Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

1662 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 38, NO.

11, NOVEMBER 1993

Technical Notes and Correspondence


Nonlinear Control of a Shunt DC Motor trated in Fig. 1 below, showing a D C motor and its associated
schematic. The TI terminal of the armature circuit is electrically
John Chiasson and Marc Bodson connected to the Ti terminal of the field circuit and, corre-
Abstract-The problem of controlling a shunt DC motor is considered.
spondingly, for the T, and Ti terminals (i.e., shunt connected).
Recent methods for nonlinear control are compared, including feedback
linearization, generalized controller canonical forms, and input-output Mathematical Model
linearization. Connections between the approaches, as well as their The equations describing this electromechanical system are [ 11
respective advantages are discussed in detail.
[51 [IO]:
I. INTRODUCTION
0 = w
In recent years, there have been significant advances in the
Jh = 7
, - Bw - rL (1)
theory of nonlinear control. Several approaches have been
reported in the literature, but the connections between them L,di,/dt + ( R a d , + R F ) i F= V,
have not been fully explored. In this note, we consider the
example of a shunt D C motor to evaluate the strengths and where T, = K,$i, is the electromagnetic torque, K , = torque
limitations of three different methods for nonlinear control. constant, I$ = K,i, is the flux in armature due to the field
The mathematical model of a shunt DC motor is nonlinear, current, KF = mutual inductance, i, = armature current, i, =
with polynomial type nonlinearities. The system consists of three field current, R a d j+ R , = field resistance ( R a d , an adjustable
state variables and one input. A traditional way of control- resistance), T~ = load torque, L a = armature inductance, L , =
ling such a system is to linearize the dynamics about a desired field inductance, J = rotor moment of inertia, B = viscous fric-
operating point [S, p. 1121. Then, using well known techniques, a tion coefficient, 0 = angular position, w = angular speed, V , =
linear controller can be designed that keeps the system in the terminal voltage.
neighborhood of its operating point. In contrast, recent methods We make the usual assumption that the armature inductance
for nonlinear control take the nonlinearities fully into account. is negligible (i.e., set L , = O), so that the armature current is
The main advantage of such approaches is that for large regions given by [see Fig. l(b)]
of the state-space, the controller does not need to be gain-
i, = (V, - V,)/R,
scheduled according to the operating point. Furthermore, this
control is achieved even for large state deviations, for example, where V, = K,$w is the back-emf, R , is the armature resis-
during the transition between operating points. tance and K , is the back-emf constant. A simple conservation of
Three methods are considered in this note: feedback linear- energy argument shows that K , = K , in MKSA units. Letting
ization (e.g., [2], [3]),generalized controller canonical forms [4], x 1 = 0, x 2 = W , x , = i,, and U = V,, system (1) becomes
and input-output linearization (e.g., [7]). Both position and
speed control are considered for comparison of the nonlinear
xi = x ,
control methods. Although the model of the shunt DC motor is x, = -cIx2 - c 2 x 2 x : + c,x,u - c6rL
relatively simple, it allows one to investigate some significant
differences in the properties of the control schemes. x, = -cqx, + cgu (2)
The note is organized as follows. In Section 11, we review the where c1 = B / J , c2 = K,K,Kj/(JR,), c3 = K,K,/(JR,),
mathematical model of the shunt DC motor. In Section 111, c4 = (R,,, + R F ) / L F c5
, = l / L F , and c6 = 1/J. We assume
we discuss the application of feedback linearization to the full that the position, speed and field current are available for
state-space model and to the restricted model used for velocity feedback. We will also assume that the load torque rL is known.
control only. In Section IV, we obtain the controller based on In practice, one can either let rL = 0 for the control algorithm
generalized controller canonical forms and, in Section 5 , based and treat the load torque as a disturbance, or assuming constant
on input-output linearization. Finally, in Section VI, we con- load torques, it is easy to see that a nonlinear observer with
clude with general remarks on the respective advantages of linear error dynamics can be constructed to estimate ?L as
the three approaches. A preliminary version of this work has shown in [9].
appeared in [12]. More compactly, we may rewrite (2) as
ILTHE SHUNTDC MOTOR
x = f ( x ) + g(x)u +d
A shunt D C motor is a D C motor in which the field circuit is
connected in parallel with the armature circuit. This is illus-

Manuscript received May 3, 1991; revised January 24, 1992 and


August 20, 1993.
J. N. Chiasson is with the Department of Electrical Engineering, Remark: There is a bound on the maximum achievable
University of Pittsburgh, Pittsburgh, PA 15261.
M. Bodson is with the Department of Electrical and Computer Engi-
steady-state speed. Consider the constant speed operating points
neering, Carnegie Mellon University, Pittsburgh, PA 15213. of the original system (2) by letting X2 = f, = 0. Eliminating the
IEEE Log Number 9212809. input U , the steady-state values (x,,, x3,,) = (w,,, i F o ) are con-

0018-9286/93$03.00 0 1993 IEEE


IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 38, NO. 11, NOVEMBER 1993 1663

n*T: which is identically zero only if x g = 0, a noninteresting trivial


situation since it results in zero torque. In other words, feedback
linearization is not applicable for position control.
If one is concerned with speed control, the dimension of the
Vb = Kb@
state-space is reduced from 3 to 2 and an interesting lineariz-
ing transformation exists. Specifically, with z , g x,, z 2 x3, the
Lr
system ( 2 ) becomes
T
2
= T'
2
i, = -C1Z1 - c2z,z; + c3zzu - C67L (4)
i, = -cqz2 + cgu
Fig. 1. (a) Direct current motor. (b) Schematic diagram of a DC shunt
motor. or, with some abuse of notation,
i =f(z) +g(z)u + d
i2
FO

I w omax = c 4 c 3/ c 5 c 2 = (Rad, + RFI/KbKF

A feedback linearizing transformation exists in an open set FY of


the state-space ( z l ,z,) if rank[g a d f g ] = 2 and the set {g}
is involutive. The involutiveness of (g} is trivially satisfied and
3 2c3c4 + 2c2c,z, + c 2 c 3 z i ) .
P ( z ) 5 det [g, a d f g ] = c g z 2 ( c 1 c -
The (load dependent) nonlinear transformation given by
Fig. 2. xio = i:o versus xz(I= w,)
ZT = T,(z) A --C,Z~ + c32f/2
strained to the curve 2; = T,(z) %$(T,(z))

(cgcq - cgc~x20)x.$l= cIcgx2()+ c5cbr1 (3) = CIC5Z1 + c2c5z1z; - c3c4z; + CgC67,- (5)
with corresponding reference input u , , ~ c4i,,/c,. For a given results in
load torque r L , this curve is plotted in Fig. 2 showing the if = 2 ;
possible steady-state operating points. For a solution to exist, we
must have i; = a(z)+ P(Z)U (6)
where a ( z ) e -(c,c, + c 2 c g z ~ ) ( c I +
z 1c 2 z , z i + c 6 r L ) -
c,z,(2c,cSz,z, - 2c3c,z2) and p ( z ) as given previously. By
which is a bound on the maximum achievable steady-state speed.
Thus, no matter how large the input voltage is, there is an
+
simply choosing the feedback U = ( - d z ) L ' ) / ~ ( Z ) ,which is
valid for all z such that P ( z ) # 0, we achieve a linear system
upper limit to the steady-state speed. This limit can be increased with new input 1 ' .
by increasing the adjustable resistance. This is just the classical This control scheme suffers from the practical issue of main-
field weakening approach to speed control [lo]. We mention this taining P ( z ) # 0 or, more specifically, P ( z ) must be bounded
bound because it will also arise in the next sections as con- away from zero, since otherwise the control command will be
straints for certain nonlinear controllers to be valid. unbounded. A straightforward calculation shows that p ( z > f 0
111. FEEDBACK
LINEARIZATION is also equivalent to the condition for the invertibility of the
transformation, i.e., d T / d z nonsingular. The set of points z
As a first approach to finding a feedback controller for this such that P ( z ) = 0 consists of the two curves:
nonlinear system, we consider the (exact) feedback linearization
approach (see, for example, [2],[3]and see [8]for an application z 2 = 0 and z, = (2c4c3- c I c 3- c , c , z ~ ) / ( 2 c , c 2 )
of the authors to stepper motors). This approach entails finding or, in terms of the original state variables, we have
a nonlinear state-space transformation such that, in the new
coordinates, the nonlinearities may be canceled out by state i, = 0 and o = (2c,c, - c l c j - c , c , i ~ ) / ( 2 c g c , ~ .
feedback. The necessary and sufficient conditions for such a Assuming that w = 0 is to be in the operating region, the
transformation to exist in some neighborhood ?Y of a point xo shaded region in Fig. 3 gives the operating region to which
are: the feedback linearization controller is limited. In this region,
rank [ g a d f g udr'g] = 3 in "I/ the transformation T ( z ) is a diffeomorphism and p ( z ) # 0. The
feedback linearization transformation is also valid above P ( z )
The set (g, ad,g} is involutive in ?Y
0, but the scheme would have to be modified to allow "jumping"
where adfg [ f ,g ] A ( d f / d x ) g - ( d g / d x ) f and adjg A over the singularity curve.
adf(adfk-'g). For the case of the DC shunt motor, these two Remark 1: Note that the requirement p ( z ) # 0 gives an upper
conditions are equivalent to limit to the speed given by
w < wmax c4c3/cgc, - c,c,/(2c,c,)
= (R,,, + R,)/K,K, - BLF/(~JK,KF)

A straightforward, but tedious computation shows that which is less than the steady-state bound wOmax calculated in
Section I1 (they are equal if the viscous-friction coefficient
B = 0).
1664 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 38, NO. 11, NOVEMBER 1993

w =
Omax c 4 c 3 I C 5 c 2 = (Rad, + R F ) / K b K F

i2 - - (c c t )/c c
FO 5 6 L 3 4
mln

w - ( R a d j + RF )/K,KF - BLF/2JKbKF
max

Fig. 3. 25 = :i versus z1 = w

Remark 2: It is interesting to note that the two dimensional in real-time. This scheme depends on x g being bounded away
system consisting of current and speed is (locally) feedback from zero but, as x 3 is the field current, this is not an unreason-
linearizable, while the addition of the position removes the able condition and it is common to all the schemes presented in
feedback linearizability property, even though the position is this note.
just the integral (linear operation) of the speed. This is easily The scheme can be easily simplified to a velocity control
explained by noting that the (exact) linearized system is in the z* scheme. The number of integrators is then reduced from three
coordinates and neither nor 2; is the speed. to two and the term a,(ql - vld)is eliminated from the expres-
Remark 3: Another interesting observation is that 2; = 0 is sion for c. In other words, this control approach can be used for
just the steady-state operating curve given in Section I1 (Fig. 2 ) both velocity and position control as opposed to the previous
and is shown again in Fig. 3. A constant operating condition approach which was only valid for velocity control.
(equilibrium point) in the z* coordinates is of the form z,* = The question of internal stability of the scheme has not been
(zT,,O) with reference input uref = c 4 z 2 0 / c s so that a ( z , ) + ascertained nor is it an easy task. This is indeed a limitation
P ( 2 , , ) U r e f = 0. to this approach. While 0 = x l , w = x 2 are guaranteed to be
bounded, such guarantees are not available for U and x3. The
IV. GENERALIZED
CONTROLLERCANONICAL
FORMS same problem arises in the context of input-output linearization
FOR NONLINEARDYNAMICS discussed in the next section, where the problem is resolved
Recently, Fliess [4] has proposed a generalization of the exact through the concept of zero dynamics.
feedback linearization technique that involves having dynamic
state feedback compensators. Specifically, as suggested in [4], we V. INPUT-OUTPUT
LINEARIZATION
define One can achieve input-output linearization with internal
771 4x1 stability by state feedback using the approach given in [7] (cf.
A . also [6] and [ll] for an adaptive version). We will see that this
772 =x, =x2 approach is simpler and guarantees trajectory tracking. In
A ..
773 = x1 = -c1xz - c 2 x 2 x : + c3x3u - ChTI-. (7) this approach, one uses full state feedback to achieve a linear
relationship between the (new) input and the output of interest.
Then, in these new coordinates, the dynamic equations are
With position as the output, let
given by
Let y 2 h ( x ) A x , so that y = Y f h + u Y g h + Y d h = Y f h =
i l = 772 x 2 and y = 9 ; h + u P g P f h + 9 d 9 f h= -cIx2 - c 2 x 2 x : +
i 2 = 773
c 3 x 3 u - c6rL. As P g h = 0 and P g T f h f- 0, the system has
relative degree 2 [ 2 , p. 1451. The control law
i73 = c,x,u + C3C$
- (2c,c,x2x, + ( c , c , + C3C& + c2c3x:)u
ClX2 + c 2 x 2 x : + C6TL + li
+ c:x2 + 2 c , c 2 x 2 x : + c ; x 2 x ; -
(11)
c3x3
+ 2 c 2 c , x 2 x : + ( c , c , + c2c,x;)T,,
results in the following (asymptotically) linear input-output
= C,X,c f (Y(X,U, 7~). (8) system
Setting y = c'.
c,x,u + a(x,U , T L ) = c (9) This control law obviously requires that x 3 be bounded away
from zero and the Same remarks regarding this condition
to define the feedback control law U ) results in a linear
here as for the previous controller. Taking the control objective
system in the new coordinates: a triple integrator from the new
to be the tracking of a prescribed reference trajectory y,(t) by
input c to the output 77, = x l . Let 77,d 6 e,, q Z d od 4i dOd/dt,
the output y ( t ) , we may set c according to
and 773d = ad 4 d o , / d t be the desired (reference) position,
speed and acceleration, respectively and set L' = y d + a2()id -)i) + al(yd - y ) (12)
L' = -(al(ql - +
q I d ) a 2 ( q 2- + a , ( q 3 - v3,))+ itd which results in the following equation for the tracking error
(10) e + a2e + a l e = 0 (13)
to obtain a trajectory tracking controller. To implement this where e ( t ) y ( t ) - y d ( t )and cy2, a 1 > 0 so that s 2 + a 2 s + als
controller, one must solve the nonlinear differential equation (9) is Hunvitz. It should also be clear that the full state must be
1665
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 38, NO. 11, NOVEMBER 1993

available (measurable) to implement this feedback law [i.e., (I 1) If x 2 = o < om, c4c3/c,c2, then W is bounded as rL and
and (1211. are bounded. This then implies that x 3 is bounded. If the
The original nonlinear system is third order (three state (linear) controller is designed such that the instantaneous speed
variables - 0, w , i,) while the input-output linearized system is does not exceed this value, bounded tracking is guaranteed. It
only second order. With the input-output linearization con- is also interesting to note that w, = c4c3/c5c2 = ( R a d ,+
troller, we have hidden (made unobservable) the third state R,)/K,K, which is the same bound derived in Section I1 for
variable xi (= i,) from the new input L’. In order to understand steady-state speed. Therefore, the condition for bounded track-
this situation and to ascertain the overall stability, consider the ing is satisfied in the normal operation of the machine.
following state-space transformation: Remark 1: This approach is obviously related to the previous
method of Fliess. Specifically, write
21, = TJx) 4x1
ZI2 = T 2 ( x ) & x, I; = f ( x , U ) .
z21 = T , ( x ) g -c+2 + c3x:/2 (14a) The input-output linearization approach consists in solving the
algebraic equation
x1 =Zll
x2 = z12 f(x,u) = L‘

so that y = U . This is particularly easy since f ( x , U ) is linear in


where 2, 4 [ z , , zI2IT= [ y >’I’ is the state of the above input- the control U. The method of Fliess reduces, in this example, to
output system and z2, e T 3 ( x ) & -c,x2 + c3x:/2 is a new letting
(transformed) state variable chosen such that ... d
Y = -f(X, U ) =g(x, U , Li)
dt
The system equations in the new coordinates are then and solving the differential equation
211 = 2 1 2 g(x, U , U ) = 1’

i,, = -C1X2 - c2x2x3‘+ c3x3u - C 6 7 L


so that y = L‘. This is essentially the same scheme except that
an extra state variable has been added, resulting in a triple inte-
grator, instead of a double integrator. The implementation is
more complicated in the Fliess method because of the need to
i,, = - c 5 ( - c 1 x 2 - c 2 x 2 x : + c3x3u - C 6 T L ) solve an extra differential equation. However, the benefit of this
approach would appear in situations where f ( x , U ) = U cannot
be solved analytically in a convenient form.
Remark 2: As in Fliess’s method, the approach can easily be
simplified to the velocity control problem. The number of inte-
grators is reduced from 2 to 1 and the scheme is easily modified
e T(z,, z 2 ) with z, [ z , , , z I 2 l T and z2 z , ~ . (15) accordingly, with no change in the singularity conditions, the
zero-dynamics or the bounded tracking.
The zero dynamics are then defined [2] by considering the
equation for z,(b z , ~ )with z , = 0, that is, V. CONCLUSIONS
In this note, we compared three approaches for nonlinear
As c4 = ( R o d f+ R F ) / L F ,it is clear that the zero dynamics are control: feedback linearization, generalized controller canonical
globally exponentially stable for rL = 0, and so zZ1is bounded if forms, and input/output linearization. T o establish a framework
7,- is bounded. By definition, this system is globally minimum
for comparison, we studied the example of a shunt DC motor. It
phase since its zero dynamics are globally asymptotically stable. was found that the conditions to apply the method of feedback
Note that z,, = 2: with zT given in Section 111 so that linearization were not satisfied for the position control problem.
i,, = z,*, with z,* also as given in Section 111. This is not In other words, the full three-dimensional state-space system
surprising since z: = T1(z) was chosen in Section I such that could not be linearized by a state transformation and feedback.
T 8 T l ( z ) = 0 and z , ~= T , ( x ) was chosen above so that This is probably not an uncommon occurrence, as the conditions
Y g T 3 ( x )= 0 was satisfied. for feedback linearization are rather stringent. While there exist
physical systems which are feedback linearizable (cf. [SI), the
Bounded Tracking conditions are often too much to ask for.
It was found for the velocity control problem, i.e., after
The above control law (11) and (12) results in bounded track-
eliminating a state variable (the rotor position) and thereby
ing, that is, x = [x,, x,, x,] is bounded in R 3 and y ( t ) converges
reducing the dimension from 3 to 2, the system was feedback
to y,(t). This follows from [6, proposition 7.2.11 whose interpre-
linearizable. In other words, the dynamics associated with the
tation gives interesting insight into the conditions that are
remaining two state variables (the rotor velocity and the field
required. To see this, let W = c3x:/2 so that
current) could be linearized by change of coordinates and state
d feedback. A similar situation occurred with the input-output
dt
w
- = c,x,X,
linearization, except that the dynamics that were linearized were
those associated with the rotor position and velocity, while the
remaining dynamics (associated with the field current) were
= -(2/C3)(C,C3 - c,c,x,)W + C S C l X , + CgL‘ +
C5C6TL. made unobservable.
1666 IEEE TRANSACTIONS O N AUTOMATIC CONTROL, VOL. 38, NO. 11, NOVEMBER 1993

To guarantee stability of the input-output linearization M. Bodson, J. Chiasson, R. T. Novotnak, and R. B. Rekowski,
scheme, it must be checked that the so-called zero dynamics are “High-performance nonlinear feedback control of a permanent
magnet stepper motor, IEEE Trans. Contr. Syst. Tech., vol. 1, no. 1,
asymptotically stable, so that their cancellation does not lead to pp. 5-14, 1993.
internal instability. This problem does not arise in the feedback A. J. Krener and A. Isidori, Linearization by output injection and
linearization approach, where the stability is trivially guaranteed nonlinear observers, Syst. Contr. Lett., vol. 3, 1983, pp. 47-52.
by the existence of a diffeomorphism between the original state W. Leonhard, Control of Electnc Driiw. New York: Springer-
and the transformed state and by the stabilizing control law Verlag, 1985.
S. Sastry and A. Isidori, “Adaptive control of linearizable systems,”
for the linear system. However, the derivation of the feedback IEEE Trans. Automat. Contr., vol. 34, no. 11, pp. 1123-1131, 1989.
linearization control law for the shunt DC motor required cer- M. Bodson and J. Chiasson, “Nonlinear and adaptive control of a
tain singularity conditions to be avoided and they were more shunt DC motor,” in Proc. IEEE Int. Conf: Syst. Eng., Dayton, OH,
stringent than for the input/output linearization approach. 1991, pp. 73-76.
P. D. Oliver, “Feedback linearization of DC motors,” IEEE Trans.
Another drawback of the feedback linearization approach is that Industrial Electronics, vol. 38, no. 6, pp. 498-501, 1991.
the transformed state is a nonlinear function of the velocity
while, for input-output linearization technique, the dynamics
were linearized from the new input to the output of interest (i.e.,
the position or the velocity). Finally, the feedback linearization
could only be used for velocity control, while both a velocity On the Observer Design for Descriptor Systems
control and a position control scheme could be derived from the
input-output linearization approach. P. C. Muller and M. Hou
The method using generalized controller canonical forms was
found to be closely related to the input-output linearization, a +
Abstrucr-For the descriptor system EX = Ax Bu, y = Cx + Du,an
conclusion which did not appear obvious a priori. The primary observer of the order p = rank E - p1 (0 I p1 I rank E ) is developed.
difference between the two methods is in the implementation It is not assumed that the descriptor systems are necessarily regular.
The derivation is straightforward using only an algebraic approach. The
of the control law and the input-output linearization approach existence conditions of the observer are also presented. These conditions
was found preferable in that respect. However, there probably are less restrictive than existing ones. Some illustrative examples are
exist examples where the generalized controller canonical form included.
approach has advantages (as, for example, when the input does
not appear linearly). 1. INTRODUCTION
Though not presented, simulations of the three control laws In recent years, a considerable amount of research has been
were performed to confirm our results. In general, the input- devoted to observer problems of descriptor systems, e.g., [1]-[9].
output linearization was found to be the simplest and least Under certain observability assumptions and with different
restrictive method. Further, a mechanism for adaptation can treatments, such as singular value decomposition [ 11, generalized
easily be implemented [6, section 7.31 which extends the useful- inverse [6], [9], geometric [7], algebraic [2]-[6], and the Drazin
ness of the scheme when parameters are unknown or vary inverse [8] approaches, observer problems of descriptor systems
slowly. were dealt with.
Note Added in Prooj5 After submission of this note, and The purpose of this note is to outline a simple observer design
publication of its preliminary version in [12], another paper [13] method for descriptor systems based on these previous works
appeared which addresses a similar problem. Results of the under less restrictive conditions. Moreover, unlike these previ-
present note not considered in [13] include the proof of the ous works, the regularity assumption (i.e., A , E are square and
boundedness for the input-output linearization controller and a IsE -AI + 0) on descriptor systems is dropped through this
comparison to the generalized canonical controller. note. This means that we consider the observer design for the
general case of descriptor systems, i.e., nonregular even non-
ACKNOWLEDGMENTS squarc descriptor systems.
The authors gratefully acknowledge helpful discussions with A n observer of the descriptor system can be used for system
Michel Fliess and Michel de Mathelin. supervision, fault diagnosis and feedback control as well as for
any other potential applications in descriptor systems. As an
example, the regularization of a nonregular descriptor system
REFERENCES needs such an observer within a feedback controller.
S. J. Chapman, Electric Machine Fundamentals. New York: The motivation for studying arbitrary descriptor systems stems
McGraw-Hill, 1985.
A. Isidori, Nonlinear Control Systems, second edition. New York:
from the following considerations. Firstly, a set of differential
Springer-Verlag, 1989. and algebraic equations of form EX = Ax + Bu arise very natu-
H. Nijmeijer and A. J. van der Schaft, Nonlinear@namical Control rally and conveniently in describing a plant. This is especially
Systems. New York: Springer-Verlag, 1990. true for a nonregular or even nonsquare system description.
M. Fliess, “Generalized controller canonical forms for linear and Therefore, in this sense we have the same reason for introducing
nonlinear dynamics,” IEEE Trans. Automat. Contr., vol. 35. no. 9,
pp. 994-1001, 1990. arbitrary descriptor systems as for introducing regular descriptor
P. C. Krause, Analysis of Electrical Machines. New York: systems at the very beginning. On the other hand, if x denotes
McGraw-Hill, 1985.
S. Sastry and M. Bodson, Adaptire Control-Stability, Coniwgence, Manuscript received January 10, 1992; revised May 22, 1992 and
and Robustness. Englewood Cliffs, NJ: Prentice-Hall, 1989. September 14, 1992.
A. Isidori and A. Ruberti, “On the synthesis of linear input-output The authors are with Safety Control Engineering, University of Wup-
responses for nonlinear systems,” Syst. Contr. Lett., vol. 4, pp. pertal, Germany.
17-22. 1984. IEEE Log Number 9208723.

0018-9286/93$03.00 6 1993 IEEE

You might also like