You are on page 1of 13

200

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 40, NO. 2, FEBRUARY 1995

Adaptive Control of Plants with Unknown Hystereses


Gang Tao, Member, IEEE, and Petar V. KokotoviC, Fellow, IEEE
Abstruct- For a system with hysteresis, we present a parameterized hysteresis model and develop a hysteresis inverse. We then design adaptive controllers with an adaptive hysteresis inverse for plants with unknown hystereses. A new adaptive controller structure is introduced which is capable of achieving a linear parameterization and a linear error model in the presence of a hysteresis nonlinearity. A robust adaptive law is used to update the controller parameters and hysteresis inverse parameters, which ensures the global boundedness of the closed-loop signals for a wide class of hysteresis models. Simulations show that the use of the adaptive hysteresis inverse leads to major improvements of system performance.

I. INTRODUCTION
YSTERESIS phenomena caused by stiction, magnetism, or gears with blacklash commonly exist in physical systems [ 11-[6], and often severely limit system performance such as giving rise to undesirable inaccuracy or oscillations, even leading to instability. Hysteresis characteristics are generally nondifferentiable nonlinearities and usually unknown. Most results of adaptive control theory are for linear or differentiable nonlinear systems, and are not applicable to systems with nondifferentiable nonlinearities. It is a challenging task of major practical interest to develop adaptive control schemes for systems with unknown hystereses. In this paper, we will pursue this task. The controlled plant consists of a linear part preceded by a hysteresis characteristic, that is, the hysteresis is present at the actuator of a linear part. The linear part can be either known or unknown, and the hysteresis is assumed to be unknown. The objective is to achieve stabilization and tracking in the presence of the unknown hysteresis. To solve this problem, we will use the adaptive inverse approach which was proposed in [7]-[9] for adaptive control of plants with nondifferentiable nonlinearities such as dead-zone and backlash. In [ 7 ] , an adaptive dead-zone inverse control scheme was proposed for systems having an unknown dead zone at the input, using full state measurement. In [8], adaptive dead-zone inverse control designs for systems with unknown dead zones were developed, using only the output measurements. In [9], the adaptive inverse approach is used for adaptive control of systems were designed: one for systems with an unknown backlash and a known linear part, and the other for both the backlash and the linear
Manuscript received March 2, 1992; revised December 2, 1993. Recommended by Past Associate Editor, A. Arapostathis. This work was supported by the National Science Foundation under Grants ECS-9203491 and ECS9307545, by the Air Force Office of Scientific Research under Grant F-4962092-5-0495, and by a Ford Motor Company grant. G. Tao is with the Department of Electrical Engineering, University of Virginia, Charlottesville, VA 22903 USA. P. V. KokotoviC is with the Department of Electrical and Computer Engineering, University of California, Santa Barbara, CA 93 106 USA. IEEE Log Number 9407562.

part unknown. In this paper, we will employ an adaptive hysteresis inverse cascaded with the plant to cancel the effects of hysteresis so that the remaining part of the controller can retain its linear structure. The paper is organized as follows. In Section 11, we present a hysteresis model characterized by a set of parameters, and formulate the control problem. In Section 111, we present a mathematical model of a hysteresis inverse and its implementation. In Section IV, we develop a parameterization for an estimate of the hysteresis inverse. In Section V, we design an adaptive controller for plants with a known linear part and an unknown hysteresis. In Section VI, we present a solution to the more general adaptive control problem in which both the linear part and hysteresis are unknown. In Section VII, we present simulation results for the proposed adaptive control schemes. Models of backlash, electromagnetic, and other types of hysteresis can be found in [1]-[6]. However, a general hysteresis model would not be convenient because of its complexity. We will use a simplified hysteresis model which captures most of the hysteresis characteristics and is useful for parameter adaptive control. We will show that our hysteresis model has a parameterizable right inverse which cancels the effect of the hysteresis when cascaded with the hysteresis. An adaptive hysteresis inverse is implemented with parameters updated on-line by adaptive laws. When the adaptive hysteresis inverse is used for control, the effect of the hysteresis may not be completely cancelled, and there is a control error we express in two parts: first, a parameterizable part, and second, an unknown but bounded part. For plants with an unknown linear part, the usual linear controller structure is modified in order to achieve a linear parameterization and a linear error model, which are both crucial for the development of robust adaptive laws. As a result of the linear parameterization of the adaptive hysteresis inverse and the new controller structure, robust adaptive laws can now be used to guarantee the closed-loop signal boundedness for a wide class of hysteresis characteristics. Our simulations show that major improvements of system performance have been achieved by the adaptive hysteresis inverse. 11.
R A N T AND CONTROL OBJECTIVE

The plant to be controlled has a linear part and a hysteresis characteristic H ( . ) as its input:
Y ( t ) = G(s)[uI(t),

4 t ) = H(v(t))

(2.1)

where G ( s ) = k p ( Z ( s ) / P ( s ) ) Z , ( s ) , P ( s ) are monic polynomials, k p is a constant scalar, y(t) is the measured output, A w(t) is the control input, and H ( . ) = H ( m t , et, mb, C b , m,,

0018-9286/95$04.00 0 1995 IEEE

TA0 AND KOKOTOVIC: PLANTS WITH UNKNOWN HYSTERESES

20 1

< w(t) < w3, u ( t ) = m t v ( t ) + ct for it@) < 0; when < mb, on the half-line (2.3) with 114 < v(t) < w2, 7i(t) = mb?l(t) f cb for it(t) > 0.
711

mt

The signs of iL(t) and it@) are not restricted on other parts et, ~ ( t2 ) w3; of these two half-lines: u ( t ) = 7ntv(t) U ( t ) = 7rib'U(t) C b , 'U@) 5 114; and U ( t ) = rrLt'U(t) f C t , 1 1 1 < v ( t ) < 7 ~ 3when mt < mb or u ( t ) = mb'u(t) Cb, 0 4 < u ( t ) < w2 when mt > mb. For example, the half-line u ( t ) = mbv(t) Cb is bidirectional when mt > mb. The motion of ~ ( tand ) v(t) on the half-lines (2.2)-(2.3) and the segments (2.4)-(2.5) and inside the hysteresis loop can be mathematically described as

mtit(t)

Fig. 1. Hysteresis model.

e,, ml , cl; .) is parameterized by constants mt, et , mb, Cb , m,, c,, ml, cl and described by two half-lines (see Fig. 1):

u ( t )= mtv(t)

+ et,

v(t)

> w1 =

et
~

+m m
mt

mb'b(t) (2.2)

7121 -

li(t) =

and two line segments:

u(t) =m,.(v(t)

e,),

112

5 v(t)< I J ~ =

et m,c, m, - wit

(2.4)

rri;ii(t)
mlii(t)

where u1, 112, 113, 114 are the values of v(t) at the four opposite "corners" of the quadrilateral. Along the segments, the time derivatives of u ( t ) ,w ( t ) are of constant sign, namely, ?i(t) > 0, it@) > 0 for u ( t ) = m,(v(t)-c,), and li(t) < O , l i ( t ) < 0 for u ( t ) = rril(w(t)-cl). The hysteresis phenomena occur inside the loop formed by the half-lines (2.2)-(2.3) and the segments (2.4H2.5). Inside the hysteresis loop, the relationship between u ( t )and v(t) is forit(t) < 0 u ( t ) = mtw(t)+ c d ( t ) mbv(t) + c U ( t ) forw(t) > 0

if ~ ( t 2)w3, or if 114 < 74t) < 'us, it(t) < 0, u ( t ) # ml(v(t) - c l ) and u ( t ) # mb?J(t) cbr or if v 4 < ~ ( t<) 713. v ( t ) < U ( t ) = m,b'!J(t) -t cb and mt < mb or if 114 < ~ ( t<) v g , v ( t ) > 0, u ( t )= mtw(t) et and mt < mb if v(t) 5 114, or if 714 < ~ ( t<) us, li(t) > 0, ~ ( t#) m,(v(t) - e,) and u ( t )# m t 4 t ) et, or if 114 < ~ ( t <)113, it(t) > 0, u ( t ) = m t v ( t ) et and m t > mb or if 114 < ~ ( t <)v3, i ( t ) < 0, u ( t ) = mbv(t) cb and m t > rrib if 114 < u ( t ) < 113, ,ii(t) > 0 and u ( t ) = m,(v(t) - e,) if w4 < w ( t ) < 113: w(t) < 0 and u,(t) = ml(v(t) - c l ) if ii(t) = 0.

+ + +

(2.9)

(2.6)

where c d ( t ) E ( e t , c l ) ?c u ( t ) E (ea, cb) are piecewise constant functions which depend on the point where v ( t ) changes its sign and on the past trajectories of (v(t), u ( t ) ) ,with
c1 =

b -

mt) cb+m'c

('VLb - Vit) %Tli21$2cbr

+ cb

Cb

formt < mb formt > 7111, for mt = rrib

(2.7)

c2

c 1 +m,c, (mt - mb) m,--mt et, for mt > 7nb = (mt - mb) ~~~~c~ et formt < mb {et for mt = m,b.

+ +

(2.8)

The relationship (2.6) holds also for a part of one of the half-lines: when mt > mb, on the half-line (2.2) with

The model of the hysteresis and its two typical minor loops are shown in Fig. 1. This model captures most of the hysteresis phenomena described by more elaborate models in [5], e.g., it is an extended approximation of the ferromagnetic hysteresis model with ml = m, = CO, m,t = mb = 0 [6]. When 7nt = mb, the major hysteresis is the out-loop hysteresis, and the inner loops degenerate into line segments. When mt = mb, 712, = nil, et = -cb > 0, c, = -cl > 0, the hysteresis model (2.2)-(2.8) becomes the approximation model of [6] for the ferromagnetic hysteresis. The assumptions about the hysteresis characteristic H ( . ) are 1) the hysteresis output u ( t ) is not available for measurement; 2) mt > 0, 7rib > 0, m, > max{mt, mb}, ml > max{mt. mb}, et 2 C b , e, 2 cl, and the values of these hysteresis parameters are unknown. The assumptions for the linear part G(s) of the plant are 3) G ( s ) is minimum phase; 4) the relative degree n* of G(s) is known; 5) the degree n of P ( s ) is known; 6) the sign of k, is known. The control objective is to design a feedback control ~ ( t so ) that all closed-loop signals are bounded and the plant output

202

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 40, NO. 2, FEBRUARY 1995

y ( t ) tracks the output yvl(t) of the reference model:

Ym (t> =w m (s>[I. (t>

(2.10)

where W m ( s )is a stable rational transfer function of relative degree n*,and r ( t ) is a bounded piecewise continuous signal. Without loss of generality, we take W m ( s )= P&'(s)]where Pm(s) is a monic Hurwitz polynomial of degree T I * . A linear model reference controller structure could be used to achieve the stated control objective of the hysteresis were absent, that is, u ( t ) = v ( t ) . In the presence of hysteresis, a linear controller alone cannot achieve the control objective. We propose to use an adaptive hysteresis inverse developed in the next sections to cancel the effect of hysteresis. 111. HYSTERESIS INVERSE Let u d ( t ) be a control signal to be designed. As the inverse of the hysteresis ( 2 . 2 ) ( 2 . 6 ) , we use the hysteresis-like characteristic H I ( . ) :
Fig. 2. Hysteresis inverse.

are allowed to have both positive and negative signs on other parts of these two half-lines. The model of the hysteresis inverse and its two typical minor loops are shown in Fig. 2. The motion of u d ( t ) and ~ ( t on ) the half-lines (3.2)-(3.3) and the segments (3.4)-(3.5) and inside the loop is mathematically described as

and the line segments:

(3.4)

(3.5)
Along the segments, the time derivatives of v ( t ) , ~ ( tare ) of constant sign, namely, &(t) > 0, .(t) > 0 for v ( t ) = ( l / m , ) u d ( t ) e,; and i L d ( t ) < 0, ;(t) < 0 for v ( t ) =
(l/ml)ud(t)

U(t)

+ Cl.

Inside the loop formed by the half-lines (3.2)-(3.3) and the segments (3.4)-(3.3, the output of the hysteresis inverse v ( t ) is defined as

where & ( t ) E ( c t , c l ) ]G ( t )E (cp, cb) are piecewise constant functions which depend on the point where u d ( t ) changes its sign and on the past trajectories of ( v ( t ) , ~ d ( t ) ) with , c1 and cp defined in (2.7) and (2.8). The relationship (3.6)holds also for a part of one of the half-lines of the hysteresis inverse H I ( . ) : when mt > m b , on the common part of the half-line (3.2)and the hysteresis loop, v(t) = ( l / m t ) ( u d ( t ) e t ) for &(t) < 0; when mt < mb, on the common part of the half-line (3.3)and the hysteresis loop, v ( t ) = ( l / m b ) ( U d ( t ) - c b ) for & ( t ) > 0. i L d ( t ) and G ( t )

The proposed hysteresis inverse has the following properties: Proposition 3.1 (Hysteresis Inverse): The characteristic H I ( . ) defined by (3.2)-(3.6) is the right inverse of the characteristic H ( .) defined by (2.2)-(2.6) in the sense that

H(HI(ud(t0))) = ,Ud(tO)

* H ( H I ( U d ( t ) ) )= W ( t ) ,
Vt

2 t o (3.8)

for any piecewise continuous ud(t) and any l o 2 0.

TA0 AND

KOKOTOVId: PLANTS WITH

UNKNOWN HYSTERESES

203

To ensure that (3.2)-(3.6) are implementable, we assume 7) positive constants m t l , mt2, mbl, mb2. m,o, mlo and constants C b l , C b t , et2, ell, cl,, cr2 are known such that m t l I rlt I ~ 2 mbi , L n L b L m b 2 . max(~nt2, mb2> < w1ro I m r , max{mtp. m b 2 ) < 71110 2 ml. and C61 I C6 I Cbt I et I ~ t 2 , c/1 I C / 2 clr I c r I C r 2 . Even when not stated explicitly, we will use projection to ensure that the estimated hysteresis parameters satisfy the above inequalities. A numerical scheme for implementing the estimated hysteresis inverse (3.8) is developed next. At time t = t k . we have the knowledge of S t ( t ) ,G ( t ) , 7 r b ( t & ) , (t). < ( t ) . 7x(t). 6 ( f ) .Ud(t). u ( t ) for f = t k - 1 and the knowledge of u d ( t ) for f = t k . With the help of Fig. 3 where S = ( u d ( t k - 1 ) . ?i(fk-1)). we introduce the quantities 5 ud/ 5 5 udt:

7rr(f).

Fig. 3 .

Adaptive hysteresis inverse.

The mapping (3.2)-(3.6) may fail to define a hysteresis inverse only if u d ( t ) is such that v ( t ) and u ( t ) never leave the hysteresis loop, that is, t o in (3.8) cannot be reached. Such a situation happens when ud(t)is inside the hysteresis inverse loop and u ( t ) # ud(t) initially, and the motion of ud(t) is such that u ( t ) never leaves the hysteresis loop to correct the error between u d ( t ) and u ( t ) .However, as Ud(t) is the design signal at our disposal, an initialization of the hysteresis inverse by an appropriate choice of u d ( t 0 ) can always make v ( t ) and u ( t ) leave the inside of the hysteresis loop at t o so that u(to) = ud(to), and then from Proposition 3.1, ~ ( t= )ud(t) for any t > to. In the adaptive control problem, the hysteresis parameters are unknown, so the exact hysteresis inverse (3.2)-(3.6) is not implementable. We propose to use an adaptive hysteresis inverse whose parameters are updated according to adaptive laws. & be the estimates of the Let s t , G , e b . 6. K,. unknown hysteresis parameters 7nt.2t, mb, Cb, m,,c,.. ml, cl. An estimated hysteresis inverse H I ( . ) is defined as

c..rx,

Fig. 3 shows its characteristic with two typical minor loops, as well as two segments indicating two possible motion of ' U ( t ) and u d ( t ) at point 6 ( t )E &(')), & ( t ) E ( G ( t ) ,c^z(t)) depend On the point ('',d(t). where & ( t ) changes the sign, and on the past trajectories of the adaptive hysteresis inverse, where G ( t ) and & ( t )are calculated from (2.5) and (2.6), but with the estimates of the hysteresis parameters. Gb, For a fixed hysteresis inverse, the estimates Cb. m,, c,.. ml, 6 are constants based on certain a priori knowledge of m t , et, mb, C6. m,, cr. nLl, CI.For an adaptive Z t ( t ) = , G ( f ) . e b = %((t), hysteresis inverse, Kt = r

''

Nowt

(c^t(t)l

The implementation is dynamic because it uses not only but also ud(tk-1) and u ( t k - 1 ) . It is important that this implementation does not need the knowledge of ?id( t ) . In our adaptive control schemes, the adaptive hysteresis inverse will be used in cascade with the plant with a hysteresis, ) be generated from a linear controller and the signal l L d ( t will structure.
ud(tk).

IV. PARAMETERIZATION

In the absence of the hysteresis, a linear controller generating u ( t ) can be used to achieve the asymptotic tracking of ! j m ( t ) by y(t). When the hysteresis is present, v ( t ) is the accessible control which, by our design, is the output of the adaptive 6 = 6 ( t )e, r = K r ( t ) ,c^, = G ( t ) .6 = r Z l ( t 6 ) , = c ( t ) hysteresis inverse with input u d ( t ) . A parameterization of the are time-varying signals generated by an adaptive update law. control error u ( f ) - ud(t) will help us to develop suitable
A h

et. c3

204

lEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 40, NO. 2, FEBRUARY 1995

controller structures for generating Ud(t) and adaptive laws S b ( t ) , G ( t ) ,S , ( t ) , for updating the estimates s t ( t ) ,% ( t ) , c^,(t) ,6 (t) , 6(t) to implement an adaptive hysteresis inverse. To proceed, we define the following indicator functions:

Z(t) =
0

G ( t ) < G(t)< G ( t ) }
otherwise.

xt(t)=

1 if(v(t), u(t>> E {(v(t), = mtv(t) C t } 0 otherwise

44) I4t)
(4.1)

(4.12) In defining these indicator functions, we do not repeatedly count any intersection of thekalf-lines and the segments, e.g., at the top right corner of H I ( . ) we define = 1 and Z(t)= 0, and when it goes down passing the G ( t )point, we define %(t) = 1 and = 0. With this constraint, only one of these functions is nonzero at any given time t

ct(t)

z(t)

All the equalities in (4.1)-(4.12) are understood in the sense of (2.2)-(2.5) and the adaptive version of (3.2)-(3.5), that is, inequalities apply whenever they are feasible. We note that the condition 7) is necessary to ensure a well-defined adaptive hysteresis inverse with the properties (4.13) and (4.14). Using (2.2>-(2.6), (4.1)-(4.6), (4.13), we express u ( t ) as

(4.7)

where

Since a projection based on the condition 7) ensures that all estimated hysteresis parameters are bounded, the adaptive hysteresis inverse loop is bounded. If v ( t ) is large such that ( U @ ) , ud(t)) is outside the hysteresis inverse loop and ( u ( t ) v(t)) , is outside the hysteresis loop, then all ~ ( t ) ' and s 2(t)'sare zero except for x , ( t ) (or x l ( t ) )and z(t) (or fl(t)), so that d l ( t ) = 0 from (4.16). This implies that d l ( t ) is bounded whenever the estimated hysteresis parameters satisfy 7), which is crucial for our designs.

205

Proposition 4.1: The unparameterizable part d h ( t ) of the control error u ( t ) - u d ( t ) is bounded for any t 2 0, and reduces to zero when the hysteresis parameter error $ h ( t ) is zero. The signals X t ( t ) . X b ( t ) , X r ( t ) , Xi@), X d ( t ) , and x u ( t ) . which describe the motion of the hysteresis output u ( t ) = H ( w ( t ) )[see the first expression for u ( t ) in (4.15)], are not available for measurement. We choose to parameterize the control error u ( t )- ud(t) in terms of the measured signals Ft(t),% ( t ) ,g(t), E(t), and [see the definition (4.26)for u h ( t ) ] which describe the motion of the adaptive hysteresis inverse output v ( t ) = H^I(ud(t)) [see (4.17)]. The parameterization (4.27)can be made simpler for some special hysteresis characteristics. (4.17) Hysteresis with Equal Slopes mt = mb: The knowledge that mt = m b can be used to choose %b(t) = 7 T t ( t ) , & ( t ) = & ( t ) . In this case, we define

ci(t),

z(t)

(4.18) (4.19) (4.20) (4.21) (4.22) (4.23)

6'; = ( 7 n t , et.

Cb,

m,, m,c,,

7111,

ml~1)~

(4.29)

Wh(t)

where In both cases, (4.27) and Proposition 4.1 hold.

z(t)

(-(x^t(t) + Z(t) + % ( t ) ) v ( t ) . -(X^t(t)- % ( t ) ) , - (%t) + 21(t))v(t), % t ) - 2i(t))T (4.35) is also defined by (4.32), but with & ( t )= - G ( t ) .

v.

ADAPTIVE HYSTERESIS INVERSE FOR G(s) KNOWN

In this section, we design an adaptive control scheme for the plant (2.1) with a known linear part.
We use the following linear controller to generate u d ( t ) : where
~ d ( t= )

OfTwl(t)

+ t 9 ; T ~ 2 ( t )+ O&w(t) + e ; ~ ( t ) . (5.1)
e;,

Thus, we have expressed the control error u ( t )- u d ( t ) as a sum of a parameterizable part and an unknown disturbance. The disturbance d h ( t ) has the following properties.

There are two designs for ST, e;,, and 0; and w l ( t ) and ~ 2 ( t ) . I) For a(.) = (1, s,...,s"-~)~, A(.) being any monic Hurwitz polynomial of degree n - 1

206

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 40, NO. 2, FEBRUARY 1995

0; = k;', and e;, 0; E Rn-l, O.$o E R satisfying the following Diophantine equation

BTTa(s)P(s) (OITa(s) B;oA(s))kpZ(s) = A ( s ) ( P ( s) k p 6 ' ; Z ( ~ ) P m ( ~ )(5.3) ).


= (1, s , . . . , s ~)- ~ , A(s) being any monic Hurwitz polynomial of degree n, w l ( t ) and ~ 2 ( t still ) have the form (5.2), and 0; = k ; ' , O ! , 0%E R", = 0 satisfying (5.3). Then the signal ud(t) is applied to the adaptive hysteresis inverse (3.8) to generate v ( t ) as the control input to the plant (2.1) which has a hysteresis at input. To implement an adaptive hysteresis inverse, we can use the adaptive version of the continuous-time hysteresis inverse (3.7) or the discrete-time version (3.13). For the controller 11, both the continuous-time and discrete-time adaptive hysteresis inverses can be implemented because, from (5.1) with eao = 0, the derivative of u d ( t ) ,i L d ( t ) , which is needed for implementing (3.7), is also available (assuming that + ( t )is bounded). For the controller 11, one can only use the discretetime inverse (3.13) since, in this case, & ( t ) depends on y ( t ) , which is not measured. For the discrete-time inverse, there is no need to assume the boundedness of i.(t ) . The parameter matching equation (5.3) can be used to derive

11) For U ( . )

e;,

Proof: With the substitution of (5.5) and (5.6) in (5.7), the estimation error th ( t ) becomes

+GOY(t)

+ s;w,-'(-S)[Yl(t). (5.4)

Using (2.9), (4.27), (5.1), (5.4), introducing e ( t ) = y ( t ) ?Im@), F ( s ) = 0~-'Wm(s)(l- OTT(a(s)/A(s))), we obtain the tracking error equation

e(t) = F(s)[4;Whl(t)

+ F(S)[dh](t).

(5.5)

We note that F ( s ) is a stable, strictly proper, and known transfer function. Introducing
Ih(t)= m t ) C h ( t ) - F ( 4 [ f l , T W h l ( t L

for some constant

k0

> 0. Then, using

(5.13) and

Ch(t) = F ( ~ ) [ w h ] ( t (5.6) )

we use the following adaptive law to update

Bh(t)

where r h = r;f > 0, and c is a "switching-sigma'' signal [lo] using a priori knowledge of an upper bound M h on the Euclidean norm 110;111 of 0; and a design parameter CO > 0:

This adaptive law has the following properties.

V we have part (2) of the lemma. Lemma 5.1 only shows the boundedness of B h ( t ) and the L2 properties (5.10)-(5.11) of the adaptive law (5.8). While the signal boundedness of the adaptive control system with a general hysteresis is still under investigation, we present the stability results for some special cases. Theorem 5.1: When the hysteresis H ( . ) in (2.1) has two equal slopes, mt = m b and (4.30), (4.31) are used for adaptation and implementation of the adaptive hysteresis inverse, all closed-loop signals are bounded. The proof for Theorem 5.1 is given in Appendix A. This boundedness result holds for the symmetric hysteresis case where (4.34), (4.35) are used for adaptation and implementation of the adaptive hysteresis inverse with a reduced system order.

TA0 AND KOKOTOVIC: PLANTS WITH UNKNOWN HYSTERESES

207

VI. ADAPTIVE HYSTERESIS INVERSE FOR G(s) UNKNOWN


The problem of adaptively controlling the plant (2.1)with an unknown hysteresis H ( . ) as well as an unknown linear part G(s ) challenges us with the task of estimating two sets of parameters: one from the hysteresis inverse, and the other from a linear controller structure. In this case, the linear controller structure (5.1) results in a nonlinear parameterization which is not suitable for parameter estimation. Facing this difficulty, we need to modify the linear structure (5.1). Using (4.13), (4.18)-(4.23),(4.25),we express the hysteresis inverse (3.2)-(3.6)as

w ( t ) = ( w ? ( t ) , y(t)l ~ ( t uz(t), ), wc(t),~ : ( t ) ) ~ (6.10) .

The expression of the tracking error allows us to use adaptive linear control theory [ 1 1 J, [ 1 2 1 to develop an adaptive law for updating parameters of the adaptive hysteresis inverse and adaptive linear controller structure which make up our adaptive controller for the nonlinear plant (2.1). We note that OHo, O;L, &o(t). 1 9 h ( t ) , ~ ( tare ) defined in different ways when the different designs of Sections IV and V are used; see (4.24)-(4.35), and (5.1). Starting with tracking error equation (6.7), we can design adaptive schemes to update the parameter vector O(t). Next, Ud(t) -4;l'wh(t) G ( t ) G ( t ) x';l(t)G(t). (6.1) we present two such schemes. Adaptive Scheme I : Letting p ( t ) be the estimate of p* and Hence, the term dT'wl(t) in (5.1) becomes L ( s ) be any Hurwitz polynomial of degree n* - 1 such that Wm(s)L(s) is strictly positive real, and introducing

< ( t )= L-l(4[w](t), [ ( t )= O'(t)<(t) - L-'(s)[O'w](t)


Introducing new regressors and parameters as
wq(t) = 4 - - [s W) h](t).

(6.11)

we define the estimation error ~ ( tfrom )


f(t)= e(t)

A($)

e: = -e;

e;;

+ Wm(s)L(s)[p
-

at(('<

+ ')](t)

(6.12)

(6.3)

where

(1

> 0, and update e ( t ) and

p ( t ) from

a(.) wg(t) = - [ X &

Ns)

+ G;c^,](t). e; = e;

O(t) = -sign[k,]I'<(t)t(t)
(6.4)
i,(t) =

I'a(0, M O ,oo)O(t) (6.13)


4 Mp, CJo)p(t)

where

denotes the Kronecker product, A ( s ) = (Ipl SI,... . . sn-'?Ip)'. I, is the p x p identity matrix with p being the dimension of O;Ll from (6.2),we obtain
@

-r(t)dt) -7

(6.14)

e;Tw,(t) = e;Tw,(t)

+ egTw:,(t).

(6.5)

We should note that the dimension of 0: and its estimate is p ( n - 1) ( p n ) for the controller I (controller 11) of Section V. It is this enlarged parameterization that enables us to obtain a linear parameterization of the error system, which is suitable for controller adaptation. Physically, the hysteresis has eight different regions, so that the feedforward part Q;'(t)w, ( t ) has eight different forms. A natural parameterization tums out to be the one given in (6.5), with the parameter vector 6;. This new parameterization brings us to a new controller structure for adaptive control

where r = I" > 0. y > 0, and CJ is the switching signal defined by (5.9). We then have the following properties of this adaptive law. Lemma 6.1: The adaptive law (6.13)-(6.14)guarantees 1) O(t), t(t) E Lm; 2) for some constants k3 > 0, kq > 0, and all t 2 > tl 2 0,

~ ( t= ) e T ( t ) w 2 ( t )+ e 2 n ( t ) p ( t ) + @ 3 ( t ) ~ ( t ) O,T(t)w&) Q,T(t)w&). (6.6)

&o(t) = 0 in (6.6)if a ( s ) , A(s) are defined by the controller I1 of Section V. In view of (4.27),(5.4), (6.1)-(6.6),the tracking error
equation now has the form

Substituting (6.12) in (6.17), we have

.(t)

=p*w"4?)T4(t)

+4t)
d(t)

~ ( t= ) W, (.)L(.)[p*4?)TC+11/-(f((T~+z)+~](t) (6.18)
(6.7)
where &t) = p * ( 1
= p*W,(s)
-

where p* = Oi-', 4?)(t) = 8 ( t ) (1 - e;T(a(s)/A(s)))[dhl(t),and

+ B*.

eI'(u(~)/A(.)))L-~(s)[dh](t), $ ( t ) =

P ( t ) - P*.

O(t) = (O,T(t),020(t), O d t ) , Q,T(t),

m, eWT
0;;')'

Let W , (s)L( s ) have a controllable realization (A,, B,, C,) and e t @ ) be its state variable. With this notation, from (6.8) (6.18), we obtain

e* = ( e ; T , e;(), e;. e;', e;'.

(6.9)

& ( t ) = A,.e,(t)

+ B,v(t).

t(t) = C,e,(t)

(6.19)

208

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 40, NO. 2, FEBRUARY 1995

where
4 t ) = P*4T(t)C(t) $ ( t ) E ( t )
-

at(t)(cT(t)<(t)('(t)) g(t).

(6.20)

L( s) implies that there Strict positive realness of W , ( s ) exist constant matrices Q = QT > 0 , P = P T > 0 , vector q , and scalar S > 0 such that A Z P P A , = -qqT SQ, P B , = C;f. Hence, the time derivative of V(P,,4, 4)= eTPe, jp*14Tr-1$ y-l@ is

v i - w 2 ( t ) ( 1 + CT(t)C(t) + E2(t))
+
J2(t)
1 CT(t)C(t) P ( t ) - 21P*14)4T(t)Q(t)- 2 a , ( t ) W ) p ( t )

(6.21)

for some constants a1 > 0, a2 > 0. This, together with a similar argument to (5.14) and (5.15), proves (1) and (2). D Adaptive Scheme I I : Instead of a choice of L ( s ) of degree n* - 1, we choose L ( s ) = P,(s) = W i l ( s ) , of degree n*, to obtain the estimation error as

still a research topic. However, for a wide class of hysteresis characteristics, we have the following stability results. Theorem 6.1: If the hysteresis H ( . ) in (2.1) has two equal slopes, mt = mb, and (4.30), (4.31), (6.9), (6.10) are used for adaptation and implementation of the adaptive hysteresis inverse and the adaptive controller, then all closed-loop signals are bounded. The proof for Theorem 6.1 is given in Appendix B. This boundedness result applies to the symmetric hysteresis case where (4.34), (4.33, (6.9), and (6.10) are used for adaptation and implementation of the adaptive hysteresis inverse and the adaptive controller (6.6), with a reduced system order. While analytically characterizing the tracking performance of the closed-loop system with an adaptive hysteresis inverse is an important future research topic, we will examine it by simulations and compare it with that of a control system without hysteresis inverse.

VII. SIMULATION RESULTS

In this section, we present an example with simulation results to illustrate the structure and effectiveness of the developed adaptive hysteresis control schemes. We consider the plant with an unstable linear part G ( s ) = where C(t) and [ ( t ) are defined in (6.11) with this new L ( s ) . 2.5/(s2 + 2s - 4) and a symmetric hysteresis characteristic We then use (6.13) and (6.14) with the new t ( t ) defined in H ( m t , ct, mb, C b , vi,, c,, m i , q ;.) where mt = m b = 1.8, (6.22) to update the estimates B(t) and p ( t ) . m, = rri1 = 3, et = -cb = 1.9, c, = -c1 = 2.5, and choose For this scheme, we have the following result. = 1/(s2 3s 2 ) . a model transfer function W m ( s ) Lemma 6.2: The adaptive (6.13), (6.14), (6.23) guarantees We study the system performance in one of the five cases: 1) O(t), d t ) ,d t ) , e 2 ( t ) ( l + 4 C T ( t ) C ( t )+ E 2 ( t ) ) )E a) a fixed linear controller without hysteresis inverse, for G(s) known; b) an adaptive linear controller without hysteresis L, 1 2) for some constant ks > 0, k6 > 0, and all t 2 2 tl 2 0, inverse, for G(s) unknown; c) a fixed linear controller with a fixed inaccurate hysteresis inverse, for G(s ) known; d) a fixed linear controller with an adaptive hysteresis inverse, for G(s) known; and e) an adaptive linear controller with an adaptive hysteresis inverse, for G(s) unknown. The controller a) has the structure

+ +

m,

Proof: With (6.7), (6.11), and (6.22), we write the time derivative of the positive-definite function V ( 4 ,4)= i(\p*lq5Tr-14 ~ - l $ ~where ) $ ( t ) = p ( t ) - p * , along (6.13) and (6.14), as

where 0; = -1, 19; = -0.4, O;, = -2.8, 6'; = 0.4, which are calculated from (5.3) with G(s) known. The controller b) is the adaptive version of a) for G ( s ) unknown:

V(t)I

1 --m(l 2 +

M(t)C(t)

+ E2W)
where O l ( t ) . & ( t ) , & o ( t ) , & ( t ) are updated from an adaptive law similar to (6.11)-(6.14) with a different O(t) = ( B l ( t ) , B z ( t ) , O , O ( t ) , &(t))*, and with O(0) = (-1.5. -0.8, -2, 0.8)T, p ( 0 ) = 1.25, and the choice of = 1. oO = 0.15, r = 101, = io. M~ = qS) = + 2, 5.53, M p = 3. These two controllers, well known in the literature, ignore the existence of the hysteresis characteristic in the plant and have no compensation for the hysteresis H ( . ) .

d2 ( t ) + 2(1 + a(CT(t)C(t) + E2(t)))


- Ip*Ig(e, Me, ~ o ) 4 * ( t M t ) - 4 P , M p , UO)lCl(t)P(t).

(6.25)

From this inequality, (6.13) and (6.14) with defined in (5.9), v (l), and (2) follow. Similar to Lemma 5.1, Lemmas 6.1 and 6.2 only show the boundedness of O ( t ) and the L2 properties (6.23)-(6.24) of the adaptive law (6.13)-(6.14). The signal boundedness of the adaptive control system with a general hysteresis is

TA0 AND KOKOTOVIC: PLANTS WITH UNKNOWN HYSTERESES

209

The controller c) is

1"

where H I ( . ) is an accurate estimate of H I ( . ) defined in (3.2)-(3.6), with Et = ?%b = 2.3, Et = -zb = 0.9, ?%, = E l = 2.5, E, = -El = 0.5, and dl, 0; are the same as that in the controller a). This controller, with H I ( . ) , has

e:,

-101

% ( t ) , qt),

f%(t),G ( t ) ;.)

5-

(7.5)

estimate of H I ( . ) , with initial parameter estimates m^t(0) = m^b(b(o) = 2.3, c(0) = -&(o) = 0.9, Z r ( 0 ) = @(U) = 2.5, G(0) = -G(O) = 0.5. The adaptive law for e h ( t ) = ( S t ( t )& , ( t ) ,Zr(t), m-r(t))T is (5.8) with CO = 0.15, rh = 101, iwh = 11. The controller (e) is with an adaptive hysteresis inverse for G ( s ) unknown
v ( t ) = H^l(Ud(t)),

10

E(.) = Hl(m^t(t),q t ) , m^b(t), q t ) ,


f%r(t),ChT(t)i m2(t)i G ( t ) :.) (7.6)
1 s+3

10

I
5

1 0

15

20

25

30

35

40

45

50

(b)

W ( t ) = ez(t)-[Yl(t)

+ &O(t)Y(t) + Qs(t)r(t)
04z(t),
1
e43(t), 0 4 4 ( t ) ) = [ W h I ( t )

Fig. 5. System responses: adaptive controller without hysteresis inverse for G(.s) unknown.

(04i(t),

+ 05(t)W5(t)
(e,@), e,@),

(7.7)

where E(.) is an adaptive hysteresis inverse with the same initial parameter estimates as that in the controller d), and the 0,0(t), 0 z ( t ) ,Os@), 0; adaptive law for 0 ( t ) = ( t ) ) Tis (6.1 1)-(6.14) with the choice of L ( s ) = s + 2 , 11 = 1, oo = 0.15, r = 101, = io, iw0 = 17.38, M~ = 3, p ( 0 ) = 1.25, and &(0) = -0.8, &0(0) = -2, &(O) = 0.8, 04(0) = (3.45, 3.75, 1.35,0.75)T, &(O) = -1.5. The control law (7.7) was developed in (6.6). Our simulations indicate 1) if either a fixed or adaptive linear controller without a hysteresis inverse is applied to this plant, then the tracking error remains significant for large t ; 2 ) an inaccurate fixed hysteresis inverse can reduce the tracking error; 3) significant improvements of system tracking performance are achieved with the use of an adaptive hysteresis inverse either with a fixed linear controller, for G ( s ) known, or with an adaptive linear controller, for G ( s ) unknown.
~~

Some typical system responses for ~ ( t=) 12.7sin(2.3t) are shown in Figs. 4-8 for the above five controllers, respectively. It is clear that with an adaptive hysteresis inverse, the tracking error is significantly reduced with less control effort (the control signal is smoother and smaller). VIII. CONCLUSIONS Our construction of an adaptive hysteresis inverse, reparameterization of an adaptive linear controller, and choice of suitable error models have led to several adaptive control schemes applicable to plants with unknown hystereses. Simulation results indicate that these adaptive inverse schemes promise to significantly improve system performance.
APPENDIX A PROOF OF THEOREM 5.1
E From Lemma 5.1 and the condition 7), we have that L , and S t ( t ) 2 mtl > 0. Introducing m = mt = mb, h(t)= Kt(t) = m ^ b ( t ) , and using u ( t ) = H ( v ( t ) ) ,w ( t ) =

210

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 40, NO. 2, FEBRUARY 1995

'th
55

i
i
5 IO
15 20 25

30

35

40

45

50

10

10

15

20

25

30

35

40

45

50

IO

I5

20

25

30

35

40

45

SO

(b)
Fig. 6. System responses: fixed controller with inaccurate hysteresis inverse for G(.s) known.

(b)

Fig. 8. System responses: adaptive controller with adaptive hysteresis inverse for G ( s ) unknown.

Using (A.l) and (A.3) in (5.1), we have

u ( t )= -qT--(S m(t)
0
5

u(s) k(.)

A(s) m

+ ao)[zo](t)+

IO

15

20

25

30

35

40

45

SO

(a)

1 - -8p--) m a(.) r j l ( . ) m(t) h ( s ) m


0

[d2](t). (A.6)

Substituting this expression for u ( t )in (AS), we see that z ( t ) and zo(t) are related by
I
0
5
10

15

20

25

30

35

40

45

50

(b)
Fig. 7. System responses: fixed controller with adaptive hysteresis inverse for G(.\) known.

u ( t ) = mw(t)+ & ( t )

('4.2)

for some bounded d2(t), & ( t ) . We then define fictitious signals zo(t), z l ( t ) , filters K l ( s ) ,K ( s ) as

Consider a linear operator T ( s , t ) with input z ( t ) .We define T ( s , t ) as a stable and proper operator if llT(s,.)[x](t)ll 5 p1 s,' , - 4 - T ) llrc(7)l)d7+pzllz(t)Il for some constants /31 2 0, /32 2 0 , and a > 0, all t 2 0, and any z ( t ) .The operator T ( s , t ) is stable and strictly proper if IIT(s,.)[z](t)lJ 5 p 1 e-"(t-')llz(.r)lldT for some constants 2 0 and a > 0, all t 2 . 0 , and any z ( t ) . With this definition, the facts that f i ( t ) , r^n(t) E L,, K l ( s ) is strictly stable, and

si

and use (A.3), (A.4), and (2.1) to obtain

zo(t)

+ aoK1(s)[zoI(t) - K1(.9)[uI(t)
= K ( s ) G - l ( s ) [ z ] ( t ) .(A.5)

TAO AND KOKOTOVIC: PLANTS WITH UNKNOWN HYSTERESES

21 1

are bounded. On the other hand, if <:(t)<h(t) Ez(t) grows unbounded, then the smallness of zl(t) in the sense of (A.16) results in the boundedness of ~ ( tgiven ) by (A.15). This, in turn, implies that zo(t) in (A.lO) is bounded, so that u ( t ) in ) (4.13), & ( t ) in (5.6), and y(t) imply that (m/7iz(t))OIT( a( s)/A( s)) (&(t)/.nz)( s ao) and (A.6), ud(t) in (A.l), ~ h ( t in in (5.7) are bounded. Hence, we conclude that all closed-loop Kl(s)(m/7?~(.))S'~(s ao) are stable and proper. By definition, the impulse response function IC1 ( t ) of K 1 (s) signals are bounded. IICl(t)ldt = n * / a . Hence, there exists u0 > 0 satisfies such that for any finite a > a', the operator APPENDIX B PROOF OF THEOREM 6.1 TO(S, t ) (1 +K1(.5)(uo- rjL0 711

A. Proof for Adaptive Scheme I


Using (A.l), (A.3), (A.11). (A.12), and (6.3), we express is stable and proper. For a fixed a

> a',

(A.7) implies that (A. 10)

O?(t)W4(t)

= O,T(t)-(s 4 s )

+ ao)[zo](t) + d5(t)

(B.1)

for some bounded O,(t) E Rn-' for the controller I (O,(t) E where Tl(s,t ) is a stable and proper operator, and & ( t ) is a R" for the controller 11) of Section V, and some bounded d 5 ( t ) . Using zo(t), K l ( s ) ,K ( s ) defined in (A.3), (A.4), we bounded signal due to r ( t ) , d2(t). As in the swapping lemma [12], we let ( A f ,B f , C,) be a obtain minimal realization of F ( s ) ,define W c ( s )= C f ( s 1 - A f ) - l , Zo(t) aoK1(s)[zoI(t) - Kl(3)[U](t) Wb(s) = ( 3 1 - A f ) - l B f , and express &(t) as 1 = K (-9)G-l( s )[y] ( t ) . s a0 E h ( t ) = WAS) [ W s + a0)= rwhT1& ( t ) . ('4.11) l l Using ( A . l ) , (6.6), (B.l), we obtain Using (4.13) and (4.31), we rewrite u h ( t ) as

zo(t) = Tl(%.)[zl(t)+ d 3 ( t )

A(s)

Wh(t)

= (-4t)
-

+ ( % ( t ) + x^l(t)b(t)l -%(t), -%(t), %(t).(t), %(t). -x2(t)v(t), x^l(WT ('4.12)


From (B.2) and (B.3), it follows that

whose components, except for - v ( t ) , are all bounded. Using (A.2) and (A.3), we note

where & ( t ) is a bounded vector signal. Filtering both sides of (5.7) by l / ( s ao) and using (A.3) again, we obtain

03.4)

Similar to (A.9), we see, from (B.4) with a sufficiently large a > 0 in K ( s ) , K l ( s ) ,that

zo(t) = Tl(S,.)[Vl(t) +

(B.5)

for some constants > 0, 7 & > 0, and any t 2 2 tl 2 0. The operator T 2 ( s , t ) is stable and strictly proper, while the operator T ~ ( s t ), is stable and proper and has a nonnegative impulse response. If <h(t), ( h ( t ) are bounded, then from Lemma 5.1 and (5.7), we see that f h ( t ) , y(t) are bounded, from (A.3), (A.lO) that z ( t ) , z o ( t ) are bounded, and from (A.6), (A.8), and ( A . l ) that u ( t )and ud(t) are bounded. Thus, all closed-loop signals

for some 2 1 ' (s, t ) stable and proper, and some 2 3 ( t )bounded. Similar to (A.lO), with (Al, Bl, Cl) being a minimal re= Cl(s1 - A[)-', W b ( s ) = alization of l / L ( s ) and W c ( s ) (SI- Al)-lBl, we express ( ( t ) as
E(t) = W c ( s ) [Wb(s) [w'lel

(t).

(B.6)

Using ( A . l l ) , (6.3), (6.4), (5.2), (BS), in (6.10), we see that


w ( t ) = Tl(% . ) [ Y ] ( t )

+ 24(t)

(B.7)

for some Tl( s , t ) stable and proper, and some 2 4 ( t )bounded.

212

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 40, NO. 2, FEBRUARY 199.5

By rewriting (6.12), we get

ACKNOWLEDGMENT

Y(t) = Ym(t)

+ 4 t )- Wm(S)L(S)

. [d -

+ <T< + E 2 ) + afl(t).

(B.8)

We are thankful to J. Winkelman and D. Rhode of Ford Motor Company for stimulating this research, and to D. Recker of the University of Illinois for many helpful discussions.

Similar to (A.15), (6.1 l), (B.6)-(B.8), and Lemma 6.1 imply

REFERENCES
[ 11 L. 0. Chua and S. C. Bass, A generalized hysteresis model, IEEE Trans. Circuit Theory, vol. CT-19, pp. 36-48, Jan. 1972. [2] G. J. Thaler and M. P. Pastel, Analysis and Design of Nonlinear Feedback Control Systems. New York: McGraw-Hill. 1962. [3] M. A. Krasnoselskii and A. V. Pokrovskii, Systems wirh Hysreresis. Berlin: Springer-Verlag, 1983. [4] N. D. Vaughan and J. B. Gamble, The modelling and simulation of a proportional solenoid valve, presented at the Winter Ann. Meet. Amer. Soc. Mech. Eng., Dallas, TX, Nov. 1990. 1 5 1 1. D. Mayergoyz, Mathematical Models ofHysteresIs. Berlin: SpringerVerlag, 199 1. (61 A. Visintin, Mathematical models of hysteresis, in Topics in Nonsmooth Mechanics. J. J. Moreau, P. D. Panagiotopoulos, and G. Strang, Ed. Berlin: Birkhauser Verlag, 1988, pp. 295-326. [7] D. A. Recker, P. V. KokotoviC, D. S. Rhode, and J. R. Winkelman, Adaptive nonlinear control of systems containing a dead-zone, in Proc. 30th IEEE Conf. Decis. Contr., Brighton, England, 1991, pp. 21 11-21 15. [XI G. Tao and P. V. KokotoviC, Adaptive control of plants with unknown dead-zones, in Proc. 1992 Amer. Contr. Conf., Chicago, IL, June 1992, pp. 2710-2714. Adaptive control of systems with backlash, Automatica. vol. [9] -, 29. pp. 323-335, Mar. 1993. [ 101 P. A. Ioannou and K. S. Tsakalis, A robust direct adaptive controller, IEEE Trans. Automat. Contr., vol. AC-31, pp. 1033-1043, Nov. 1986. [ l I ] K. S. Narendra and A. M. Annaswamy, Stable Adaprive Systems. Englewood Cliffs, NJ: Prentice-Hall, 1989. [ 12) S. S. Sastry and M. Bodson, Adaptive Control: Stability, Conwrgence and Robustness. Englewood Cliffs, NJ: Prentice-Hall, 1989.

for some constants E 3 > 0, > 0, and any t 2 2 t l 2 0. The operator T ~ ( t s ), is stable and strictly proper, while the operator T ~ ( t s), is stable and proper and has a nonnegative impulse response. From here on, the closed-loop signal boundedness follows from a contradiction argument which is similar to that in the proof of Theorem 5.1.

B . Proof for Adaptive Scheme I1


Using (A.]), (6.6), (B.l), we obtain

u ( t ) = -eT(t)-(s m 4s)
h(t)

+ ao)[zo](t)
+ ao)[zl(t)
+ .o)[.](t)

A(s)

m + -O,(t)r(t) + --d5(t) rn(t) m(t)


7n

Vl + -Oif(t)-(s m(t) rn + -&o(t)(s m(t)

4s)
A(.s)

+ dz(t).

(B.ll)

From

(AS)

and (B.l1), it follows that

Gang Tao (S384-M89) received the B.S. degree from the University of Science and Technology of China in 1982, and the Ph.D. degree from the University of Southern California in 1989, both in electrical engineering. He was a Visiting Assistant Professor at the Washington State University from 1989 to 1991, and an Assistant Research Engineer and a Lecturer at the University of Califomia at Santa Barbara from 1991 to 1992. Since September 1992, he has been an Assistant Professor with the Department of Electrical Engineering, University of Virginia. His main research area is adaptive control, and currently he is working on the adaptive control of systems with nonsmooth nonlinearities, and on developing new adaptive algorithms for control and estimation.

(B.12)
where Kl(S)(7./7;1(.))e,,(.)(S because

+ U O ) is

stable and proper

and 7jL(t),f n ( t ) ,&,(t), H 2 0 ( 2 ) are all bounded. The remaining part of the proof is analogous to (A.9)-(A. 14), followed by a contradiction argument similar to that used in the proof of Theorem 5.1.

Petar V. KokotoviC (SM74-F80) has been active for more than 30 years as a control engineer. researcher, and educator, first in his native Yugoslavia, and then, from 1966 through 1990, at the University of Illinois, where he held the endowed Grainger Chair. Since 1991 he has been Co-Director (with A. J. Laub) of the newly formed Center for Control Engineering and Computation at the University of California, Santa Barbara. He has coauthored eight books and numerous articles contributing to sensitivity analysis, singular perturbation methods, and robust adaptive and nonlinear control. He is also active in industrial applications of control theory. As a consultant to Ford, he was involved in the development of the first series of automotive computer controls, and at General Electric, he participated in large-scale systems studies. Dr. KokotoviC received the 1990 Quazza Medal, the 1983 and 1993 Outstanding IEEE TRANSACTIONS Paper Awards, and presented the 1991 Bode Prize Lecture. He is the recipient of the 1995 IEEE Control Systems Award.

You might also like