Professional Documents
Culture Documents
Wilson J. Rugh
A framework for the design of "idealized" not used for design purposes because of its design problems, the control laws thatmust be
gain-scheduled controllers for nonlinear sys- complexity. scheduled are becoming more complex. As
tems is discussed. Based on this framework complexity increases, scheduling will become
Characteristics of Gain Scheduling
questions are raised, and implications are more difficult to address in a satisfactory man-
A number of characteristics o f gain ner by ad hoc methods.
drawn for 'practical' design situations.
scheduling are more-or-less apparent, even Another aspect of gain scheduling is that it
Introduction from the brief description above. These will is inherently local in nature. The overall per-
be discussed briefly in terms of general ad- formance of the control system must be check-
Realistic models of engineering systems
vantages, and disadvantages. ed by extensive simulation studies. In a sense,
often are nonlinear. One consequence is that
The main advantage of the gain scheduling this is the trade-off that results from the ad-
the dynamical behavior of a system to be
approach is that linear design methods are vantage of using linear design methods. How-
controlled changes with the operating region.
applied to the linearized system at each operat- ever it should be noted that in actual
A typical approach to this situation has been
ing condition, and the wealth of linear control applications most direct nonlinear methods
to apply the notion of gain schedulin<g.The
methods, performance measures, design intui- will be local as well. That is, the assumptions
first step is to linearize the model about one or
tion, and computational tools can be brought required for global application of, say, feed-
more operating points (also called set points,
to bear on control design for general, multi- back linearization seldom will be satisfied by
equilibrium points, or operating conditions).
variable nonlinear systems. For example, fre- complex, practical design problems. In any
Then linear design methods are applied to the q u e n c y - d o m a i n n o t i o n s , or q u a d r a t i c case, application of gain scheduling to com-
linearized model at each operating point in performance indices, can be used in specify- plex design problems can lead to a large
order to arrive at a set of linear feedback ing performance, and output-feedback control simulation burden.
control laws that perform satisfactorily when techniques can be used, while of course these
the closed-loop system is operated near the tools are much less well understood in an Toward a Framework
respective operating points. The final step is explicitly nonlinear setting. Also, modern It should be clear that gain scheduling
the actual gain scheduling, which is intended methods of robust design for linear systems yields a closed-loop system that is nonlinear,
to handle the nonlinear aspects of the design can be applied to counter uncertainty in the at least in the scheduling variables. Interest-
problem. The basic idea involves interpolating plant parameters, with the expectation, or at ingly enough, given the wide application of
the linear control law designs at intermediate least hope, that this robustness will be in- this approach, there has been very little study
operating conditions. That is, a scheme is herited by the final gain-scheduled design. of gain scheduling as a nonlinear control
d e v i sed f o r c han g i ng (scheduling) t h e Other advantages include the fact that a method that derives from certain basic prin-
parameters (gains) in the linear control law gain-scheduled control system has the poten- ciples of nonlinear systems and control theory.
structure based on the monitored operating tial to respond rapidly to changing operating Indeed, there is scant information available in
condition. conditions, and that the real-time computa- the literature on gain scheduling from any
Usually the model is arranged so that the tional burden typically is within modern point of view. An elementary introduction to
operating condition is specified by the values capabilities. However, these significant ad- gain scheduling is provided in [ 1, ch. 91, along
of one or more exogenous signals, or vari- vantages are partly offset by some difficulties. with references to some particular applica-
ables, and the gains are scheduled according One difficulty is the selection of ap- tions. A conceptual viewpoint involving al-
to the instantaneous values of these schedul- propriate scheduling variables. General char- gebraic representations is advanced in [2]. The
ing variables. Of course, this description is acterizations and prescriptions have not been recent doctoral dissertation [3] contains an
simplified - sometimes an internal variable particularly useful, except for rules of thumb analysis of gain scheduling from the perspec-
also is used for scheduling, or perhaps a vari- such as "schedule on a slow variable." In tive of robust, linear stability theory. Gain
able from another 'channel' of the design that practice, scheduling variable selection usually scheduling is discussed from a viewpoint
was decoupled for reasons of simplicity. In is based on the "physics" of the situation, and similar to that here in [4]. However there are
fact, usually the exogenous variables would be on particular characteristics of the model. important differences in the formulation, par-
state variables in a more complete model of Perhaps the major difficulty in gain- ticularly relating to the specification of "trim"
the situation, but the more complete model is scheduling design is the selection of the conditions. Finally, it should be noted that
scheduling procedure, that is, the program by results from research in 'families of linear
which control-law gains are changed as a systems' can have application in a gain-
Presented at the American Control Con- function of the scheduling variables. This scheduling context. See, for example, the sur-
ference. M a y 23 - 25, 1990, San Diego, aspect is seldom discussed in the literature, but vey in [ 5 ] . However, the focus on global
California. Wilson J . Rirgh is with the Depart- it appears that scheduling currently is an art, results, and on the algebraic nature of the
ment of Electrical & Computer Engineering. and simple curve-fitting approaches are used. parameter dependence, are not particularly
The Johns Hopkins University.Baltimore. M D In fact, linear interpolation seems to be the relevant in typical gain-scheduling problems.
21218. This ~ " was 4 sponsosed by the Air standard method. However, with the introduc-
Force Ofice of Scienti5c. Research, Air For-<,e tion, and increasing adoption, of more power- The objective in developing an analytical
Systems Command. USAF. ful techniques for robust, multivariable, linear framework for gain scheduling is to make use
January I99 I 81
w(t) -+ W E r as t 4 00 yields x(f) -+ x ( w ) and In fact, we can choose an arbitrary, smooth (27) that K 3 ( w ) # 0 for WE^. Therefore, in
y ( t ) --i, y ( w ) as t 4 w. function z ( w ) , and let this choice fix K 2 ( w ) as addition to the eigenvalue placement, K ~ ( w )
per (29). Then a natural solution is provided can be chosen to improve regulator response
Regulator Problem: to the scheduling signal.
by the gain scheduled control law
Dynamic Feedback Example 3: For the plant in Example 1, the
It tums out that even in the simple regulator scalar, integral-error compensator takes the
problem there are advantages to using form
dynamic state feedback. It is natural to con-
sider smooth feedback control laws of the Z(f) = y(t)
form
Placing the three linearized, closed-loop
u ( f ) = k ( x ( t ) , w(t), z ( f ) ) system eigenvalues at - 2, regardless of W ,
leads to
(Verification of (25) and (26) is quite similar
to the static feedback case.) Again, this control
Obviously ; ( f ) is thep-dimensional state of
law includes terms linear in the plant and
an integral-error compensator.
compensator states, with a rather complicated-
In this setting, the requirement that the
looking bias term (though there is no reason,
closed-loop system ( I ) , (24) has an operating
at this point, to pick a nonzero function z(w).
point with operating-point output y ( w ) for
Each of these terms is scheduled by the instan-
each w(t) = W E r reduces to the existence of
taneous value of the scheduling variable,
a smooth function z ( w ) such that
while the compensator simply integrates the
error between the actual output and the instan-
taneous trim value.
Also, the requirement that the correspond- Theorem 2 can be applied to show that the
ing closed-loop linearization has specified, closed-loop system (I), (30) has stability
negative-real-part eigenvalues, for each W E r, properties of the same type as in the static
can be described as follows. Let feedback case. The interest here is on the
additional properties provided by the dynamic
Kl(W) = ak/ax(x(w),w , z ( w ) ) control law, and one is a basic robustness
property that can be described as follows.
Suppose the plant (1) undergoes a local
perturbation. resulting in a perturbation of the
operating point functions x ( w ) and u ( w ) , as
well as a perturbation of the linearized plant
Then the eigenvalues of the closed-loop coefficients A ( w ) , B(w). and so on. Assuming
linearized system are the eigenvalues of the that these perturbations are such that (27)
(n+ p) x (n+ p ) matrix retains negative-real-part eigenvalues for W E r,
the perturbed closed-loop operating point
family will remain asymptotically stable. At
such an operating point, corresponding to a
particular value of w , :(t) has a constant value,
For each WE^, these eigenvalues can be though a value that is perturbed from the
placed as desired, by choice of (smooth) K , ( w ) original z(w). In any case, from (24). since z ( t )
and K l ( w ) , since one consequence of the is constant and ~ f t =j w , the corresponding
standing assumptions is that operating point output is forced to be the con-
stant y ( w ) . This argument shows that the
integral-error compensator provides robust-
ness of the output trim in the face of small
plant perturbat ions.
is a controllable pair. Another advantage of a dynamic control
It remains todefineK2(w) andz(w), and then law is that the constraint that ties K 2 ( w ) to
compute a smooth k (x, MI, -7) such that (25) and K l ( w ) in the case of static control laws some-
(26) are satisfied. As in the static feedback case, times can be relaxed. While more detailed .Y,,(t)
the key to this is to note that differentiation of study is needed, it is clear that the integral-
(25) implies, upon using (26). error compensator (24) permits this in certain .?(I) = '
(31)
cases. For example, if m = q = 1 (scalar plant
K,(w)dx(w)l a w + Kz(w)+ Ki(w)az(w)/aw input and scheduling variable), then (29) can X!,,(t)
be satisfied for independently specified K i ( w ) ,
K2(w), and K 3 ( w ) by choice of : ( w ) since it
follows from the eigenvalue requirement on
82 /E€
Control Systems
- Kz(P'($(t)))3'($(t)) a set of state feedback gains such that each A
(w,, ) + B ( w f )Ki(wf ) , p = 1 ,..., P, has the
- K,(:~'($(t)))z(:~'(~:(r),)] desired eigenvalues. Then, in terms of devia-
tion variables, near the operating condition
i ( t )= v(t) -y(:-'($(f))) (34) specified by w,, the linear state feedback con-
trol law
for which the function $ ( w ) is invertible. Then In contrast to (30)this control law is linear in
at any operating point, knowledge of the con- w ( t ) and z(t), though typically a complicated,
stant state value i(~) is equivalent to nonlinear function of x ( t ) .
knowledge of the constant value of the Sometimes more can be done. For ex- is to be used. Using the relations ug(f)= u(t) -
scheduling variable w . Since gain scheduling ample, in the case of a single input, single U ( w , ) a n d a s ( t ) = . u ( t ) - x ( w , ) , thecontrol law
is based solely on operating point considera- output, and a single scheduling variable, m = in terms of the original plant variables has the
tions, it follows that the instantaneous values p = 4 = 1, (29) can be satisfied by setting form of a linear state feedback component and
of $ ( f ) can be used for scheduling. a constant bias:
Pursuit of this notion can lead to many K?(w ) = 0
different nonlinear control laws that satisfy the
requisite operating-point and linearization re-
quirements. Again, the most natural choice Of course (37) is simply an evaluation of (18)
from a scheduling viewpoint can be written as at the appropriate, constant value of the
scheduling variable, which is the reason that
(18) was chosen in the first place.
u(t) = K l ( 2 ( $ ( t ) ) X ( f ) + K&2(t))M.(t) r (35)
For intermediate operating conditions, the
Then the surrogate-scheduled control law set of control laws in (37) is scheduled (inter-
(34) is independent of the scheduling variable, polated). That is, functions k l ( w ) and & w )
(33) and appears as simply a nonlinear, dynamic are computed - often by linear interpolation -
state feedback control law! In other words, the such that
where again there are linear gains that place gain scheduling is completely implicit, and kI(,,
= K,(w,)
the linearized closed-loop eigenvalues, and a measurements of the scheduling variables are
bias term, all depending on the instantaneous not required. &wJ = u(w,) - Kl(w,)x(w,)
value of the components of ,?(t). Note, in com-
parison with ( 1 8), that this control law is linear 5.€.iumple: With K i ( w ) and K3(w) as in p = 1 , ... , P (38)
in the scheduling variable, but typically is a Example 3, ( 3 5 )gives
rather complicated, nonlinear function of the and then the gain-scheduled control law
state. Verification of (14) and (15) in this case K2(W) = 0 . z ( w ) = swig
takes a little more work. While (14) and the
second part of (15) are easy (since at the The corresponding control law (34) is
is used. Notice that at intermediate operating
o p e r a t i n g p o i n t w , :-'($(t)) becomes
A A conditions neither the output trim nor the
x-'(x(w)) = w ) , the verification of the first part
linearized closed-loop eigenvalues will be
of (15) is a lengthy exercise in the chain rule.
:(t) = J ( t ) correct in the typical case.
One approach to improving the situation is
4. E.uample: Returning to Example 1, since Less-Idealized Gain Scheduling
to make use of additional plant information. In
x2(w) = w , an obvious choice for surrogate, In a realistic design setting, the informa- many settings it is not difficult to obtain the
internal-variable scheduling is 2?(f)=.vl(t). tion available about the plant often is much operating point functions U ( w ) and x ( w ) in
Then ;-'($(r)) = x2(t), and substituting the more limited than that assumed so far. For consonance with the giveny ( w ) .Then it clear-
various computed quantities from Example 1 example, suppose that from the given trim ly makes sense to interpolate the state feed-
into (33) with this argument change gives the conditiony ( w ) . only a set of distinct operating back gain values only. For having determined
control law points is computed - perhaps from experi- k i ( w ) , the control law
ments on a simulation. Furthermore, suppose
that the only additional plant information is U ( ? ) = kI(M'(t))X(t)
the set of plant linearization coefficient
Exactly the same considerations can be matrices at these distinct operating p i n t s . We + [U(Mft)) - kx(Mff))x(Nf))] (40)
applied in the dynamic control law case, where assume, of course, that the standing assump-
the general form for a surrogate-scheduled tions are satisfied at these operating condi- will provide the correct output trim at both the
control becomes tions. design and intermediate operating conditions.
To be specific, suppose that corresponding (Simplyevaluate(40)atw(t)=w , s ( t ) = x ( w ) . )
to the distinct values w l , . . . . w p of t h e However, the eigenvalue locations at inter-
scheduling variable, the available information mediate operating conditions remain incor-
is the set of coefficient matrices A(w,, ), B(w,,). rect, in general.
E(w,, A C(wP ), D ( w p 1, and F(wP )for p = Another course of action, one that does not
1 ,.... P. and the valuesx (w,, )and U (w,, ).p = require more plant information, is to use an
1 ,..., P. This information suffices to compute integral-error compensator to correct the output
Januaw I99 I 83
trim. First compute Ki(w,)andK~(w,)toplace servations can be extracted in summary form, References
the eigenvalues of (27) at desired locations for and put into the gap. I, I .I K.J. Astrom. B. Wittenmark. Adontire Control.
the operating points wi ,...,w p Then these two Potentially the most important observation is Addison-Wesley Publishing Co., Reading. MA,
sets of gains can be interpolated to obtain kl(w) that dependent gains should not be scheduled Chapter 9. 1989.
and k3(w). Also the bias term in ( 3 0 ) can be independently. For example the bias term 121 M.K. Sain. S. Yurkovtch, "Controller Scheduling:
interpolated to obtain &w) such that should be scheduled in terms of the scheduled A Possible Algebraic Viewpoint," A-oceedin,qrofthe
eigenvalue-placement gain, and the functions Americuii Control Confereiic.e, pp. 261 - 269. 1982.
&wp) = u(w,) - Kdw,)x(w,) - K4wp)z(w,) describing the operating point variation. . 131 J.S. Shamma, "Analysis and Design of Gain-
Even for the simple state-feedback regulator Scheduled Control Systems," PhD Thesis, Depar-
problem. the use of integral-error feedback in ment of Mechanical Engineering, Massachusetts
nstitute of Technology, May, 1988.
the linear control law designs at fixed operat-
ing points can provide usehl performance 14)W.T. Baumann. "Discrete-Time Control of Con-
The corresponding control law, based on (30). tinuous-Time Nonlinear Systems," Proceedinp of
properties when incorporated into the gain
the Coiifrrriir.e on lrformution Sciences urid Sys-
scheduling. These include robustness proper-
tenis, Baltimore. MD, pp. 119 - 124, 1989.
ties. and also freedom to independently
[SI E.D. Sontag, "An Introduction to the Stabiliza-
specify various terms in the control law.
~ ( r =) y ( r ) - y ( M ' ( r ) ) (42) tion Problem for Parameterized Families of Linear
In some cases - see Example 4 - a nonlinear Systems." ContemporuryMathematics, Vol. 47, pp.
again will supply the correct output trim value control law can be found that satisfies the re- 369 - 399. 19x5.
at intermediate operating coriditions. How- quirements of a gain scheduled control law, but 161 D.A. Lawrence, W.J. Rugh, "On a Stability
ever the eigenvalue locations of the linearized that does not depend explicitly on the scheduling Theorem for Nonlinear Systems with Slowly-Vary-
closed-loop system at intermediate operating variable. This issue of so-called surrogate ing Inputs,'' IEEE Trcrrisactionson Automatic Con-
conditions typically remain incorrect. In any scheduling isone thing that makes gain schedul- trol, Vol. 35,No. 7. pp. 860 - 864, 1990.
case, trim correction at intermediate operating ing difficult to pin down from a theoretical [71 W.T. Baumann, W.J. Rugh, "Feedback Control
conditions again provides an important reason perspective. On the other hand, implementing a of Nonlinear Systems by Extended Linearization."
to consider dynamic feedback. gain-scheduled control law in part by sut~ogate IEEE Tr-umac~rioiis oii Automatic Control, Vol. AC-