Articulo Higher Order Sliding Modes, Differentiation and Output Feedback Control

Higher-order sliding modes, differentiation and output-feedback control

ARIE LEVANT{

Being a motion on a discontinuity set of a dynamic system, sliding mode is used to keep accurately a given constraint andfeatures theoretically-infinite-frequency switching. Standard sliding modes provide for finite-time convergence, precisekeeping of the constraint and robustness with respect to internal and external disturbances. Yet the relative degree of theconstraint has to be 1 and a dangerous chattering effect is possible. Higher-order sliding modes preserve or generalize themain properties of the standard sliding mode and remove the above restrictions. r-Sliding mode realization provides forup to the rth order of sliding precision with respect to the sampling interval compared with the first order of the standardsliding mode. Such controllers require higher-order real-time derivatives of the outputs to be available. The lackinginformation is achieved by means of proposed arbitrary-order robust exact differentiators with finite-time convergence.These differentiators feature optimal asymptotics with respect to input noises and can be used for numerical differentia-tion as well. The resulting controllers provide for the full output-feedback real-time control of any output variable of anuncertain dynamic system, if its relative degree is known and constant. The theoretical results are confirmed by computersimulation.

1. Introduction

Control under uncertainty condition is one of the

main topics of the modern control theory. In spite of

extensive and successful development of robust adaptive

control and backstepping technique (Landau et al. 1998,

Kokotovic and Arcak 2001) the sliding-mode control

approach stays probably the main choice when one

needs to deal with non-parametric uncertainties and

unmodelled dynamics. That approach is based on keep-

ing exactly a properly chosen constraint by means of

high-frequency control switching. It exploits the main

features of the sliding mode: its insensitivity to external

and internal disturbances, ultimate accuracy and finite-

time transient. However, the standard sliding-mode

usage is bounded by some restrictions. The constraint

being given by an equality of an output variable tozero, the standard sliding mode may be implemented

only if the relative degree of is 1. In other words,control has to appear explicitly already in the first

total derivative _. Also, high frequency control switch-ing leads to the so-called chattering effect which is exhib-

ited by high frequency vibration of the controlled plant

and can be dangerous in applications.

A number of methods were proposed to overcome

these difficulties. In particular, high-gain control with

saturation approximates the sign-function and

diminishes the chattering, while on-line estimation of

the so-called equivalent control (Utkin 1992) is used to

reduce the discontinuous-control component (Slotine

and Li 1991), the sliding-sector method (Furuta and

Pan 2000) is suitable to control disturbed linear time-

invariant systems. Yet, the sliding-mode order approach

(Levantovsky 1985, Levant 1993) seems to be the most

comprehensive, for it allows to remove all the above

restrictions, while preserving the main sliding-mode fea-

tures and improving its accuracy. Independently devel-

oped dynamical (Sira-Ramrez 1993, Rios-Bolvar et al.

1997) and terminal (Man et al. 1994) sliding modes are

closely related to this approach.

Suppose that 0 is kept by a discontinuousdynamic system. While successively differentiating along trajectories, a discontinuity will be encountered

sooner or later in the general case. Thus, sliding

modes 0 may be classified by the number r of thefirst successive total derivative r which is not a con-tinuous function of the state space variables or does not

exist due to some reason like trajectory non-uniqueness.

That number is called sliding order (Levant 1993,

Fridman and Levant 1996, Bartolini et al. 1999 a). The

standard sliding mode on which most variable structure

systems (VSS) are based is of the first order ( _ is dis-continuous). While the standard modes feature finite

time convergence, convergence to higher-order sliding

modes (HOSM) may be asymptotic as well. While the

standard sliding mode precision is proportional to the

time interval between the measurements or to the

switching delay, r-sliding mode realization may provide

for up to the rth order of sliding precision with respect

to the measurement interval (Levant 1993). Properly

used, HOSM totally removes the chattering effect.

Trivial cases of asymptotically stable HOSM are

often found in standard VSSs. For example there is an

asymptotically stable 2-sliding mode with respect to the

constraint x 0 at the origin x _xx = 0 (at one pointonly) of a two-dimensional VSS keeping the constraint

x _xx 0 in a standard 1-sliding mode. Asymptoticallystable or unstable HOSMs inevitably appear in VSSs

International Journal of Control ISSN 00207179 print/ISSN 13665820 online # 2003 Taylor & Francis Ltdhttp://www.tandf.co.uk/journals

DOI: 10.1080/0020717031000099029

INT. J. CONTROL, 2003, VOL. 76, NOS 9/10, 924941

Received 15 May 2002. Accepted 2 September 2002.{School of Mathematical Sciences, Tel-Aviv University,

Ramat-Aviv, 69978 Tel-Aviv, Israel. e-mail: [email protected]

with fast actuators (Fridman 1990, Fridman and Levant1996, 2002). Stable HOSM leads in that case to sponta-neous disappearance of the chattering effect.Asymptotically stable or unstable sliding modes of anyorder are well known (Emelyanov et al. 1986, Elmaliand Olgac 1992, Fridman and Levant 1996). Dynamicsliding modes (Sira-Ramrez 1993, Spurgeon and Lu1997) produce asymptotically stable higher-order slidingmodes and are to be specially mentioned here.

A family of finite-time convergent sliding mode con-trollers is based on so-called terminal sliding modes(Man et al. 1994, Wu et al. 1998). Having been indepen-dently developed, the first version of these controllers isclose to the so-called 2-sliding algorithm with a pre-scribed convergence law (Emelyanov et al. 1986,Levant 1993). The latter version is intended actually toprovide for arbitrary-order sliding mode with finite-timeconvergence. Unfortunately, the resulting closed-loopsystems have unbounded right-hand sides, which pre-vents the very implementation of the Filippov theory.Thus, such a mode cannot be considered as HOSM. Thecorresponding control is formally bounded along eachtransient trajectory, but takes on infinite values in anyvicinity of the steady state. In order to avoid infinitecontrol values all trajectories are to start from a pre-scribed sector of the state space. The very definitionand the existence of the solution require some specialstudy here.

Arbitrary-order sliding controllers with finite-timeconvergence were only recently demonstrated (Levant1998 b, 2001 a). The proofs of these results are for thefirst time published in the present paper. These control-lers provide for full output control of uncertain single-inputsingle-output (SISO) weakly-minimum-phasedynamic systems with a known constant relative degreer. The control influence is a discontinuous function ofthe output and its r 1 real-time-calculated successivederivatives. The controller parameters may be chosen inadvance, so that only one parameter is to be adjusted inorder to control any system with a given relative degree.No detailed mathematical model is needed. The systemsrelative degree being artificially increased, sliding con-trol of arbitrary smoothness order can be achieved,completely removing the chattering effect. Since manymechanical systems have constant relative degree due,actually, to the Newton law, the application area forthese controllers is very wide.

Any implementation of the above controllersrequires real-time robust estimation of the higher-order total output derivatives. The popular high-gainobservers (Dabroom and Khalil 1997) would destroythe exactness and finite-time-convergence features ofthe proposed controllers. The first-order robust exactdifferentiator (Levant 1998 a) can be used here, but itssuccessive application is cumbersome and not effective.

The problem is solved by recently presented arbitrary-order robust exact finite-time-convergent differentiators(Levant 1999, 2001 a,b). The proposed lth-order differ-entiator allows real-time robust exact differentiation upto the order l, provided the next l 1th input deriva-tive is bounded. Its performance is proved to be asymp-totically optimal in the presence of small Lebesgue-measurable input noises. This paper is the first regularpublication of these differentiators and of the corre-sponding proofs. Their features allow broad implemen-tation in the non-linear feedback control theory due tothe separation principle (Atassi and Khalil 2000) trivi-ality. They can be also successfully applied for numericaldifferentiation.

The rth-order sliding controller combined with ther 1th-order differentiator produce an output-feed-back universal controller for SISO processes with per-manent relative degree (Levant 2002). Having been onlyrecently obtained, the corresponding results are onlybriefly described in this paper, for the author intendsto devote a special paper to this subject. The featuresof the proposed universal controllers and differentiatorsare illustrated by computer simulation.

2. Preliminaries: higher-order sliding modes

Let us recall first that according to the definition byFilippov (1988) any discontinuous differential equation_xx vx, where x 2 Rn and v is a locally boundedmeasurable vector function, is replaced by an equivalentdifferential inclusion _xx 2 Vx. In the simplest case,when v is continuous almost everywhere, Vx is theconvex closure of the set of all possible limits of vyas y ! x, while fyg are continuity points of v. Any sol-ution of the equation is defined as an absolutely contin-uous function xt, satisfying the differential inclusionalmost everywhere.

Consider a smooth dynamic system _xx vxwith asmooth output function , and let the system be closedby some possibly-dynamical discontinuous feedback(figure 1). Then, provided that successive total time deri-vatives ; _; . . . ; r1 are continuous functions of theclosed-system state space variables, and the r-slidingpoint set

_ r1 0 1

is non-empty and consists locally of Filippov trajec-tories, the motion on set (1) is called r-sliding mode(rth-order sliding mode, Levantovsky 1985, Levant1993, Fridman and Levant 1996).

The additional condition of the Filippov velocity setV containing more than one vector may be imposed inorder to exclude some trivial cases. It is natural to callthe sliding order r strict if r is discontinuous or does

Higher-order sliding modes 925

not exist in a vicinity of the r-sliding point set, but slid-ing-mode orders are mostly considered strict by default.

Hence, r-sliding modes are determined by equalities(1) which impose an r-dimensional condition on thestate of the dynamic system. The sliding order charac-terizes the dynamics smoothness degree in some vicinityof the sliding mode.

Suppose that ; _; ; . . . ; r1 are differentiablefunctions of x and that

rank r;r _; . . . ;rr1 r 2

Equality (2) together with the requirement for the cor-responding derivatives of to be differentiable functionsof x is referred to as r-sliding regularity condition. Ifregularity condition (2) holds, then the r-sliding set isa differentiable manifold and ; _; . . . ; r1 may be sup-plemented up to new local coordinates.

Proposition 1 (Fridman and Levant 1996): Let regu-larity condition (2) be fulfilled and r-sliding manifold (1)be non-empty. Then an r-sliding mode with respect tothe constraint function exists if and only if the inter-section of the Filippov vector-set field with the tangentialspace to manifold (1) is not empty for any r-slidingpoint.

Proof: The intersection of the Filippov set of admissi-ble velocities with the tangential space to the slidingmanifold (1) induces a differential inclusion on thismanifold (figure 1). This inclusion satisfies all the Fi-lippov conditions for solution existence. Thereforemanifold (1) is an integral one. &

A sliding mode is called stable if the correspondingintegral sliding set is stable. The above definitions areeasily extended to include non-autonomous differential

equations by introduction of the fictitious equation_tt 1. All the considerations are literally translated tothe case of the closed-loop controlled system_xx f t; x; u; u Ut; x with discontinuous U andsmooth f ; .

Real sliding: Up to this moment only ideal slidingmodes were considered which keep 0. In reality,however, switching imperfections being present, idealsliding cannot be attained. The simplest switchingimperfection is discrete switching caused by discretemeasurements. It was proved (Levant 1993) that thebest possible sliding accuracy attainable with discreteswitching in r is given by the relation jj r,where > 0 is the minimal switching time interval.Moreover, the relations jkj rk; k 0; 1; . . . ; r aresatisfied at the same time 0 . Thus, in order toachieve the rth order of sliding precision in discrete rea-lization, the sliding mode order in the continuous-timeVSS has to be at least r. The standard sliding modesprovide for the first-order real sliding only. The secondorder of real sliding was achieved by discrete switchingmodifications of the 2-sliding algorithms (Levant 1993,Bartolini et al. 1998) and by a special discrete switchingalgorithm (Su et al. 1994). Real sliding of higher ordersis demonstrated in Fridman and Levant (1998) andLevant (1998 b, 2001 a).

In practice the final sliding accuracy is alwaysachieved in finite time. When asymptotically stablemodes are considered, however, it is not observable atany fixed moment, for the convergence time tends toinfinity with the rise in accuracy. In the known casesthe limit accuracy of the asymptotically stable modescan be shown to be of the first order only (Slotine andLi 1991). The above-mentioned highest precision isprobably obtained only with finite-time-convergencesliding modes.

3. The problem statement and its solutions with lowrelative degrees

Consider a dynamic system of the form

_xx at; x bt; xu; t; x 3

where x 2 Rn, u 2 R, a, b and are smooth unknownfunctions, the dimension n is also unavailable. The rela-tive degree r of the system is assumed to be constant andknown. The task is to fulfil the constraint t; x 0 infinite time and to keep it exactly by some feedback.

Extend system (3) by introduction of a fictitious vari-able xn1 t; _xxn1 1. Denote ae a; 1t; be b; 0t,where the last component corresponds to xn1. Theequality of the relative degree to r means that the Liederivatives Lbe, LbeLae; . . . ;LbeL

r2ae

equal zero identi-

926 A. Levant

Figure 1. 2-sliding mode.

cally in a vicinity of a given point and LbeLr1ae

is notzero at the point (Isidori 1989).

In a simplified way the equality of the relative degreeto r means that u first appears explicitly only in the rthtotal derivative of . In that case regularity condition (2)is satisfied (Isidori 1989) and @=@ur 6 0 at the givenpoint. The output satisfies an equation of the form

r ht; x gt; xu 4

It is easy to check that g LbeLr1ae

@=@ur;h Lrae. Obviously, h is the rth total time derivativeof calculated with u 0. In other words, unknownfunctions h and g may be defined using only inputout-put relations. The heavy uncertainty of the problem pre-vents immediate reduction of (3) to any standard formby means of standard approaches based on the knowl-edge of a, b and . Nevertheless, the very existence ofstandard form (4) is important here.

Proposition 2 (Fridman and Levant 1996): Let system(3) have relative degree r with respect to the outputfunction at some r-sliding point t0; x0. Let, also,u Ut; x, the discontinuous function U taking onvalues from sets K ;1 and 1;K on some sets ofnon-zero measure in any vicinity of each r-sliding pointnear point t0; x0. Then this provides, with sufficientlylarge K, for the existence of r-sliding mode in some vici-nity of the point t0; x0:

Proof: Proposition 2 is a straight-forward conse-quence of Proposition 1 and equation (4). &

The trivial controller u Ksign satisfiesProposition 2. Usually, however, such a mode is notstable. The r-sliding mode motion is described by theequivalent control method (Utkin 1992), on the otherhand, this dynamics coincides with the zero-dynamics(Isidori 1989) of the corresponding systems.

The problem is to find a discontinuous feedbacku Ut; x causing the appearance of a finite-time con-vergent r-sliding mode in (3). That new controller has togeneralize the standard 1-sliding relay controlleru Ksign . Hence, gt; y and ht; y in (4) areassumed to be bounded, g > 0. Thus, it is requiredthat for some Km;KM;C > 0

0 < Km @

@ur KM; jrju0j C 5

3.1. Solutions of the problem with relative degreesr 1; 2

In case r 1 calculation shows that

_t; x; u 0t t; x 0xt; xat; x 0xt; xbt; xu

and the problem is easily solved by the standard relaycontroller u sign , with a > C=Km. Hereh _ju0

0t 0xa is globally bounded and

g @=@u _ 0xb. The first-order real-sliding accuracywith respect to the sampling interval is ensured.

Let r 2. The following list includes only few mostknown controllers. The so-called twisting controller(Levantovsky 1985, Emelyanov et al. 1986, Levant1993) and the convergence conditions are given by

u r1sign r2 sign _; r1 > r2 > 0; 6

r1 r2Km C > r1 r2KM C;

r1 r2Km > C

A particular case of the controller with prescribed con-vergence law (Emelyanov et al. 1986, Levant 1993) isgiven by

u sign _ jj1=2 sign ; ; > 0; 7

Km C > 2=2

Controller (7) is close to terminal sliding mode control-lers (Man et al. 1994). The so-called sub-optimal con-troller (Bartolini et al. 1998, 1999 a, b) is given by

u r1 sign =2 r2 sign ; r1 > r2 > 0 8

2r1 r2Km C > r1 r2KM C;

r1 r2Km > C 9

where is the current value of detected at the closesttime when _ was 0. The initial value of is 0. Anycomputer implementation of this controller requires suc-cessive measurements of _ or with some time step.Usually the detection of the moments when _ changesits sign is performed. The control value u depends actu-ally on the history of _ and measurements, i.e. on _and .

Theorem 1 (Levant 1993, Bartolini et al. 1998): 2-slid-ing controllers (6)(8) provide for finite-time convergenceof any trajectory of (3) and (5) to 2-sliding mode 0.The convergence time is a locally bounded function ofthe initial conditions.

Let the measurements be carried out at times ti withconstant step > 0, i ti; xti, i i i1,t 2 ti; ti1. Substituting i for , sign i for sign _, andsign (i jij1=2 signi) for sign ( _ jj1=2 sign)achieve discrete-sampling versions of the controllers.

Theorem 2 (Levant 1993, Bartolini et al. 1998): Dis-crete-sampling versions of controllers (6)(8) provide for


the establishment of the inequalities jj < 02; j _j 1; _uu ; _ with juj 1.

Consider the general case. Let u be defined from theequality

d

dtPr1

d

dt

sign Pr1

dt

dt

;

where Pr1 r1 1r2 r1 is astable polynomial, i 2 R. In the case r 1, P0 1achieve the standard 1-sliding mode. Let r > 1. The r-sliding mode exists here at the origin and is asymptoti-cally stable. There is also a 1-sliding mode on the mani-fold Pr1d=dt 0. Trajectories transfer in finite timeinto the 1-sliding mode on the manifold Pr1d=dt 0and then exponentially converge to the r-sliding mode.Dynamic-sliding-mode controllers (Sira-Ramrez 1993,Spurgeon and Lu 1997) are based on such modes.Unfortunately, due to the dependence on higher-orderderivatives of the control is not bounded here evenwith small . Also the accuracy here is the same as ofthe 1-sliding mode.

4. Building an arbitrary-order sliding controller

Let p be any positive number, p r. Denote

N1;r jjr1=r

Ni;r jjp=r j _jp=r1 ji1jp=ri1ri=p;

i 1; . . . ; r 1

Nr1;r jjp=r _jp=r1 jr2jp=21=p

0;r

1;r _ 1N1;r sign

i;r i iNi;r sign i1;r; i 1; . . . ; r 1

where 1; . . . ; r1 are positive numbers.

Theorem 3 (Levant 1998 a, 2001): Let system (3) haverelative degree r with respect to the output function and (5) be fulfilled. Suppose also that trajectories of sys-tem (3) are infinitely extendible in time for any Lebes-

gue-measurable bounded control. Then with properlychosen positive parameters 1; . . . ; r1, the controller

u sign r1;r; _; . . . ; r1 10

leads to the establishment of an r-sliding mode 0attracting each trajectory in finite time. The convergencetime is a locally bounded function of initial conditions.

The proof of Theorem 3 is given in Appendix 1. It isthe first publication of the proof. The assumption on thesolution extension possibility means in practice that thesystem be weakly minimum phase. The positive par-ameters 1; . . . ; r1 are to be chosen sufficiently largein the index order. They determine a controller familyapplicable to all systems (3) of relative degree r satisfy-ing (5) for some C, Km and KM. Parameter > 0 is to bechosen specifically for any fixed C, Km and KM. Theproposed controller may be generalized in many ways.For example, coefficients of Ni;r may be any positivenumbers, equation (10) can be smoothed (Levant 1999).

Certainly, the number of choices of i is infinite.Here are a few examples with i tested for r 4, pbeing the least common multiple of 1; 2; . . . ; r. The firstis the relay controller, the second coincides with (7).

1: u sign ;

2: u sign _ jj1=2sign

3: u 2j _j3 j21=6sign _ jj2=3sign

4: u f 36 _4 jj31=12sign _4

jj31=6sign _ 0:5jj3=4sign g

5: u sign 4 4jj12 j _j15 jj20

jj301=60sign 3jj12 j _j15

jj201=30sign 2jj12

j _j151=20sign _ 1jj4=5sign

Obviously, parameter is to be taken negative with@=@ur < 0. Controller (10) is certainly insensitiveto any disturbance which keeps the relative degree and(5). No matching condition having been supposed, theresidual uncertainty reveals itself in the r-sliding motionequations (in other words, in zero dynamics).

The idea of the controller is that a 1-sliding mode isestablished on the smooth parts of the discontinuity setG of (10) (figures 2, 3). That sliding mode is described bythe differential equation r1;r 0 providing in its turnfor the existence of a 1-sliding mode r2;r 0. But theprimary sliding mode disappears at the moment whenthe secondary one is to appear. The resulting movement

928 A. Levant

takes place in some vicinity of the cylindrical subset of Gsatisfying r2;r 0, transfers in finite time into somevicinity of the subset satisfying r3;r 0 and so on.While the trajectory approaches the r-sliding set, set Gretracts to the origin in the coordinates ; _; . . . ; r1.

Controller (10) requires the availability of; _; . . . ; r1. That information demand may be low-ered. Let the measurements be carried out at times tiwith constant step > 0. Consider the controller

ut sign r2i r1Nr1;ri; _i; . . . ; r2i

sign r2;ri; _i; . . . ; r2i 11

where ji

jti; xti;r2i

r2i

r2i1 ;

t 2 ti; ti1.

Theorem 4 (Levant 1998 b, 2001 a): Under conditionsof Theorem 1 with discrete measurements both algo-rithms (10) and (11) provide in finite time for fulfilmentof the inequalities

jj < a0 r; j _j < a1 r1; . . . ; jr1j < ar1for some positive constants a0; a1; . . . ; ar1. The conver-gence time is a locally bounded function of initial con-ditions.

That is the best possible accuracy attainable withdiscontinuous r separated from zero (Levant 1993).The proof of Theorem 4 is given in Appendix 1.Following are some remarks on the usage of the pro-posed controllers.

Convergence time may be reduced increasing thecoefficients j. Another way is to substitute

j j for j, r for and for in (10) and (11), > 1,causing convergence time to be diminished approxi-mately by times. As a result the coefficients of Ni;rwill differ from 1.

Local application of the controller. In practical appli-cations condition (5) is often invalid globally, but stillholds in some restricted area of the state space contain-ing the actual region of the system operation. In fact,that is always true, if the constraint keeping problem iswell posed from the engineering point of view. The prac-tical implementation of the controller is straight-for-ward in that case and is based on the following simpleproposition proved in Appendix 1.

Proposition 3: Under the conditions of Theorem 3, forany RM > 0 there exists such Rm, RM > Rm > 0, thatany trajectory starting in the disc of radius Rm which iscentred at the origin of the space ; _; . . . ; r1 doesnot leave the larger disc of the radius RM while conver-ging to the origin. Moreover, RM Rm and the conver-gence time can be made arbitrarily small choosing1; . . . ; r1; , sufficiently large in the list order.

Implementation of r-sliding controller when the rela-tive degree is less than r. Let the relative degree k of theprocess be less than r. Introducing successive time deri-vatives u; _uu; . . . ; urk1 as a new auxiliary variables andurk as a new control, achieve a system with relativedegree r. Condition (5) is locally satisfied for the newcontrol. The standard r-sliding controller may be nowlocally applied. The resulting control ut is anr k 1-smooth function of time with k < r 1, aLipschitz function with k r 1 and a bounded infi-nite-frequency switching function with k r. A globalcontroller was developed by Levant (1993) forr 2; k 1.

Chattering removal. The same trick removes thechattering effect. For example, substituting ur1 for uin (10), receive a local r-sliding controller to be usedinstead of the relay controller u sign and attainthe rth-order sliding precision with respect to bymeans of an (r 2)-times differentiable control with aLipschitzian (r 2)th time derivative. It has to be modi-fied like in (Levant 1993) in order to provide for theglobal boundedness of u and global convergence.

Controlling systems non-linear on control. Consider asystem _xx f t; x; u non-linear on control. Let


Figure 2. The idea of the r-sliding controller.

Figure 3. The 3-sliding controller discontinuity set.

@=@uit; x; u 0 for i 1; :::; r 1; @=@urt;x; u > 0. It is easy to check that

r1 Lr1u @

@ur _uu;

Lu @

@t @

@xf t; x; u

The problem is now reduced to that considered abovewith relative degree r 1 by introducing a new auxiliaryvariable u and a new control v _uu.

Real-time control of output variables. The implemen-tation of the above-listed r-sliding controllers requiresreal-time observation of the successive derivatives _,; . . . ; r1. In case system (3) is known and the fullstate is available, these derivatives may be directly cal-culated. In the real uncertainty case the derivatives arestill to be real-time evaluated in some way. Thus, onewould not theoretically need to know any model of thecontrolled process, only the relative degree and threeconstants from (5) were needed in order to adjust thecontroller. Unfortunately, the problem of successivereal-time exact differentiation is usually considered aspractically insoluble. Nevertheless, as is shown in thenext section, the boundedness of r, which followsfrom (4) and (5) allows robust exact estimation of _,; . . . ; r1 in real time.

5. Arbitrary-order exact robust differentiator

Real-time differentiation is an old and well-studiedproblem. The main difficulty is the obvious differentia-tion sensitivity to input noises. The popular high-gaindifferentiators (Atassi and Khalil 2000) provide for anexact derivative when their gains tend to infinity.Unfortunately at the same time their sensitivity tosmall high-frequency noises also infinitely grows. Withany finite gain values such a differentiator has also afinite bandwidth. Thus, being not exact, it is, at thesame time, insensitive with respect to high-frequencynoises. Such insensitivity may be considered both asadvantage or disadvantage depending on the circum-stances. Another drawback of the high-gain differentia-tors is their peaking effect: the maximal output valueduring the transient grows infinitely when the gainstend to infinity.

The main problem of the differentiator feedbackapplication is the so-called separation problem. Theseparation principle means that a controller and anobserver (differentiator) can be designed separately, sothat the combined observer-controller output feedbackpreserve the main features of the controller with the fullstate available. The separation principle was proved forasymptotic continuous-feedback stabilization of auton-omous systems with high-gain observers (Atassi andKhalil 2000, Isidori et al. 2000). That important result

is realized in spite of the non-exactness of high-gainobservers with any fixed finite gain values. The qualita-tive explanation is that the output derivatives of allorders vanish during the continuous-feedback stabiliz-ation. Thus, the frequency of the signal to be differen-tiated also vanishes and the differentiator provides forasymptotically exact derivatives. On the contrary, in thecase considered in the previous section r is chatteringwith a finite magnitude and a frequency tending to infi-nity while approaching the r-sliding mode. Thus, thesignal is problematic for a high-gain differentiator.Indeed, the closer to the r-sliding mode, the highergain is needed to produce a good derivative estimationof r1. Actually, the high-gain differentiator will dif-ferentiate only the slowly changing average output com-ponent. As a result, convergence into some vicinity ofthe r-sliding mode could only be attained.

The sliding-mode differentiators (Golembo et al.1976, Yu and Xu 1996) also do not provide for exactdifferentiation with finite-time convergence due to theoutput filtration. The differentiator by Bartolini et al.(2000) is based on a 2-sliding-mode controller usingthe real-time measured sign of the derivative to be cal-culated. Therefore, the first finite difference of the differ-entiator input is used with the sampling stepproportional to the square root of the maximal noisemagnitude. That is rather inconvenient and requirespossibly lacking information on the noise.

Exact derivatives may be calculated by successiveimplementation of a robust exact first-order differentia-tor (Levant 1998 a) with finite-time convergence. Thatdifferentiator is based on 2-sliding mode and is provedto feature the best possible asymptotics in the presenceof infinitesimal Lebesgue-measurable measurementnoises, if the second time derivative of the unknownbase signal is bounded. The accuracy of that differentia-tor is proportional to "1=2, where " is the maximal meas-urement-noise magnitude and is also assumed to beunknown. Therefore, having been n times successivelyimplemented, that differentiator will provide for the nth-order differentiation accuracy of the order of "2

n.Thus, the differentiation accuracy deteriorates rapidly.On the other hand, it is proved by Levant (1998 a) thatwhen the Lipschitz constant of the nth derivative ofthe unknown clear-of-noise signal is bounded bya given constant L, the best possible differentiationaccuracy of the ith derivative is proportional toLi=n1"n1i=n1, i 0; 1; . . . ; n. Therefore, a specialdifferentiator is to be designed for each differentiationorder.

Let input signal f t be a function defined on 0;1consisting of a bounded Lebesgue-measurable noisewith unknown features and an unknown base signalf0t with the nth derivative having a known Lipschitzconstant L > 0. The problem is to find real-time robust

930 A. Levant

estimations of _ff0t; ff0t; . . . ; fn0 t being exact in the

absence of measurement noises.Two similar recursive schemes of the differentiator

are proposed here. Let an n 1th-order differentiatorDn1f ;L produce outputs Din1; i 0; 1; . . . ; n 1,being estimations of f0; _ff0; ff0; . . . ; f

n10 for any input

f t with f n10 having Lipschitz constant L > 0. Thenthe nth-order differentiator has the outputs zi Din,i 0; 1; . . . ; n, defined as

_zz0 v; v 0jz0 f tjn=n1sign z0 f t z1

z1 D0n1v;L; . . . ; zn Dn1n1v;L

9=;

12Here D0f ;L is a simple non-linear filter

D0 : _zz sign z f t; > L 13Thus, the first-order differentiator coincides here

with the above-mentioned differentiator (Levant1998 a) :

_zz0 v; v 0jz0 f tj1=2sign z0 f t z1

_zz1 1sign z1 v 1sign z0 f t

9=;

14Another recursive scheme is based on the differentia-

tor (14) as the basic one. Let ~DDn1f ;L be such a newn 1th-order differentiator, n 1; ~DD1f ;L coincid-ing with the differentiator D1f ;L) given by (14).Then the new scheme is defined as

_zz0 v

v 0jz0 f tjn=n1sign z0 f t w0 z1

_ww0 0jz0 f tjn1=n1sign z0 f t

z1 ~DD0n1v;L; . . . ; zn ~DDn1n1v;L

9>>>>>>>>=>>>>>>>>;15

The resulting 2nd-order differentiator (Levant 1999) is

_zz0 v0; v0 0jz0 f tj2=3sign z0 f t w0 z1

_ww0 0jz0 f tj1=3sign z0 f t

_zz1 v1; v1 1jz1 v0j1=2sign z1 v0 w1

_ww1 1sign z1 v0; z2 w1Similarly, a 2nd-order differentiator from each of

these sequences may be used as a base for a newrecursive scheme. An infinite number of differentiatorschemes may be constructed in this way. The onlyrequirement is that the resulting systems be homo-geneous in a sense described further. While the author

has checked only the above two schemes (12), (13)and (14), (15), the conjecture is that all suchschemes produce working differentiators, providedsuitable parameter choice. Differentiator (12) takes onthe form

_zz0 v0;

v0 0jz0 f tjn=n1sign z0 f t z1

_zz1 v1; v1 1jz1 v0jn1=nsign z1 v0 z2

..

.

_zzn1 vn1;

vn1 n1jzn1 vn2j1=2sign zn1 vn2 zn

_zzn n sign zn vn1

9>>>>>>>>>>>>>>>>>>>>>=>>>>>>>>>>>>>>>>>>>>>;

16

Theorem 5: The parameters being properly chosen, thefollowing equalities are true in the absence of inputnoises after a finite time of a transient process

z0 f0t; zi vi1 fi0 t; i 1; . . . ; n

Moreover, the corresponding solutions of the dynamicsystems are Lyapunov stable, i.e. finite-time stable(Rosier 1992). The theorem means that the equalitieszi f

i0 t are kept in 2-sliding mode, i 0; . . . ; n 1.

Here and further all Theorems are proved in Appendix2.

Theorem 6: Let the input noise satisfy the inequalityjf t f0tj ". Then the following inequalities areestablished in finite time for some positive constants i,i depending exclusively on the parameters of thedifferentiator

jzi fi0 tj i"

ni1=n1; i 0; . . . ; n

jvi f i10 tj i"ni=n1; i 0; . . . ; n 1

Consider the discrete-sampling case, when z0tj f tjis substituted for z0 f t with tj t < tj1; tj1 tj > 0.

Theorem 7: Let > 0 be the constant input samplinginterval in the absence of noises. Then the following in-equalities are established in finite time for some positiveconstants i, i depending exclusively on the parametersof the differentiator


jzi fi0 tj i

ni1; i 0; . . . ; n

jvi fi10 tj i

ni; i 0; . . . ; n 1

In particular, the nth derivative error is proportionalto . The latter theorem means that there are a numberof real sliding modes of different orders. Nevertheless,nothing can be said on the derivatives of vi and of_zzi, because they are not continuously differentiablefunctions.

Homogeneity of the differentiators. The differentia-tors are invariant with respect to the transformation

t; f ; zi; vi;wi7!t; n1f ; ni1zi; nivi; niwi

The parameters i, i are to be chosen recursively insuch a way that 1; 1; . . . ; n; n, n provide for theconvergence of the n 1th-order differentiator withthe same Lipschitz constant L, and 0, 0 be sufficientlylarge (0 is chosen first). The best way is to choose themby computer simulation. A choice of the 5th-order dif-ferentiator parameters with L 1 is demonstrated in } 7.Recall that it contains parameters of all lower-orderdifferentiators. Substituting f t=L for f t and takingnew coordinates z 0i Lzi, v 0i Lvi, w 0i Lwi achieve thefollowing proposition.

Proposition 4: Let parameters 0i, 0i; i 0; 1; . . . ; n,of differentiators (12), (13) or (14), (15) provide for ex-act nth-order differentiation with L 1. Then the para-meters i 0iL2=ni1, i 0iL1=ni1 are valid forany L > 0 and provide for the accuracyjzi f i0 tj iLi=n1"ni1=n1 for some i 1.

The separation principle is trivially fulfilled for theproposed differentiator. Indeed, the differentiatorbeing exact, the only requirements for its implementa-tion are the boundedness of some higher-order deriva-tive of its input and the impossibility of the finite-timeescape during the differentiator transient. Hence, thedifferentiator may be used in almost any feedback.Mark that the differentiator transient may be made arbi-trarily short by means of the parameter transformationfrom Proposition 4, and the differentiator does not fea-ture peaking effect (see the proof of Theorem 5 in theAppendices).

Remarks: It is easy to see that the kth-order differen-tiator provides for a much better accuracy of the lthderivative, l < k, than the lth-order differentiator (The-orem 6). A similar idea is realized to improve the firstderivative by Krupp et al. (2001). It is easy to checkthat after exclusion of the variables vi differentiator(16) may be rewritten in the non-recursive form

_zz0 0jz0 f tjn=n1sign z0 f t z1

_zzi ijz0 f tjni=n1sign z0 f t zi1;

i 1; . . . ; n 1

_zzn n sign z0 f t

for some positive i calculated on the basis of 0 . . . ; n.

6. Universal output-feedback SISO controller

The results of this section have been just recentlyobtained (Levant 2002), and the author supposes todescribe them in a special paper. Thus, only a briefdescription is provided. Consider uncertain system (3),(5). Combining controller (10) and differentiator (16)achieve a combined single-inputsingle-output (SISO)controller

u sign r1;rz0; z1; . . . ; zr1

_zz0 v0; v0 0;0L1=rjz0 jr1=r sign z0 z1

_zz1 v1;

v1 0;1L1=r1jz1 v0jr2=r1 sign z1 v0 z2

..

.

_zzr2 vr2;

vr2 0;r2L1=2jzr2 vr3j1=2 sign zr2 vr3 zr1

_zzr1 0;r1L sign zr1 vr2

where parameters i 0;iL1=rI of the differentiatorare chosen according to the conditionjrj L;L C KM. In their turn, parameters 0iare chosen in advance for L 1 (Proposition 4). Thus,parameters of controller (10) are chosen separately ofthe differentiator. In case when C and KM are known,only one parameter is really needed to be tuned, other-wise both L and might be found in computer simula-tion. Theorems 3 and 4 hold also for the combinedoutput-feedback controller. In particular, under the con-ditions of Theorem 3 the combined controller providesfor the global convergence to the r-sliding mode 0with the transient time being a locally bounded functionof the initial conditions.

On the other hand, let the initial conditions of thedifferentiator belong to some compact set. Then for anytwo embedded discs centred at the origin of the space; _; . . . ; r1 the parameters of the combined control-ler can be chosen in such a way that all trajectoriesstarting in the smaller disc do not leave the larger discduring their finite-time convergence to the origin. The

932 A. Levant

convergence time can be made arbitrarily small. Thatallows for the local controller application.

With discrete measurements, in the absence of inputnoises, the controller provides for the rth-order real slid-ing sup jj r, where is the sampling interval.Therefore, the differentiator does not spoil the r-slidingasymptotics if the input noises are absent. It is alsoproved that the resulting controller is robust and pro-vides for the accuracy proportional to the maximal errorof the input measurement (the input noise magnitude).Note once more that the proposed controller does notrequire detailed mathematical model of the process to beknown.

7. Simulation examples

7.1. Numeric differentiation

Following are equations of the 5th-order differentia-tor with simulation-tested coefficients for L 1

_zz0 v0; v0 12jz0 f tj5=6sign z0 f t z1 17

_zz1 v1; v1 8jz1 v0j4=5 sign z1 v0 z2 18



_zz4 v4; v4 1:5jz4 v3j1=2 sign z4 v3 z5 21

_zz5 1:1 sign z5 v4 22

The differentiator parameters can be easily changed, forit is not very sensitive to their values. The tradeoff is asfollows: the larger the parameters, the faster the conver-gence and the higher sensitivity to input noises and thesampling step.

As mentioned, differentiator (17)(22) also containsdifferentiators of the lower orders. For example, accord-ing to Proposition 4, the second-order differentiator forthe input f with |

f j L takes on the form

_zz0 v0; v0 3L1=3jz0 f j2=3 sign z0 f z1

_zz1 v1; v1 1:5L1=2jz1 v0j1=2 sign z1 v0 z2

_zz2 1:1L sign z2 v1

Differentiator (17)(22) and its 3rd-order sub-differen-tiator (19)(22) (differentiating here the internal variablev1) were used for simulation. Initial values of the differ-entiator state were taken zero with exception for theinitial estimation z0 of f , which is taken equal to theinitial measured value of f. The base input signal

f0t 0:5 sin 0:5t 0:5 cos t 23was taken for the differentiator testing. Derivatives off0t do not exceed 1 in absolute value.

Third-order differentiator. The measurement step

103 was taken, noises are absent. The attainedaccuracies are 5:8 1012, 1:4 108, 1:0 105 and0.0031 for the signal tracking, the first, second and

third derivatives respectively. The derivative tracking

deviations changed to 8:3 1016, 1:8 1011,1:2 107 and 0.00036 respectively after was reducedto 104. That corresponds to Theorem 7.

Fifth-order differentiator. The attained accuracies are

1:1 1016, 1:29 1012, 7:87 1010, 5:3 107,2:0 104 and 0.014 for tracking the signal, the first,second, third, fourth and fifth derivatives respectively

with 104 (figure 4(a)). There is no significantimprovement with further reduction of . The authorwanted to demonstrate the 10th-order differentiation,

but found that differentiation of the order exceeding 5

is unlikely to be performed with the standard software.

Further calculations are to be carried out with precision

higher than the standard long double precision (128 bits

per number).


Figure 4. Fifth-order differentiation.

Sensitivity to noises. The main problem of the differ-entiation is certainly its well-known sensitivity to noises.As we have seen, even small computer calculation errorsappear to be a considerable noise in the calculation ofthe fifth derivative. Recall that, when the nth derivativehas the Lipschitz constant 1 and the noise magnitude is", the best possible accuracy of the ith-order differen-tiation, i n, is ki; n"ni1=n1 (Levant 1998 a),where ki; n > 1 is a constant independent on the differ-entiation realization. That is a minimax (worst case)evaluation. Since differentiator (17)(22) assumes thisLipschitz input condition, it satisfies this accuracyrestriction as well (see also Theorems 6 and 7). In par-ticular, with the noise magnitude " 106 the maximal5th derivative error exceeds "1=6 0:1. For comparison,if the successive first-order differentiation were used, therespective maximal error would be at least "25 0:649and some additional conditions on the input signalwould be required. Taking 10% as a border, achievethat the direct successive differentiation does not givereliable results starting with the order 3, while the pro-posed differentiator may be used up to the order 5.

With the noise magnitude 0.01 and the noise fre-quency about 1000 the 5th-order differentiator producesestimation errors 0.000 42, 0.0088, 0.076, 0.20, 0.34 and0.52 for signal (23) and its five derivatives respectively(figure 4(b)). The differentiator performance does notsignificantly depend on the noise frequency. The authorfound that the second differentiation scheme (15) pro-vides for slightly better accuracies.

7.2. Output-feedback control simulation

Consider a simple kinematic model of car control(Murray and Sastry 1993)

_xx v cos; _yy v sin

_ v=l tan

_ u

where x and y are Cartesian coordinates of the rear-axlemiddle point, is the orientation angle, v is the long-itudinal velocity, l is the length between the two axlesand is the steering angle (figure 5). The task is to steerthe car from a given initial position to the trajectoryy gx, while x; gx and y are assumed to be meas-ured in real time. Note that the actual control here is and _ u is used as a new control in order to avoiddiscontinuities of . Any practical implementation ofthe developed here controller would require some real-time coordinate transformation with approaching=2. Define

y gx 24

Let v const 10m/s, l 5m, gx 10 sin 0:05x5, x y 0 at t 0. The relative degree ofthe system is 3 and the listed 3-sliding controller maybe applied here. The resulting steering angle dependenceon time is not sufficiently smooth (Levant 2001 b),therefore the relative degree is artificially increased upto 4, _uu having been considered as a new control. The 4-sliding controller from the list in } 3 is applied now, 20 is taken. The following 3rd order differentiatorwas implemented:

_zz0 v0; v0 25jz0 j3=4 sign z0 z1 25



_zz3 500 sign z3 v2 28

The coefficient in (28) is large due to the large values of4, other coefficients were taken according toProposition 4 and (17)(22). During the first half-secondthe control is not applied in order to allow the conver-gence of the differentiator. Substituting z0, z1, z2 and z3for , _, and respectively, obtain the following 4-sliding controller

u 0; 0 t < 0:5

u 20 sign fz3 3z62 z41 jz0j31=12 sign z2

z41 jz0j31=6 sign z1 0:5jz0j3=4 sign z0g

t 0:5

The trajectory and function y gx with thesampling step 104 are shown in figure 6(a). Theintegration was carried out according to the Eulermethod, the only method reliable with sliding-modesimulation. Graphs of , _, , are shown in figure

934 A. Levant

Figure 5. Kinematic car model.

6(b). The differentiator performance within the first 1.5 sis demonstrated in figure 6(c). The steering angle graph(actual control) is presented in figure 6(d). The slidingaccuracies jj 9:3 108, j _j 7:8 105,jj 6:6 104, jj 0:43 were attained with thesampling time 104.

8. Conclusions and discussion of the obtained results

Arbitrary-order real-time exact differentiationtogether with the arbitrary-order sliding controllers pro-vide for full SISO control based on the input measure-ments only, when the only information on the controlleduncertain weakly-minimum-phase process is actually itsrelative degree.

A family of r-sliding controllers with finite time con-vergence is presented for any natural number r, provid-ing for the full real-time control of the output variable ifthe relative degree r of the dynamic system is constantand known. Whereas 1- and 2-sliding modes were usedmainly to keep auxiliary constraints, arbitrary-ordersliding controllers may be considered as general-purposecontrollers. In case the mathematical model of thesystem is known and the full state is available, thereal-time derivatives of the output variable are directlycalculated, and the controller implementation is

straightforward and does not require reduction of the

dynamic system to any specific form. If boundedness

restrictions (5) are globally satisfied, the control is also

global and the input is globally bounded. Otherwise the

controller is still locally applicable.

In the uncertainty case a detailed mathematical

model of the process is not needed. Necessary time deri-

vatives of the output can be obtained by means of the

proposed robust exact differentiator with finite-time

convergence. The proposed differentiator allows real-

time robust exact differentiation up to any given order

l, provided the next l 1th derivative is bounded by aknown constant. These features allow wide application

of the differentiator in non-linear control theory.

Indeed, after finite time transient in the absence of

input noises its outputs can be considered as exact direct

measurements of the derivatives. Therefore, the separa-

tion principle is trivially true for almost any feedback.

At the same time, in the presence of measurement noises

the differentiation accuracy inevitably deteriorates

rapidly with the growth of the differentiation order

(Levant 1998 a), and direct observation of the deriva-

tives is preferable. The exact derivative estimation does

not require tending some parameters to infinity or to

zero. Even when treating noisy signals, the differentiator


Figure 6. 4-sliding car control.

performance only improves with the sampling stepreduction.

The resulting controller provides for extremely hightracking accuracy in the absence of noises. The slidingaccuracy is proportional to r, being a sampling periodand r being the relative degree. That is the best possibleaccuracy with discontinuous rth derivative of the output(Levant 1993). It may be further improved increasingthe relative degree artificially, which produces arbitrarilysmooth control and removes also the chattering effect.

The proposed controllers are easily developed forany relative degree, at the same time most of the practi-cally important problems in output control are coveredby the cases when relative degree equals 2, 3, 4 and 5.Indeed, according to the Newton law, the relative degreeof a spatial variable with respect to a force, being under-stood as a control, is 2. Taking into account somedynamic actuators, achieve relative degree 3 or 4. Ifthe actuator input is required to be a continuousLipschitz function, the relative degree is artificiallyincreased to 4 or 5. Recent results (Bartolini et al.1999 b) seem to allow the implementation of the devel-oped controllers for general multi-inputmulti-outputsystems.

Appendix 1. Proofs of Theorems 3, 4 and Proposition 3

The general idea of the proofs is presented in } 4 andis illustrated by figures 2 and 3.

Preliminary notions: The following notions areneeded to understand the proof. They are based on re-sults by Filippov (1988).

Differential inclusion _ 2 X, 2 Rm is calledfurther Filippov inclusion if for any :

1. X is a closed non-empty convex set;2. X fv 2 Rmjkvk g, where is a con-

tinuous function;

3. the maximal distance of the points of X 0 fromX tends to zero when 0 ! .

Recall that any solution of a differential inclusion isan absolutely continuous function satisfying the inclu-sion almost everywhere, and that any differential equa-tion with a discontinuous right-hand side is understoodas equivalent to some Filippov inclusion.

The graph of a differential inclusion _ 2 X, 2 Rmis the set f; _ 2 Rm Rmj _ 2 Xg. A differentialinclusion _ 2 X 0 is called "-close to the Filippov inclu-sion _ 2 X in some region if any point of the graph of_ 2 X 0 is distanced by not more than " from the graphof _ 2 X. It is known that within any compact regionsolutions of _ 2 X 0 tend to some solutions of _ 2 Xuniformly on any finite time interval with " ! 0. In the

special case m 2 X; _; . . . ; m1, 2 R, which isconsidered in the present paper, the graph may be con-sidered as a set from Rm R. An inclusionm 2 X 0; _; . . . ; m1 corresponding to the closed"-vicinity of that graph is further called the "-swolleninclusion. It is easy to see that this is a Filippov inclu-sion. An "-swollen differential equation is the inclusioncorresponding to the "-vicinity of the correspondingFilippov inclusion.

Proof of Theorem 3: Consider the motion of a projec-tion trajectory of (3) and (10) in coordinates ,_; . . . ; r1

r Lraet; x u@

@urt; x; u 29

Taking into account (5) achieve a differential inclusion

r 2 C;C Km;KMu 30

which will be considered from now on instead of the realequality (29). The operations on sets are naturallyunderstood here as sets of operation results for all poss-ible combinations of the operand set elements. Control uis given by (10) or (11).

A given point P is called here a discontinuity point ofa given function, if for any point set N of zero measureand any vicinity O of P there are at least two differentlimit values of the function when point p 2 O=Napproaches P. Let G be the closure of the discontinuityset of sign r1;r; _; . . . ; r1 or, in other words, ofcontrol (10).

Lemma 1: Set G partitions the whole space; _; . . . ; r1 into two connected open componentssatisfying r1;r; . . . ; r1 > 0 and r1;r; . . . ;r1 < 0 respectively. Any curve connecting pointsfrom different components has a non-empty intersectionwith G.

Proof: Consider any equation i;r; _; . . . ; i 0;i 1; . . . ; r 1. It may be rewritten in the form

i i; _; . . . ; i1

iNi;r; _; . . . ; i1 sign i1;r; _; . . . ; i1

Let Si be the closure of the discontinuity set of sign

i;r; i 0; . . . ; r 1;G Sr1, and S0 f0g R (figure7). Each set Si lies in the space ; _; . . . ;

i and is, actu-ally, a modification of the graph ofi i; _; . . . ; i1.

The Lemma is proved by the induction principle.Obviously, S0 partitions R into two open connectedcomponents. Let Si1 divide the space R

i with coordi-nates ; _; . . . ; i1 into two open connected compon-

936 A. Levant

ents Oi1 and Oi1 with i1;r > 0 and i1;r < 0 respect-

ively. It is easy to see (figure 7) that

Si f; _; . . . ; ij jij

iNi;r; _; . . . ; i1& ; _; . . . ; i1 2 Si1_

i iNi;r; _; . . . ; i1 sign i1;r; _; . . . ; i1

& ; _; . . . ; i1 =2Si1g

Oi f; _; . . . ; iji > iNi;r; _; . . . ; i1_

i > iNi;r; _; . . . ; i1

& ; _; . . . ; i1 2 Oi1g

Oi f; _; . . . ; iji < iNi;r; _; . . . ; i1_

i < iNi;r; _; . . . ; i1

& ; _; . . . ; i1 2 Oi1g

Considering the division of the whole space ; _; . . . ; i

in the sets jij iNi;r; i < iNi;r and i > iNi;rit is easily proved that Oi and O

i are connected and

open. &

Consider the transformation Gv : t; ; _; . . . ; r17!vt; vr; vr1 _; . . . ; vr1. It is easy to see that this lineartransformation complies with understanding j as

coordinates and as derivatives as well. It is also easy tocheck that inclusion (30) and (10) is invariant withrespect to Gv; v > 0. That invariance implies that anystatement invariant with respect to Gv is globally true ifit is true on some set E satisfying the condition[v0GvE Rr. That reasoning is called further homo-geneity reasoning. For example, it is sufficient to provethe following lemma only for trajectories with initial con-ditions close to the origin.

Lemma 2: With sufficiently large any trajectory ofinclusion (30) and (10) hits G in finite time.

Proof: Obviously, it takes finite time to reach the re-gion jr1j r1Nr1;r. It is also easy to check thatit takes finite time for any trajectory in a sufficientlysmall vicinity of the origin to cross the entire regionjr1j r1Nr1;r if no switching happens. Thus, ac-cording to Lemma 1 set G is also encountered on theway. The homogeneity reasoning completes theproof. &

Lemma 3: There is a 1-sliding mode on r1;r in thecontinuity points of r1;r with sufficiently large .There is such a choice of j that each differential equa-tion i;r 0; i 1; . . . ; r 1 provides for the existenceof a 1-sliding mode in the space ; _; . . . ; i1 on themanifold i1 i1; _; . . . ; i2 in continuitypoints of i1.

Proof: It is needed to prove that _r1;r signr1;r 0 the inequality C KM < "Nr1;r issatisfied outside of some bounded vicinity of the origin.Thus, similarly to the Theorem 3 proof, sufficientlysmall " being taken, finite-time convergence into somevicinity D of the origin is provided. Obviously, the trans-formation G: t; ; _; . . . ; r17!t; r; r1 _; . . . ;r1 transfers (30), (11) into the same inclusion, butwith the new measurement interval . Thus, GD is anattracting set corresponding to that new value of themeasurement interval. &

Proof of Proposition 3: As is seen from the proof ofTheorems 3 and 4 the only restriction on the choice of1 is formulated in Lemma 3, and is chosen with re-spect to Lemma 3 and is sufficiently enlarged after-wards. Thus, at first the existence of the virtual slidingmodes i;r 0, i 1; . . . ; r 1 is provided and than is taken so large that the real motion will take place inarbitrarily small vicinity of these modes. Thus, theparameters may be chosen sufficiently large in the or-der 1; . . . ; r1, so that the convergence from anyfixed compact region of initial conditions be arbitrarilyfast, and at the same time the overregulation be arbi-trarily small. &

Appendix 2. Proofs of Theorems 57

Consider for simplicity differentiator (16). The prooffor differentiator (15) is very similar. Introducefunctions 0 z0 f0t, 1 z1 _ff0t; . . . ; n zn f

n0 t, f t f0t. Then any solution of (16)

satisfies the following differential inclusion understoodin the Filippov sense

_0 0j0 tjn=n1sign 0 t 1 31

_1 1j1 _0jn1=nsign 1 _0 2

..

.

_n1 n1jn1 _n2j1=2sign n1 _n2 n;

_n 2 n sign n _n1 L;L

9>>>>>>>>>>=>>>>>>>>>>;

32

where t 2 "; " is a Lebesgue-measurable noisefunction. It is important to mark that (31), (32) doesnot remember anything on the unknown input basic

signal f0t:System (31), (32) is homogeneous with " 0, its tra-

jectories are invariant with respect to the transformation

G : t; i; ; "7!t; ni1i; n1; n1" 33

Define the main features of differential inclusion (31),(32) which hold with a proper choice of the parameters

i.

Lemma 7: Let t satisfy the condition that the inte-gral

jtjdt over a time interval is less than some

fixed K > 0. Then for any 0 < Si < S0i ; i 0; . . . ; n,

each trajectory of (31), (32) starting from the regionjij Si does not leave the region jij S 0i during thistime interval if is sufficiently small.

Lemma 8: For each set of numbers Si > 0; i 0; . . . ; n,there exist such numbers Si > Si; ki > 0 and T >0; "M 0 that for any t 2 "; "; " "M, any trajec-tory of (31), (32) starting from the region jij Sienters within the time T, and without leaving the regionjij Si, the region jij ki"ni1 and stays thereforever.

Mark that _0 plays rule of the disturbance for the(n 1)th-order system (32). With n 0 (31), (32) isreduced to _0 2 0 sign 0 t L;L, andLemmas 7 and 8 are obviously true with n 0; 0 > L.The lemmas are proved by induction. Let their state-ments be true for the system of the order n 1 withsome choice of parameters

i, i 0; . . . ; n 1. Provetheir statements for the nth order system (31), (32)with sufficiently large 0, and i

i1; i 1; . . . ; n.

Proof of Lemma 7: Choose some SMi;Si < S0i < SMi,

i 0; . . . ; n. Then


j _0j 0ktj j0kn=n1 j1j

0jtjn=n1 0Sn=n1M0 SM1

Thus, according to the Holder inequality

j _0jdt 01=n1

jtdt

n=n10S

n=n1M0 SM1

Hence, j0j S 00 with small . On the other hand _0serves as the input disturbance for the (n 1)th ordersystem (32) and satisfies the conditions of Lemma 7,thus due to the induction assumption jij S 0i ;i 1; . . . ; n with small . &

Lemma 9: If for some Si < S0i ; i 0; . . . ; n; 0 and

T > 0 any trajectory of (31), (32) starting from the re-gion jij S 0i enters within the time T the regionjij Si and stays there forever, then the system (31),(32) is finite-time stable with 0.

Lemma 9 is a simple consequence of the invarianceof (31), (32) with respect to transformation (33). Theconvergence time is estimated as a sum of a geometricseries.

Proof of Lemma 8: Consider first the case "M 0; i.e. 0. Choose some larger region jij S 0i ;Si < S 0i ,i 0; . . . ; n, and let Si > S 0i ; i 0; . . . ; n 1, be someupper bounds chosen with respect to Lemma 8 for (32)and that region. It is easy to check that for any q > 1with sufficiently large 0 the trajectory enters the re-gion j0j qS1=0n1=n in arbitrarily small time.During that time _0 does not change its sign. There-fore,

j _0jdt S 00 for sufficiently large 0. Thus, the

disturbance _0 entering subsystem (32) satisfies Lem-ma 7 and the inequalities jij S 0i are kept with i 1.As follows from (31), from that moment on j _0j 3S1is kept with a properly chosen q.

Differentiating (31) with 0 achieve

0 0j0j1=n1 _0 _1where according to (32) the inequality j _1j 41S1 S2holds. Thus, with sufficiently large 0 within arbitrarilysmall time the inequality j _0j 10 41S1S2qS1=01=n is established. Its right-hand side maybe made arbitrarily small with large 0. Thus, due toLemma 8 for (n 1)-order system (32) and to Lemma9, the statement of Lemma 8 is proved with "M 0.

Let now "M > 0. As follows from the continuousdependence of the solutions of a differential inclusionon the right-hand side (Filippov 1988) with sufficientlysmall "M all trajectories concentrate within a finite timein a small vicinity of the origin. The asymptotic featuresof that vicinity with "M changing follow now from thehomogeneity of the system. &

Theorems 5 and 6 are simple consequences ofLemma 8 and the homogeneity of the system. Toprove Theorem 7 it is sufficient to consider

_0 0j0tjjn=n1 sign 0t 1 34

instead of (31) with tj t < tj1 tj . The resultinghybrid system (34), (32) is invariant with respect to thetransformation

t; ; i7!t; ; ni1i 35On the other hand, it may be considered as system (31),(32), jj ", with arbitrarily small " for any fixed regionof the initial values when is sufficiently small. Applyingsuccessively Lemma 8 and the homogeneity reasoningwith transformation (35) achieve Theorem 7. &

Remark: It is easy to see that the above proof maybe transformed in order to obtain a constructive upperestimation of the convergence time. Also the accuracymay be estimated by means of the suggested inductiveapproach.

References

Atassi, A. N., and Khalil, H. K., 2000, Separation resultsfor the stabilization of nonlinear systems using differenthigh-gain observer designs. Systems and Control Letters,39, 183191.

Bartolini, G., Ferrara, A., Levant, A., and Usai, E.,1999 a, On second order sliding mode controllers. In K. D.Young and U. Ozguner (Eds) Variable Structure Systems,Sliding Mode and Nonlinear Control (Lecture Notes inControl and Information Science, 247) (London: Springer-Verlag), pp. 329350.

Bartolini, G., Ferrara, A., and Usai, E., 1998, Chatteringavoidance by second-order sliding mode control. IEEETransactions on Automatic Control, 43, 241246.

Bartolini, G., Ferrara, A., and Usai, E., 1999 b, On multi-input second order sliding mode control of nonlinear sys-tems with uncertainty. Proceedings of the 38th Conference onDecision and Control, Phoenix, AZ, USA, December 1999.

Bartolini, G., Pisano, A., and Usai, E., 2000, First andsecond derivative estimation by sliding mode technique.Journal of Signal Processing, 4, 167176.

Dabroom, A., and Khalil, H. K., 1997, Numerical differen-tiation using high-gain observers. Proceedings of the 37thConference on Decision and Control (CDC97), San Diego,CA, USA, December 1997, pp. 47904795.

Elmali, H., and Olgac, N., 1992, Robust output trackingcontrol of nonlinear MIMO systems via sliding mode tech-nique. Automatica, 28, 145151.

Emelyanov, S. V., Korovin, S. K., and Levantovsky, L.V., 1986, Higher order sliding regimes in the binary controlsystems. Soviet Physics, Doklady, 31, 291293.

Filippov, A. F., 1988, Differential Equations withDiscontinuous Right-Hand Side (Dordrecht, TheNetherlands: Kluwer).

Fridman, L. M., 1990, Singular extension of the definition ofdiscontinuous systems and stability. Differential Equations,10, 13071312.

940 A. Levant

Fridman, L., and Levant, A., 1996 , Sliding modes of higherorder as a natural phenomenon in control theory. In F.Garofalo and L. Glielmo (Eds) Robust Control viaVariable Structure and Lyapunov Techniques (LectureNotes in Control and Information Science, 217) (London:Springer-Verlag), pp. 107133.

Fridman, L., and Levant, A., 2002, Higher order slidingmodes. In W. Perruquetti and J. P. Barbot (Eds) SlidingMode Control in Engineering (New York: Marcel Dekker),pp. 53101.

Furuta, K., and Pan, Y., 2000, Variable structure controlwith sliding sector. Automatica, 36, 211228.

Golembo, B. Z., Emelyanov, S. V., Utkin, V. I., andShubladze, A. M., 1976, Application of piecewise-contin-uous dynamic systems to filtering problems, Automation andRemote Control, 37, Part 1, 369377.

Isidori, A., 1989, Nonlinear Control Systems, 2nd edn (NewYork: Springer Verlag).

Isidori, A., Teelc, A. R., and Praly, L., 2000, A note on theproblem of semiglobal practical stabilization of uncertainnonlinear systems via dynamic output feedback. Systemsand Control Letters, 39, 165171.

Krupp, D., Shkolnikov, I. A., and Shtessel, Y. B., 2001, 2-sliding mode control for nonlinear plants with parametricand dynamic uncertainties. Proceedings of AIAA Guidance,Navigation, and Control Conference, Denver, CO, USA,AIAA paper No. 20003965.

Kokotovic, P., and Arcak,M., 2001, Constructive nonlinearcontrol: a historical perspective. Automatica, 37, 637662.

Landau, I. D., Lozano, R., and MSaad, M., 1998, AdaptiveControl (London: Springer-Verlag).

Levant, A. (Levantovsky, L.V.), 1993, Sliding order and slid-ing accuracy in sliding mode control. International Journalof Control, 58, 12471263.

Levant, A., 1998 a, Robust exact differentiation via slidingmode technique. Automatica, 34, 379384.

Levant, A., 1998 b, Arbitrary-order sliding modes with finitetime convergence. Proceedings of the 6th IEEEMediterranean Conference on Control and Systems, June 911, Alghero, Sardinia, Italy.

Levant, A., 1999, Controlling output variables via higherorder sliding modes. Proceedings of the European ControlConference, September, Karlsruhe, Germany.

Levant, A., 2001 a, Universal SISO sliding-mode controllerswith finite-time convergence. IEEE Transactions onAutomatic Control, 46, 14471451.

Levant, A., 2001 b, Higher order sliding modes and arbitrary-order exact robust differentiation. Proceedings of theEuropean Control Conference, Porto, Portugal, pp. 9961001.

Levant, A., 2002, Universal SISO output-feedback controller.Proceedings of IFAC2002, Barcelona, Spain, July 2126.

Levantovsky, L. V., 1985, Second order sliding algorithms.Their realization. In Dynamics of Heterogeneous Systems(Moscow: Institute for System Studies), pp. 3243 (inRussian).

Man, Z., Paplinski, A. P., and Wu, H. R., 1994, A robustMIMO terminal sliding mode control for rigid roboticmanipulators. IEEE Transactions on Automatic Control,39, 24642468.

Murray, R., and Sastry, S., 1993, Nonholonomic motionplanning: steering using sinusoids. IEEE Transactions onAutomatic Control, 38, 700716.

Rios-BolIvar, M., Zinober, A. S. I., and Sira-RamIrez, H.,1997, Dynamical adaptive sliding mode output trackingcontrol of a class of nonlinear systems. InternationalJournal of Robust and Nonlinear Control, 7, 387405.

Rosier, L., 1992, Homogeneous Lyapunov function forhomogeneous continuous vector field. Systems ControlLetters, 19, 467473.

Sira-RamIrez, H., 1993, On the dynamical sliding mode con-trol of nonlinear systems. International Journal of Control,57, 10391061.

Slotine, J.-J. E., and Li, W., 1991, Applied Nonlinear Control(London: Prentice-Hall).

Spurgeon, S. K., and Lu, X. Y., 1997, Output tracking usingdynamic sliding mode techniques. International Journal ofRobust and Nonlinear Control, Special Issue on New Trendsin Sliding Mode Control, 7, 407427.

Su, W.-C., Drakunov, S. V., and O zgu ner, U ., 1994,Implementation of variable structure control for sampled-data systems. Proceedings of IEEE Workshop on RobustControl via Variable Structure and Lyapunov Techniques,Benevento, Italy, pp. 166173.

Utkin, V. I., 1992, Sliding Modes in Optimization and ControlProblems (New York: Springer Verlag).

Wu, Y., Yu, X., and Man, Z., 1998, Terminal sliding modecontrol design for uncertain dynamic systems. Systems andControl Letters, 34, 281287.

Yu, X., and Xu, J. X., 1996, Nonlinear derivative estimator.Electronic Letters, 32, 14451447.


first

Documents

Articulo Higher Order Sliding Modes, Differentiation and Output Feedback Control