the
following transfer functions:

5
G ~ ( s )= -
sys + 2) (3)
Figure 2: Fuzzy Logic Controller Rule Base
E. Performance Criteria
The designed controllers must be robust to varia The step response of this fuzzy logic controller ap-
tions of system parameters. Specifically, for the plied to the nominal and perturbed plants is shown
fuzzy logic, PID, and sliding control designs, one in Appendix A. The step responses for all three
single controller must render acceptable the closed plants have short rise times and no overshoot. Thus,
they meet the specified robust performance criteria. We consider the following sliding surface for this
The control effort generated by the fuzzy logic con- design:'
troller is also shown in Appendix A.
s = (i - i d ) + A l ( j l - aid) +A2(y - Yd) (7)
v. ANALYSIS
AND SYNTHESIS OF AROBUST PID
CONTROLLER The choice of sliding surface is based on the follow-
The fuzzy logic controller designed in [l,141 is sim- ing considerations:
ilar to a nonlinear PD controller [15]. We design 0 The relative degree of the system (6) with S as

a linear PID controller t o obtain satisfactory step output should be 1. This ensures that the input
response for all three transfer functions. U appears explicitly on the right hand side of S
A linear PID controller can be characterized by equation.
the following transfer function: System dynamics on the sliding surface should
be stable. (This requires AI, A2 > 0)
(4) vd(t) specifies the reference trajectory to be tracked
by output. If one knows the entire trajectory of y d ,
Because, it is very difficult to realize a pure differ- then the derivative information i d ( t ) , &(t) can be
entiator, we use the following transfer function for extracted off line and used in the design t o improve
implementation. the performance. If the information is not available
(or as in this case, Y d ( t ) being unit step, is nondif-
C ( S )= K p +-
K d S
+-
T S + 1
K i
s (5) ferentiable) we can assume jld and lid to be zero and
let the robustness property of the sliding mode con-
To have a fair comparison, we use a small value for troller take care of the mismatch.
K i = 0.001, just enough to keep steady state error
zero. We use a small value for T = 0.01, so that its With the definition of sliding surface as above, we
effect on the dynamics is minimal, but the derivative have reduced the design requirement from tracking
controller is still realizable. y d ( t ) to being on the surface s = 0. Once on the
The selection of K p and K d is based on the root sliding surface, the dynamics (equation 7) is e x p e
locus of C(s)Gl(s). The root locus gives us the nentially stable and asymptotic trajectory tracking
locations of closed loop poles. The proportional is achieved. The control U is designed to make the
and derivative gain values are chosen such that the surface S = 0 attractive and to reach the surface in
closed loop pole locations are in the left half com- finite time.
plex plain for all three plants. Another criterion for
selection of these gains is the step response of the Consider the following lyapunov function:
closed loop system.
1
We select the following gains: v = - s 2
2
K p = 2, K d = 5
With these gains, the closed loop system is stable for Its time derivative is given by
the open loop gain up to 51. Thus in particular, the
above PID controller stabilizes the systems given by V=SS (9)
the transfer functions ('p2)* For the Of (3)7 As the system has relative degree 1 with S as output,
we have closed loop stability for the open loop gain
we can solve for from the equation
upto 30. The step responses have no overshoot and
are critically damped. S = -K sgn(S) (10)
VI. ANALYSIS
AND ROBUST
SYNTHESIS OF
This results in a negative definite V and also guar-
MODECONTROLLER
SLIDING entees finite time convergence to the sliding surface.
To design a sliding mode controller, we convert the But, the controller will result in high frequency chat-
transfer function model into state space format. The tering near sliding surface. To avoid this, we use the
controllable canonical form realization of the model following expression to solve for u . ~
of (1) is given by:

= [; 0 -"2
1 z+[i]u 'Refer to [16]for details of sliding mode controller design.
2With this controller, the trajectory is not guarenteedt o reach
y = [ l 0 012 (6) the sliding surface in finite time.

1656
Expanding both sides of the equation, we get results. Software tools were also available for the
design and simulation of the PID and sliding mode
s = -222 - 3x3 + 5u + A123 + A 2 2 2 controllers for this problems, and so these two con-
= -K(23 + AIXZ + &XI - X2yd) Desired trollers were easy to implement.
The U obtained from this equation will be used for B. Analytical Tools for the Analysis of Fuzzy Logic
the control. Control
A control engineer adding the fuzzy logic control
This controller needs access t o all three states, tool t o her toolbox already full of classical control
whereas the fuzzy logic controller of [l, 141 make tools, might ask “Why does this method demon-
use of only x1 and 2 2 . To have a fair comparison, strate robustness?”, and “When will it fail?” These
we will construct an observer to get an estimate of questions often have mathematically analytical an-
the state variable 23. swers for classical control tools. To satisfy this in-
quiry, we call for the continued development of math-
Standard Luenburger observer is designed t o get ematically analytical tools to answer such questions
an estimate of 2 3 . We use the model of (1) to design to the satisfaction of the control engineering com-
the observer as follows: munity.
f = AZ+bu+ L(y-y) (13) C. Fuzzy Logic Control: A Useful Tool
y = cx From our brief study, we conclude that fuzzy logic
control should have a place in the control engineer’s
where, the matrices A , b and c correspond to the toolbox. Only time, experience, and further anal-
model in equation (1). The matrix L is chosen such ysis will determine whether fuzzy logic control be-
that the eigenvalues of observer error dynamics (A- comes a prominent tool in the control toolbox used
LC) are in the left half plane and the error dynamics by devotees of classical control techniques.
is faster than the dynamics on the sliding surface.
The closed loop step responses with this controller VIII. FUTUREWORK
are shown in Appendix A. The controller is robust We encourage more extensive fair and unbiased bench-
to input gain changes as well as to the changes in mark comparisons of fuzzy logic control with classi-
system dynamics while keeping the input magnitude cal control techniques for the benefit of the practic-
small. ing control engineer, and we hope t o become a part
Note: With the use of equation ( l l ) , the con- of this effort.
troller is essentially a linear state feedback controller. In the future, we hope to contribute t o the cre-
It results in closed loop eigenvalues of -1.3392 & ation of new analytical tools for the analysis of fuzzy
-4.33f.j 1.91 for model (1). The same controller, logic controllers. We also hope t o be able t o con-
when applied to other plants will give the following tribute to the continuing effort t o identify classes of
closed loop eigenvalue locations: problems where rule based fuzzy logic control tech-
Model of (2) + -1.32, -3.55, -19.12 niques have advantages over and are more appropri-
Thus, by design, the increase in gain makes the ate than classical control techniques.
closed loop system more stable ...
Model of (3) 3 -1.4, -3.79 f j2.62 IX. ACKNOWLEDGMENTS
VII. CONCLUSIONS The authors would like to thank Professor S. Shankar
Sastry for his gracious support and encouragement.
A . Comments on Controller Designs The authors would also like to thank Dr. Shahram
The control engineer proficient in PID and sliding M. Shahruz for helpful discussions.
mode control techniques can readily synthesize ro-
bust controllers to perform the task dictated by this REFERENCES
example. As demonstrated in [l] and verified in this [l] H . T. Nguyen, C.-W. Tao, W. E. Thompson,
study, the control engineer utilizing fuzzy control “An Empirical Study of Robustness of Fuzzy
techniques can readily achieve the same goal. Thus, Systems”, Proc. of 2nd IEEE Intl. Conf. on
all three control tools are available to the control en- Fuzzy Systems, pp. 1340-1345.
gineer for use in synthesizing an acceptable robust [2] R. M Tong, “An Annotated Bibliography
controller for this control problem. of Fuzzy Control”, in Industrial Applications
With the appropriate software tools available, the of Fuzzy Control, M. Sugeno, Ed., North-
fuzzy logic controller for this problem was relatively Holland, Amsterdam, Holland, 1985, pp249-
simple to implement, and it provided satisfactory 269.

1657
[3] T. Terano, K. Asai, M. Sugeno. Fuzzy Systems A CONTROLLER PERFORMANCE RESULTS
Theory and Its Applications. Academic Press,
Baeton, MA, USA.
[4] W. Pedrycz. Fuzzy Control and Fuzzy Sys-
tems, 2nd extended ed. John Wiley & Sons,
New York, NY, USA.
[5] H. Berenji, “FUZZYLogic Controllers”, in
A n Introduction to Fuzzy Logic Applications
in Intelligent Systems, R. R. Yager and
L. A. Zadeh, Eds., Kluwer Academic Publish-
ers, Boston, MA, USA, 1992, pp 69-96.
[6] S. Chiu, S. Chand, D. Moore, A. Chaudhary,
“FUZZY Logic for Control of Roll and Moment
for a Flexible Wing Aircraft”, IEEE Control
Systems Magazine, pp 42-48, Vol 11, No 4,
June 1991.
[7] A. L. Schwartz, “Comments on Fuzzy Logic
Control of Roll and Moment for a Flexible
Wing Aircraft”, IEEE Control Systems Mag-
azine, pp 61-62, Vol 12, No 1, February 1992.
[8] S. Chiu, “Author’s Reply”, IEEE ControlSys-
tems Magazine, pp 62-63, Vol12, No 1, Febru-
ary 1992. T i

[9] E. Cox, “Adaptive Fuzzy Systems”, IEEE


Spectrum, pp 27-31, Vol 30, No 2, February Figure 3: Closed Loop Step Response of nominal
1993. plant
[lo] C. J. Herget, Ed., “Reader’s Forum”, IEEE
Control Systems Magazine, pp 5-7, Vol13, No
3, June 1993.
[ll] M. Athans, “Control- The Adventure Contin-
ues”, Bode Lecture, 92nd IEEE Conference on
Decision and Control, San Antonio, TX, USA,
December 15-17, 1993.
[12] E. H. Mamdani, “Twenty Years of Fuzzy
Control: Experiences Gained and Lessons
Learnt”? Proc. of 2nd IEEE Intl. Conf on
Fuzzy Systems, pp. 339-344.
[13] M. Tomizuka, “FUZZYControl in Control
Engineer’s Tool Box”, Lecture, A R U / N A S A
Workshop on Formal Models f o r Intelligent
Control, MIT, Cambridge, MA 02139, USA,
September 30 - October 2, 1993.
[14] C.-W. Tao, R. Mamlook, W. E. Thompson,
“Reduction of Complexity for a Robust Fuzzy
Controller”, Proc. of 2nd IEEE Intl. Conf on
Fuzzy Systems, pp. 1346-1349.
[15] H. Ying, W. Siler, J. J. Buckley, “Fuzzy Con-
trol Theory: A Nonlinear Case”, Aulomatica,
Vol. 26, No. 3, pp 513-420, 1990.
[16] J.-J. E. Slotine and W. Li, Applied Nonlinear
Control. Prentice Hall, 2nd ed., 1991. Figure 4: Closed Loop Step Response of the per-
turbed model Gz(s)
11

i Od

iK 0.6

I 0.4

02

0 5 10 15 25

Tim

Figure 5: Closed Loop Step Response of the per- Figure 7: PID Controller Control Effort
turbed model G ~ ( s )

Figure 6 : Fuzzy Logic Controller Control Effort Figure 8: Sliding Mode Controller Control Effort

1659

You might also like