Professional Documents
Culture Documents
00
Printed in Great Britain. Pergamon Press Ltd.
~;) 1984 International Federation of Automatic Control
Key Words--Model reference adaptive control; nonlinear control systems; control system design; state
feedback; synthesis methods.
Abstract--The use of discontinuous control in adaptive model- parameter structure. Since they do not require any
following continuous control systems with nonlinear time- global convergence properties, these approaches are
varying plant is described. The first approach is based on the
theory of variable structure systems.Hyperstabilitytheory is the particularly well suited to the class of uncertain
basis of the secondscheme;the third one results by combiningthe time-varying plants under consideration.
previous approaches. The link among the three approaches is After defining the system and control aims, the
explored by studying the error dynamics. Practical implemen-
tation and applications are discussed. theory of VSS control is described in Section 3. The
hyperstability approach is elaborated in the
following section and a comparison between the
1. INTRODUCTION
VSS and the hyperstability approaches is given in
THE TECHNIQUE of model reference adaptive
Section 5, including a discussion of the mixed
control (MRAC) using Lyapunov and hyper-
approach. Aspects linked with the practical
stability theory has been widely described in the
implementation of such systems and some further
literature (Monopoli, 1974; Landau and Courtiol,
extensions of the approach are discussed in Section
1974; Landau, 1979). The designer specifies a model
6. A simple example is worked out both theoretically
the dynamic response of which he desires his actual
and by simulation in Section 7 in order to clarify and
plant to follow. M R A C relies upon the convergence
compare the three approaches.
of parameter identification schemes or a suitable
determination of auxiliary signals. The global
2. THE SYSTEM
stability of M R A C systems in the case of uncertain
The nonlinear time-varying plant satisfies the
nonlinear plants is problematic. In this review paper
differential equations
we shall describe three approaches to the synthesis
of adaptive model-following controllers (AMFC)
Y~(t) = A(x,t)x(t) + Bu(t) + h(x,t) (1)
for nonlinear time-varying plants with disturbances.
The first technique is based on the theory of
where x ~ R" is the plant state, u ~ R" is the control,
multivariable variable-structure systems (VSS)
h e R t is a disturbance vector and B0 = BoR(x,t)
(Utkin, 1974) which has desirable properties such as
where Bo is constant and R(x, t) is a square full rank
parameter invariance and disturbance rejection.
matrix. This structural assumption is equivalent to
Hyperstability theory is the basis for the second
the condition rank(B~, AB) = rank(B~) where
approach, whereas the third one is a suitable
B = B~ + AB. Clearly B = BoR = Bo(Ro + AR) =
combination of the previous approaches. These
B~ + AB. The additive form B = B~ + AB will be
techniques lead to control laws which are discon-
used in Section 3 and the multiplicative form
tinuous on linear switching hyperplanes. The
B = BoR in Section 4.
schemes do not incorporate any parameter
The designer specifies a linear time-invariant
identification algorithms, but the adaptation is
model
r e a l i z e d - t h r o u g h signal synthesis in a fixed
:~ (t) = / l ~ ( t ) +/~fi(t) (2)
* Received 17 October 1983; revised4 April 1984.The original where ~ ~ R" is the model state and fi E R m is the
version of this paper was not presented at any IFAC meeting.
This paper was recommendedfor publication in revised form by reference input./[ is a stable matrix and the pair (/i,/~)
guest editor L. Ljung. is assumed completely controllable. The objective of
t Dipartimento di Informatica e Sistemistica, University of thecontrolleristoguaranteethattheerrore = ~ - x
Naples, via Claudio 21, 80125 Naples, Italy.
:[:Department of Applied and Computational Mathematics, between the plant and the model states tends
University of Sheffield, Sheffield SI0 2TN, U.K. asymptotically to zero, i.e. lira e(t) = 0.
t~oo
559
560 A. BALESTRINO, G. DE MARIA and A. S. I. ZINOBER
without precise knowledge of the plant parameters independently of B. Shaked (1977) has shown that in
and disturbances. Applications of Ueq in the form of the linear time-invariant model-following control
a linear feedback controller, on the surface v = 0, problem order reduction is also the main feature of
would achieve the desired dynamics only in the case the controller design.
of known constant parameter values and distur- To ensure that the set of the switching
bances. The value of Ueq is effectively the average hyperplanes is reached and that subsequently the
value of u which maintains the state error on the sliding mode is ensured, it is necessary to choose a
surface v -- 0. The actual control u consists of a low- suitable control law (see Section 3.3).
frequency (average) component, Ueq, and a high-
frequency (chatter)component. Substitution of (10)
into (7) yields the nth order equation 3.2. Selection o f switching hyperplanes
The selection of the switching hyperplanes
: {I - B ( G B ) - ' G } { , 4 e + (,4 - A)x +/~ii - h}. v = G e - - 0 , is the central feature of VSS design.
(11) Utkin and Young (1978) have described the
technique of arbitrary eigenvalue assignment and
When the structural conditions (4) hold, we obtain links with the optimal quadratic regulator problem.
(Young, 1978a) Young, Kokotovic and Utkin (1977) have shown
that if the pair (/LB) is controllable then the
= {I - B ( G B ) - ' G } , 4 e . (12) eigenvalues of the reduced system (12) can be placed
in desired locations. The designer, having first
The invariance of the error dynamics with respect chosen the stable model (2) by taking into account
to x and fi is due the fact that the matching (or the structural constraints (4), chooses the desired
perfect model-following) conditions coincide with asymptotically stable error transient in the sliding
the invariance of the sliding mode if x and fi are mode by specifying the eigenvalues of the sliding
considered as 'disturbances' to the error dynamics. mode. The geometric approach (E1-Ghezawi,
For total disturbance rejection, namely Zinober and Billings, 1983) yields algorithms for
eigenvalue assignment, arbitrary specification of the
{I - B ( G B ) - I G } h = 0 (13) matrix GB and also allows partial selection of the
eigenvectors wi. The eigenvalues 2i are the
we require (Drazenovic, 1969) rank(B,h) = rank(B), transmission zeros (MacFarlane and Karkanias,
which is equivalent to (4). If (13) does not hold, 1976) of the triple (,4,/~,G) (Young, 1977; E1-
suitable error response may still be possible Ghezawi and co-workers, 1982; E1-Ghezawi,
(Zinober, E1-Ghezawi and Billings, 1982). Billings and Zinober, 1983). The determination of
VSS design is particularly useful in the case of the matrix G is equivalent to the output zeroing
time-varying systems where the system parameters problem in linear multivariable control theory
are uncertain. For instance, if B -- B¢ + AB, where where one desires to choose a suitable output matrix
/~ is the nominal value, invariance to all variations C where the output is given by y = Cx (Karkanias
AB results if and Kouvaritakis, 1979). The matrix
P = {I - B ( G B ) - 1G} is a projector which maps all
rank(B~,AB) = rank(B). (14) the columns of .4 into the nullspace of G. Some
useful relations between B, P and W = {Wl,W2.... ,
Even if (14) does not hold, Young (1978a) has shown w,_,,}, where wi are the eigenvectors spanning the
how suitable design can yield an asymptotically nullspace of G, are given in El-Ghezawi, Zinober
stable error transient. With structural constraints and Billings (1983) and a straightforward algorithm
(4), variations from a nominal value of A may also to determine a suitable set of switching hyperplanes
be handled. is presented.
Equation (12) describes the dynamic behaviour of
the error system whilst sliding on the intersection of
the m switching hyperplanes. A suitable coordinate 3.3. T h e control law
transformation (Young, 1977) decomposes the It is now necessary to choose the nonunique
system into n - m 'slow' modes with the eigenvalues control law (5) to ensure that the error attains the
2~eA characterizing the motion of the error switching hyperplanes v - - 0 in a finite time and
transient and a further m 'fast' modes. E1-Ghezawi, thereafter maintains the error e on the sliding
Zinober and Billings (1983) have used a geometric manifold. The control may have many forms; for
approach with projectors to study the properties of instance consider the following:
the sliding mode. The system during the sliding
mode has an effective reduced order n - m v~ith the ui = gi(e,x,fi) + fi(e,x,fi)sgn vi,
eigenvalue set A specified by the matrix G, i-- 1,2..... m. (15)
562 A. BALESTRINO, G. DE MAR1A and A. S. 1. ZINOBER
g = Ke (16)
with (~2 and ~P2 positive scalar functions. In order to Another possible structure is to use m fixed vectors f~
satisfy the inequality (25) we must have in
m
d~[Ivl[ < vrRM~bl(V) < d211vtl + d3llvl[' (28) ~ i fisgn(vT RMfi) (36)
1
d411vll < vTRMffl(v) < dsllvll + d6llvllp (29) which is more appealing when using micro-
processors in the implementation of the controllers.
where d~ are positive constants and r and p are Many other interesting adaptation laws can be
positive integers. Assuming the structure obtained by combining the above forms. A par-
ticularly appealing structure from the theoretical
~2 = Co + cxl[~ll + c21lxlt (30) viewpoint has the form
a similar argument applies to the choice of the other (A T + - PBoNoB~/p)P + P(A + ctI
constants ci. In order to complete the design of the - BoNoB~P/p) = - HTH (39)
plant control u (24) we must also choose the matrix
M(x). First suppose that a matrix M(x) exists such where H is any full rank matrix.
that In equation (39) we can put ct = 0, p = ~ , or, if we
R M + MTR T = N(x) > No > 0. (33) choose 0t > 0, a suitable p must be specified in order
to get a positive definite matrix P (Anderson and
Equation (33) leads in a natural way to classify the Moore, 1971).
different adaptation laws. If we know R(x) exactly
and can compute in real time R - l ( x ) an obvious 5. DESIGN PROCEDURES: COMPARISON AND
ERROR DYNAMICS
choice is M ( x ) = R - l ( x ) leading to perfect non-
linear compensation (Freund, 1975). If the com- 5.1. Stability
putation time of R - l ( x ) is too long we can use By using the hyperstability approach the global
M(x) = RT(x); but ifR(x) is not exactly known, as in asymptotic stability is directly ensured; indeed,
the case of uncertain systems, other choices must be choosing as Lyapunov function
made trying for instance constant matrices M. Of
course, if we know that R(x) is a positive definite V(e) = eXpe (40)
matrix, an obvious choice is M = I (Balestrino, De
Maria and Sciavicco, 1981). Having determined a with P solution of the Riccati equation (39), and
matrix M(x) satisfying (33) some possible choices assuming the plant control as
for the structure of ~b~(v) and ff~ (v) are (Balestrino,
De Maria and Sciavicco, 1982b; Gutman, 1982) u = M(x){KI~ - K2x + g3fi + q~ + ~ + Ge/p}
(41)
v/(vTQv) l/z, any Q > 0; v/~ilvd) (34)
from (19), (20) and (40) we obtain
or v/llvll (unit vector adaptation law) where Ilvll is
any norm. If the matrix L = R M is diagonally l?(e) = er(PA + ArP)e - er(GTRMG
dominant, i.e. lil> Efli~],i ~ j we can choose the + GTMrRTG)e/p + 2eTGTwl (42)
relay type adaptation law (Monopoli, 1966; Devaud
and Caron, 1975; Balestrino, De Maria and where
Sciavicco, 1980) w~ = B~(A - A)x + B ~ - B~h
sgn(v) = {sgnvl, sgnv2 . . . . . . sgnvm}T. (35) - R M ( K ~ - K2x + K3fi + ~b + ~,). (43)
564 A. BALESTRINO, G. DE MARIA and A. S. I. Z1NOBER
Choosing tho Lyapunov function V(v) = VTpm¥, controllers are fixed with no on-line identification
as in (40), we obtain an inequality similar to (44) being required. Robustness is achieved by the
where the feedforward and feedback terms in (54) inherent adaptation arising from the switching
along with the adaptation terms ~b and ff must be structure which effectively alters the adaptive
chosen so that feedback paths. An open problem is the relaxing of
vXpmB0mw,n < 0 (55) the full state availability requirement in connection
with a nonlinear time-varying plant.
in order to ensure that v = 0 is asymptotically stable The plant model (1) will not have included any
in the large with ~ as an upper bound on the rate of high frequency parasitic dynamics present and
convergence. Note that the mixed approach may other imperfections in the adaptation mechanism
prove computationally more expedient because the and switching structures. The high frequency
solution of a Riccati equation of lower order is chatter component u¢ in the sliding mode is
required. undesirable since high order modes dynamic may be
excited. The inclusion of a small time delay,
5.4. Summary of design procedures hysteresis or a low pass filter in the input channels
The above procedures are summarized as follows. yields nonideal sliding motion without excessive
chatter motion and Utkin (1974) has proved that
Required prior information. the state path remains within an ~-neighbourhood of
v = 0. Ambrosino, Celentano and Garofalo (1983)
(1) The plant must be specified in the state variable
give an example with the control input of the form
form; all the state variables must be available.
(2) A model, whose order is equal to the order of the
v/Ovl + 6) rather than the pure relay v/Iv] = sgnv; the
high frequency component is effectivelyremoved and
plant, must be assigned; the perfect model
the system response is satisfactory.
following conditions (4) must be satisfied.
Alternatively, we may ensure that the chatter
(3) Specify the structure of the plant input matrix B,
frequency is outside the bandwidth of the plant by
i.e. B = Bc + AB or B = BoR.
injecting high frequency dither (Zames and
Shneydor, 1976; Balestrino, De Maria and
Design procedure. As a design procedure select the
Sciavicco, 1983a,b,c).
VSS or the hyperstability approach or the mixed
The most significant applications of VSS and
one.
hyperstability approaches to the design of AMFC
(1) VSS approach. for nonlinear or uncertain plants occur whenever
(i) Choose G, i.e. the sliding hyperplanes, the plant dynamics are so fast that identification or,
taking into account that the dynamics more generally, on-line computations make other
during the sliding motion are given by approaches infeasible. The main areas of appli-
(5.12). cations are electrical power systems, including
(ii) Select u in order to ensure the asymptotic electrical drives, robotics and aerospace. High
stability in the large ofv = 0; now use can be performance electrical drives must ensure a good
made of the Lyapunov method or of the time response independent of the variations of
hierarchy of controls method. electrical and mechanical parameters. Practical
(2) Hyperstability approach. schemes using discontinuous controllers have been
(i) Choose an upper bound on the convergence described by Balestrino, De Maria and Sciavicco
rate to zero of the generalized error e, say ~. (1983a), De Maria and Marino (1983), Sabanovic
(ii) Find a solution M of (33) with N > 0. and co-workers (1983), Bilalovic and co-workers
(iii) Solve the Riccati equation (39) for P and a (1983), and Dote, Manabe and Murakani (1983). In
suitable p. robots with fast dynamics we must tackle the highly
(iv) Let be G = B~P; select u (41) so that the nonlinear structure of dynamical equations and the
passivity condition (23) is satisfied, or if it is lack of knowledge of the payloads. An effective
required that v = Ge = 0 be invariant and reduction of the computer burden is achieved by
attractive select u in accordance with (47). equipping the mechanical structure with an
(3) Mixed approach. adaptive model reference control (Young, 1978b;
(i) Choose G taking into account (51). Balestrino, De Maria and Sciavicco, 1983b; Slotine
(ii) With the error dynamics in the form (50) and Sastry, 1983). Young and Kwatny (1982) have
apply the hyperstability approach to the studied turbine overspeed protection.
subsystem (50b).
7. EXAMPLE
6. IMPLEMENTATIONAND APPLICATIONS In order to illuminate more clearly the three
The proposed design techniques are readily reviewed approaches and the kinship between them,
implemented using microprocessors since the let us consider the position control of the pendulum
566 A. BALESTRINO,G. DE MARIA and A. S. I. ZINOBER
e v = (el,Ol) = (0 - 0, (i - 0t (641
v = Ge = glel + el = 0. (65)
y
where g is the gravity acceleration and where g(e,x,~) must be chosen in order to satisfy the
sliding condition. Assuming gl = 7, ao = 2, a~ = 3,
A(0) --- (cos0 + tolt:)2Z~/t~o. (S8) equation (67) becomes
Choosing ll/lo -~ 1/2, g/lo = 10, and rnlg = 1 yields u = 2.25 sgnv(5.8 + 0.24 0
+ 12el - 4~1 + 20 + 30 - ~1). (68)
0 = 0.5 sin0(1 + 0.5cos0)0z/A(0)
- 10sin0(1 + c o s 0 ) / A ( 0 ) + T/A(O) (59) 7.2. Hyperstability design procedure
In order to set an upper b o u n d on the
and
convergence rate of the error e to zero a suitable
• m a x A(0) = 2.25; rain A(0) = 0.25. (60) feedback M G e / p must be chosen. Assuming
0 O
c~ = g2 = 49, M >_ 1.125, No = 1 (69)
Assuming x T = (0 0), iT = (0 0) and u = T plant
and model equations can be rewritten as follows:
from the solution of the Riccati equation (39) we
have
0
- 10(1 + cos0) sin0 sin0(1 +1 0.5cos0)0) 1 416
P=- t 80 ; l/p> 136.4. (70)
A(0)
7.3. M i x e d procedure
Assuming G = (7, 1), ~ = 49 and taking into
account that the error dynamics can be rewritten as -.0~
et = - 7 e l + v
(75)
= 4v + w - 30e~ = 4v + w,~ -.18[,.
0 .3 .6 .9 1.2
FIG. 5. Plant control input for the VSS approach.
the Riccati equation with H T H = 1 = P yields
l/p = 107.
Setting Bo,,, = 1, choosing M = 1.125 in order to
satisfy (33) with No = 1, the input u (54) becomes 8. C O N C L U S I O N S
Three schemes for adaptive model-following
u = 120v + 2.54(130et + 20 + 30 - ~1 control of nonlinear uncertain systems have been
discussed; one based directly on variable structure
+ 5.8 + 0.2402)sgnv. (76)
methods, the second derived from a hyperstability
approach, and the third obtained by a suitable
Figure 4 shows the response of the plant with i = 0,
combination of the previous ones. All give robust
and x(0) = (n/2,0), when the VSS approach (curve
discontinuous control with related error dynamics
a), mixed approach (curve b) or hyperstability
and do not require any adaptive identification
design with a = 49 and v = 0 attractive and
schemes.
invariant (curve c) are used. Figures 5 and 6 show
The adaptation is realized by a suitable signal
the torque T applied in the case of VSS and mixed
synthesis in a fixed parameter structure; adaptivity
approach respectively.
or robustness is achieved through 'high gain' effects
Some worked examples of design for the multi-
locally in the neighbourhood of the switching
input-multi-output case are reported in Young
hyperplanes.
(1978a) and Zinober, E1-Ghezawi and Billings
The design procedures are theoretically well-
(1982) for the VSS approach, and in Balestrino, De
founded for nonlinear time-varying or uncertain
Maria and Sciavicco (1983a,b,c) for the hyper-
plants while the classical model-following control
stability approach.
techniques can be used only if the plant may be
considered linear and time-invariant during the
adaptation process.
"~/2
,;,'r/4
0 .3 .6 .9 1.2
FIG. 4. Free evolution of the plant output 0: (a) VSS design, (b) -I
mixed approach, (c) hyperstability design with ct = 49 and v = 0 .3 .6 .9 1.2
attractive and invariant. FIG. 6. Plant control input for the mixed approach.
568 A. BALESTRINO, O. DE MAR1A and A. S. I. ZINOBER