discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/253785103
CITATIONS READS
878 247
1 AUTHOR:
Arie Levant
Tel Aviv University
108 PUBLICATIONS 6,169 CITATIONS
SEE PROFILE
Being a motion on a discontinuity set of a dynamic system, sliding mode is used to keep accurately a given constraint and
features theoretically-innite-frequency switching. Standard sliding modes provide for nite-time convergence, precise
keeping of the constraint and robustness with respect to internal and external disturbances. Yet the relative degree of the
constraint has to be 1 and a dangerous chattering eect is possible. Higher-order sliding modes preserve or generalize the
main properties of the standard sliding mode and remove the above restrictions. r-Sliding mode realization provides for
up to the rth order of sliding precision with respect to the sampling interval compared with the rst order of the standard
sliding mode. Such controllers require higher-order real-time derivatives of the outputs to be available. The lacking
information is achieved by means of proposed arbitrary-order robust exact dierentiators with nite-time convergence.
These dierentiators feature optimal asymptotics with respect to input noises and can be used for numerical dierentia-
tion as well. The resulting controllers provide for the full output-feedback real-time control of any output variable of an
uncertain dynamic system, if its relative degree is known and constant. The theoretical results are conrmed by computer
simulation.
International Journal of Control ISSN 00207179 print/ISSN 13665820 online # 2003 Taylor & Francis Ltd
http://www.tandf.co.uk/journals
DOI: 10.1080/0020717031000099029
Higher-order sliding modes 925
with fast actuators (Fridman 1990, Fridman and Levant The problem is solved by recently presented arbitrary-
1996, 2002). Stable HOSM leads in that case to sponta- order robust exact nite-time-convergent dierentiators
neous disappearance of the chattering eect. (Levant 1999, 2001 a,b). The proposed lth-order dier-
Asymptotically stable or unstable sliding modes of any entiator allows real-time robust exact dierentiation up
order are well known (Emelyanov et al. 1986, Elmali to the order l, provided the next l 1th input deriva-
and Olgac 1992, Fridman and Levant 1996). Dynamic tive is bounded. Its performance is proved to be asymp-
sliding modes (Sira-Ram rez 1993, Spurgeon and Lu totically optimal in the presence of small Lebesgue-
1997) produce asymptotically stable higher-order sliding measurable input noises. This paper is the rst regular
modes and are to be specially mentioned here. publication of these dierentiators and of the corre-
A family of nite-time convergent sliding mode con- sponding proofs. Their features allow broad implemen-
trollers is based on so-called terminal sliding modes tation in the non-linear feedback control theory due to
(Man et al. 1994, Wu et al. 1998). Having been indepen- the separation principle (Atassi and Khalil 2000) trivi-
dently developed, the rst version of these controllers is ality. They can be also successfully applied for numerical
close to the so-called 2-sliding algorithm with a pre- dierentiation.
scribed convergence law (Emelyanov et al. 1986, The rth-order sliding controller combined with the
Levant 1993). The latter version is intended actually to r 1th-order dierentiator produce an output-feed-
provide for arbitrary-order sliding mode with nite-time back universal controller for SISO processes with per-
convergence. Unfortunately, the resulting closed-loop manent relative degree (Levant 2002). Having been only
systems have unbounded right-hand sides, which pre- recently obtained, the corresponding results are only
vents the very implementation of the Filippov theory. briey described in this paper, for the author intends
Thus, such a mode cannot be considered as HOSM. The to devote a special paper to this subject. The features
corresponding control is formally bounded along each of the proposed universal controllers and dierentiators
transient trajectory, but takes on innite values in any are illustrated by computer simulation.
vicinity of the steady state. In order to avoid innite
control values all trajectories are to start from a pre-
scribed sector of the state space. The very denition
2. Preliminaries: higher-order sliding modes
and the existence of the solution require some special
study here. Let us recall rst that according to the denition by
Arbitrary-order sliding controllers with nite-time Filippov (1988) any discontinuous dierential equation
convergence were only recently demonstrated (Levant x_ vx, where x 2 Rn and v is a locally bounded
1998 b, 2001 a). The proofs of these results are for the measurable vector function, is replaced by an equivalent
rst time published in the present paper. These control- dierential inclusion x_ 2 Vx. In the simplest case,
lers provide for full output control of uncertain single- when v is continuous almost everywhere, Vx is the
inputsingle-output (SISO) weakly-minimum-phase convex closure of the set of all possible limits of vy
dynamic systems with a known constant relative degree as y ! x, while fyg are continuity points of v. Any sol-
r. The control inuence is a discontinuous function of ution of the equation is dened as an absolutely contin-
the output and its r 1 real-time-calculated successive uous function xt, satisfying the dierential inclusion
derivatives. The controller parameters may be chosen in almost everywhere.
advance, so that only one parameter is to be adjusted in Consider a smooth dynamic system x_ vxwith a
order to control any system with a given relative degree. smooth output function , and let the system be closed
No detailed mathematical model is needed. The systems by some possibly-dynamical discontinuous feedback
relative degree being articially increased, sliding con- (gure 1). Then, provided that successive total time deri-
trol of arbitrary smoothness order can be achieved, vatives ; _ ; . . . ; r1 are continuous functions of the
completely removing the chattering eect. Since many closed-system state space variables, and the r-sliding
mechanical systems have constant relative degree due, point set
actually, to the Newton law, the application area for
_ r1 0 1
these controllers is very wide.
Any implementation of the above controllers is non-empty and consists locally of Filippov trajec-
requires real-time robust estimation of the higher- tories, the motion on set (1) is called r-sliding mode
order total output derivatives. The popular high-gain (rth-order sliding mode, Levantovsky 1985, Levant
observers (Dabroom and Khalil 1997) would destroy 1993, Fridman and Levant 1996).
the exactness and nite-time-convergence features of The additional condition of the Filippov velocity set
the proposed controllers. The rst-order robust exact V containing more than one vector may be imposed in
dierentiator (Levant 1998 a) can be used here, but its order to exclude some trivial cases. It is natural to call
successive application is cumbersome and not eective. the sliding order r strict if r is discontinuous or does
926 A. Levant
cally in a vicinity of a given point and Lbe Lr1ae is not and the problem is easily solved by the standard relay
zero at the point (Isidori 1989). controller u sign , with a > C=Km . Here
In a simplied way the equality of the relative degree h _ ju0 t0 x0 a is globally bounded and
to r means that u rst appears explicitly only in the rth g @=@u_ x0 b. The rst-order real-sliding accuracy
total derivative of . In that case regularity condition (2) with respect to the sampling interval is ensured.
is satised (Isidori 1989) and @=@ur 6 0 at the given Let r 2. The following list includes only few most
point. The output satises an equation of the form known controllers. The so-called twisting controller
(Levantovsky 1985, Emelyanov et al. 1986, Levant
r ht; x gt; xu 4 1993) and the convergence conditions are given by
r
It is easy to check that g Lbe Lr1 ae @=@u ; u r1 sign r2 sign _ ; r1 > r2 > 0; 6
r
h Lae . Obviously, h is the rth total time derivative
of calculated with u 0. In other words, unknown r1 r2 Km C > r1 r2 KM C;
functions h and g may be dened using only inputout-
put relations. The heavy uncertainty of the problem pre- r1 r2 Km > C
vents immediate reduction of (3) to any standard form
by means of standard approaches based on the knowl- A particular case of the controller with prescribed con-
edge of a, b and . Nevertheless, the very existence of vergence law (Emelyanov et al. 1986, Levant 1993) is
standard form (4) is important here. given by
the establishment of the inequalities jj < 0 2 ; j_ j < gue-measurable bounded control. Then with properly
1 for some positive 0 ; 1 . chosen positive parameters 1 ; . . . ; r1 , the controller
All listed controllers may be used also with relative u sign
r1;r ; _ ; . . . ; r1 10
degree 1 in order to remove the chattering and improve leads to the establishment of an r-sliding mode 0
the sliding accuracy. Indeed, let u ; _ be one attracting each trajectory in nite time. The convergence
of controllers (6)(8), depending possibly on the pre- time is a locally bounded function of initial conditions.
vious measurements as in (8), then under certain natural
conditions (Levant 1993, Bartolini et al. 1999 a) the con- The proof of Theorem 3 is given in Appendix 1. It is
troller u sign can be replaced by the controller the rst publication of the proof. The assumption on the
u_ u with juj > 1; u_ ; _ with juj 1. solution extension possibility means in practice that the
Consider the general case. Let u be dened from the system be weakly minimum phase. The positive par-
equality ameters 1 ; . . . ; r1 are to be chosen suciently large
d d dt in the index order. They determine a controller family
P sign Pr1 ; applicable to all systems (3) of relative degree r satisfy-
dt r1 dt dt
ing (5) for some C, Km and KM . Parameter > 0 is to be
where Pr1
r1 1
r2 r1 is a chosen specically for any xed C, Km and KM . The
stable polynomial, i 2 R. In the case r 1, P0 1 proposed controller may be generalized in many ways.
achieve the standard 1-sliding mode. Let r > 1. The r- For example, coecients of Ni;r may be any positive
sliding mode exists here at the origin and is asymptoti- numbers, equation (10) can be smoothed (Levant 1999).
cally stable. There is also a 1-sliding mode on the mani- Certainly, the number of choices of i is innite.
fold Pr1 d=dt 0. Trajectories transfer in nite time Here are a few examples with i tested for r 4, p
into the 1-sliding mode on the manifold Pr1 d=dt 0 being the least common multiple of 1; 2; . . . ; r. The rst
and then exponentially converge to the r-sliding mode. is the relay controller, the second coincides with (7).
Dynamic-sliding-mode controllers (Sira-Ram rez 1993,
Spurgeon and Lu 1997) are based on such modes. 1: u sign ;
Unfortunately, due to the dependence on higher-order
derivatives of the control is not bounded here even 2: u sign _ jj1=2 sign
with small . Also the accuracy here is the same as of
the 1-sliding mode. 2j_ j3 j2 1=6 sign _ jj2=3 sign
3: u
4. Building an arbitrary-order sliding controller jj3 1=6 sign _ 0:5jj3=4 sign g
Let p be any positive number, p r. Denote
5: u sign 4 4 jj12 j_ j15 j
j20
N1;r jjr1=r
j
j30 1=60 sign
3 jj12 j_ j15
Ni;r jjp=r j_ jp=r1 ji1 jp=ri1 ri=p ;
j20 1=30 sign
j 2 jj12
i 1; . . . ; r 1
j_ j15 1=20 sign _ 1 jj4=5 sign
p=r p=r1 r2 p=2 1=p
Nr1;r jj _ j j j
Obviously, parameter is to be taken negative with
0;r @=@ur < 0. Controller (10) is certainly insensitive
to any disturbance which keeps the relative degree and
1;r _ 1 N1;r sign (5). No matching condition having been supposed, the
residual uncertainty reveals itself in the r-sliding motion
i;r i i Ni;r sign
i1;r ; i 1; . . . ; r 1 equations (in other words, in zero dynamics).
The idea of the controller is that a 1-sliding mode is
where 1 ; . . . ; r1 are positive numbers.
established on the smooth parts of the discontinuity set
G of (10) (gures 2, 3). That sliding mode is described by
Theorem 3 (Levant 1998 a, 2001): Let system (3) have the dierential equation
r1;r 0 providing in its turn
relative degree r with respect to the output function for the existence of a 1-sliding mode
r2;r 0. But the
and (5) be fullled. Suppose also that trajectories of sys- primary sliding mode disappears at the moment when
tem (3) are innitely extendible in time for any Lebes- the secondary one is to appear. The resulting movement
Higher-order sliding modes 929
z_1 v1 ; v1
1 jz1 v0 j1=2 sign z1 v0 w1 Consider the discrete-sampling case, when z0 tj f tj
is substituted for z0 f t with tj t < tj1 ; tj1 tj
w_ 1 1 sign z1 v0 ; z2 w 1 > 0.
Similarly, a 2nd-order dierentiator from each of
these sequences may be used as a base for a new Theorem 7: Let > 0 be the constant input sampling
recursive scheme. An innite number of dierentiator interval in the absence of noises. Then the following in-
schemes may be constructed in this way. The only equalities are established in nite time for some positive
requirement is that the resulting systems be homo- constants i , i depending exclusively on the parameters
geneous in a sense described further. While the author of the dierentiator
932 A. Levant
i
jzi f0 tj i ni1 ; i 0; . . . ; n z_0 0 jz0 f tjn=n1 sign z0 f t z1
jvi f0
i1
tj i ni ; i 0; . . . ; n 1 z_ i i jz0 f tjni=n1 sign z0 f t zi1 ;
convergence time can be made arbitrarily small. That Third-order dierentiator. The measurement step
allows for the local controller application. 103 was taken, noises are absent. The attained
With discrete measurements, in the absence of input accuracies are 5:8 1012 , 1:4 108 , 1:0 105 and
noises, the controller provides for the rth-order real slid- 0.0031 for the signal tracking, the rst, second and
ing sup jj r , where is the sampling interval. third derivatives respectively. The derivative tracking
Therefore, the dierentiator does not spoil the r-sliding deviations changed to 8:3 1016 , 1:8 1011 ,
asymptotics if the input noises are absent. It is also 1:2 107 and 0.00036 respectively after was reduced
proved that the resulting controller is robust and pro- to 104 . That corresponds to Theorem 7.
vides for the accuracy proportional to the maximal error Fifth-order dierentiator. The attained accuracies are
of the input measurement (the input noise magnitude). 1:1 1016 , 1:29 1012 , 7:87 1010 , 5:3 107 ,
Note once more that the proposed controller does not 2:0 104 and 0.014 for tracking the signal, the rst,
require detailed mathematical model of the process to be second, third, fourth and fth derivatives respectively
known. with 104 (gure 4(a)). There is no signicant
improvement with further reduction of . The author
7. Simulation examples wanted to demonstrate the 10th-order dierentiation,
7.1. Numeric dierentiation but found that dierentiation of the order exceeding 5
Following are equations of the 5th-order dierentia- is unlikely to be performed with the standard software.
tor with simulation-tested coecients for L 1 Further calculations are to be carried out with precision
higher than the standard long double precision (128 bits
z_0 v0 ; v0 12jz0 f tj5=6 sign z0 f t z1 17 per number).
6(b). The dierentiator performance within the rst 1.5 s straightforward and does not require reduction of the
is demonstrated in gure 6(c). The steering angle graph dynamic system to any specic form. If boundedness
(actual control) is presented in gure 6(d). The sliding restrictions (5) are globally satised, the control is also
accuracies jj 9:3 108 , j_ j 7:8 105 , global and the input is globally bounded. Otherwise the
4
j
j 6:6 10 , jj 0:43 were attained with the controller is still locally applicable.
sampling time 104 . In the uncertainty case a detailed mathematical
model of the process is not needed. Necessary time deri-
8. Conclusions and discussion of the obtained results vatives of the output can be obtained by means of the
proposed robust exact dierentiator with nite-time
Arbitrary-order real-time exact dierentiation
together with the arbitrary-order sliding controllers pro- convergence. The proposed dierentiator allows real-
vide for full SISO control based on the input measure- time robust exact dierentiation up to any given order
ments only, when the only information on the controlled l, provided the next l 1th derivative is bounded by a
uncertain weakly-minimum-phase process is actually its known constant. These features allow wide application
relative degree. of the dierentiator in non-linear control theory.
A family of r-sliding controllers with nite time con- Indeed, after nite time transient in the absence of
vergence is presented for any natural number r, provid- input noises its outputs can be considered as exact direct
ing for the full real-time control of the output variable if measurements of the derivatives. Therefore, the separa-
the relative degree r of the dynamic system is constant tion principle is trivially true for almost any feedback.
and known. Whereas 1- and 2-sliding modes were used At the same time, in the presence of measurement noises
mainly to keep auxiliary constraints, arbitrary-order the dierentiation accuracy inevitably deteriorates
sliding controllers may be considered as general-purpose rapidly with the growth of the dierentiation order
controllers. In case the mathematical model of the (Levant 1998 a), and direct observation of the deriva-
system is known and the full state is available, the tives is preferable. The exact derivative estimation does
real-time derivatives of the output variable are directly not require tending some parameters to innity or to
calculated, and the controller implementation is zero. Even when treating noisy signals, the dierentiator
performance only improves with the sampling step special case m 2 X; _; . . . ; m1 , 2 R, which is
reduction. considered in the present paper, the graph may be con-
The resulting controller provides for extremely high sidered as a set from Rm R. An inclusion
tracking accuracy in the absence of noises. The sliding m 2 X 0 ; _; . . . ; m1 corresponding to the closed
accuracy is proportional to r , being a sampling period "-vicinity of that graph is further called the "-swollen
and r being the relative degree. That is the best possible inclusion. It is easy to see that this is a Filippov inclu-
accuracy with discontinuous rth derivative of the output sion. An "-swollen dierential equation is the inclusion
(Levant 1993). It may be further improved increasing corresponding to the "-vicinity of the corresponding
the relative degree articially, which produces arbitrarily Filippov inclusion.
smooth control and removes also the chattering eect.
The proposed controllers are easily developed for Proof of Theorem 3: Consider the motion of a projec-
any relative degree, at the same time most of the practi- tion trajectory of (3) and (10) in coordinates ,
cally important problems in output control are covered _ ; . . . ; r1
by the cases when relative degree equals 2, 3, 4 and 5.
@ r
Indeed, according to the Newton law, the relative degree r Lrae t; x u t; x; u 29
of a spatial variable with respect to a force, being under- @u
stood as a control, is 2. Taking into account some Taking into account (5) achieve a dierential inclusion
dynamic actuators, achieve relative degree 3 or 4. If
the actuator input is required to be a continuous r 2 C; C Km ; KM u 30
Lipschitz function, the relative degree is articially
which will be considered from now on instead of the real
increased to 4 or 5. Recent results (Bartolini et al.
equality (29). The operations on sets are naturally
1999 b) seem to allow the implementation of the devel-
understood here as sets of operation results for all poss-
oped controllers for general multi-inputmulti-output
ible combinations of the operand set elements. Control u
systems.
is given by (10) or (11).
A given point P is called here a discontinuity point of
Appendix 1. Proofs of Theorems 3, 4 and Proposition 3 a given function, if for any point set N of zero measure
and any vicinity O of P there are at least two dierent
The general idea of the proofs is presented in } 4 and
limit values of the function when point p 2 O=N
is illustrated by gures 2 and 3.
approaches P. Let G be the closure of the discontinuity
set of sign
r1;r ; _ ; . . . ; r1 or, in other words, of
Preliminary notions: The following notions are
control (10).
needed to understand the proof. They are based on re-
sults by Filippov (1988).
Lemma 1: Set G partitions the whole space
Dierential inclusion _ 2 X, 2 Rm is called
; _ ; . . . ; r1 into two connected open components
further Filippov inclusion if for any :
satisfying
r1;r ; . . . ; r1 > 0 and
r1;r ; . . . ;
1. X is a closed non-empty convex set; r1 < 0 respectively. Any curve connecting points
2. X fv 2 Rm jkvk g, where is a con- from dierent components has a non-empty intersection
tinuous function; with G.
3. the maximal distance of the points of X 0 from
Proof: Consider any equation
i;r ; _ ; . . . ; i 0;
X tends to zero when 0 ! .
i 1; . . . ; r 1. It may be rewritten in the form
Recall that any solution of a dierential inclusion is
i i ; _ ; . . . ; i1
an absolutely continuous function satisfying the inclu-
sion almost everywhere, and that any dierential equa-
i Ni;r ; _ ; . . . ; i1 sign
i1;r ; _ ; . . . ; i1
tion with a discontinuous right-hand side is understood
as equivalent to some Filippov inclusion. Let Si be the closure of the discontinuity set of sign
The graph of a dierential inclusion _ 2 X, 2 Rm
i;r ; i 0; . . . ; r 1; G Sr1 , and S0 f0g R (gure
is the set f; _ 2 Rm Rm j_ 2 Xg. A dierential 7). Each set Si lies in the space ; _ ; . . . ; i and is, actu-
inclusion _ 2 X 0 is called "-close to the Filippov inclu- ally, a modication of the graph of
sion _ 2 X in some region if any point of the graph of i i ; _ ; . . . ; i1 .
_ 2 X 0 is distanced by not more than " from the graph The Lemma is proved by the induction principle.
of _ 2 X. It is known that within any compact region Obviously, S0 partitions R into two open connected
solutions of _ 2 X 0 tend to some solutions of _ 2 X components. Let Si1 divide the space Ri with coordi-
uniformly on any nite time interval with " ! 0. In the nates ; _ ; . . . ; i1 into two open connected compon-
Higher-order sliding modes 937
ents O
i1 and Oi1 with
i1;r > 0 and
i1;r < 0 respect- coordinates and as derivatives as well. It is also easy to
ively. It is easy to see (gure 7) that check that inclusion (30) and (10) is invariant with
respect to Gv ; v > 0. That invariance implies that any
Si f; _ ; . . . ; i j ji j statement invariant with respect to Gv is globally true if
it is true on some set E satisfying the condition
i Ni;r ; _ ; . . . ; i1 & ; _ ; . . . ; i1 2 Si1 _ [v0 Gv E Rr . That reasoning is called further homo-
geneity reasoning. For example, it is sucient to prove
i i Ni;r ; _ ; . . . ; i1 sign
i1;r ; _ ; . . . ; i1 the following lemma only for trajectories with initial con-
ditions close to the origin.
& ; _ ; . . . ; i1 2
= Si1 g
d
N_ r1;r jjp=r j_ jp=r1 jr2 jp=2 1=p
dt
1 d p=r p=r1 r2 p=2
jj j_ j j j
p dt
d j p=rj
t1 t0 c0 1 Nr1;r t0 ; t0 ; _ t0 ; . . . ; r2 t0
j j =jj p=r
j _
j p=r1
j r2 p=2 p1=p
dt
rj
ji j cj 1 Nr1;r t0 ; t0 ; _ t0 ; . . . ; r2 t0 j ;
j p=rj1 j1 p=r p=r1
p=r jj j j sign j=jj j_ j
j 1; 2; . . . ; r
r2 p=2 p1=p
j j where cj are some positive constants.
j p=rj1 j p=rjp1rj1=p j1
p=r jj j =j j j j=
Proof: Considering a small vicinity of the origin, we
j1 p=rj1rj1=p found that the rst inequality is true when the initial
j j
conditions belong to a set of the form Nr1;r const.
Thus, according to the homogeneity reasoning, it is
p=r jjj jprj=rj=jj jprj=rj jj1 j=jj1 j
true everywhere. Other inequalities are results of suc-
p=r j cessive integration and the homogeneity reasoning with
the same assumption Nr1;r const with t t0 . &
With j r 2 the equality jr1 j r1 jjp=r
The following lemma is a simple consequence of
j_ jp=r1 jr2 jp=2 1=p is to be used. &
Lemma 4. It is illustrated by gure 2.
Let Gi be the corresponding subset of G which is
projected into Si ; G Gr1 Sr1 . Denote Lemma 4a: With suciently large any trajectory of
the inclusion (30), (10) transfers in nite time into some
vicinity of G and stays there. That vicinity contracts to
i f; _ ; . . . ; r1 j jj j i Nj;r ; _ ; . . . ; j1 j
G with ! 1 uniformly within any compact area.
i 1; i 2; . . . ; r 1;
Let
i i Ni;r ; _ ; . . . ; i1 sign
i1;r ; _ ; . . . ; i1 ;
M f; _ ; . . . ; r1 jNr1;r ; _ ; . . . ; r1 g:
; _ ; . . . ; i1 2
= Si1 g;
Lemma 5: There is such 1 > 1 that with suciently
where i 1; . . . ; r 1; 0 contains only the origin large any trajectory of the inclusion (30), (10) starting
from M(1) never leaves M(1 ).
_ r1 0. It is easy to check that
[
i Lemma 5 is an obvious consequence of Lemmas 3a
Gi j ; i 1; . . . ; r 1 and 4a.
j0
Lemma 6: There are such 2 < 1 and T > 0 that with
suciently large any trajectory of inclusion (30), (10)
Lemma 3a: Let j be chosen as in Lemma 3. There is starting from M(1) enters M(2 ) within time T and stays
such a vicinity O of the origin _ i 0 in it.
and such " > 0 that each "-swollen dierential equa-
tion
i;r 0; i 1; . . . ; r 1, provides for nite-time Proof: As follows from Lemmas 3a, 4a and 5, there
attraction of all trajectories from O into the "-vicinity of are such successively embedded suciently small vici-
Gi1 . nities Oi Gi of Gi ; i 0; 1; . . . ; r 1; Gr1 G, that
with suciently large any trajectory of inclusion
Proof: Obviously, the "-vicinity of Gi divides the (30), (10) hits one of these vicinities in nite time and
space ; _ ; . . . ; i in two parts (Lemma 1). Also 1-slid- then transfers from one set to another in nite time ac-
ing mode on i1 i ; _ ; . . . ; i1 is ensured at cording to the diagram:
continuity points of i out of the "-vicinity of Gt1 : & OG Or1 Gr1 ! Or2 Gr2 ! ! O0 G0
O0 0 &
Lemma 4: With suciently large any trajectory of
the inclusion (30), (10) which leaves G at time t0 returns The invariance of inclusion (10), (30) with respect to
in nite time to G at some time t1 . The time t1 t0 and
G: t; ; _ ; . . . ; r1 7!t; r ; r1 _ ; . . . ; r1
the maximal coordinate deviations from the initial point
at the moment t0 during t1 t0 satisfy inequalities of implies that if M1 transfers in time T into M2 ,
the form M2 transfers in time 2 T into M22 , M22 transfers
Higher-order sliding modes 939
For any " > 0 the inequality C KM < "Nr1;r is G : t; i ; ; "7!t; ni1 i ; n1 ; n1 " 33
satised outside of some bounded vicinity of the origin.
Thus, similarly to the Theorem 3 proof, suciently Dene the main features of dierential inclusion (31),
small " being taken, nite-time convergence into some (32) which hold with a proper choice of the parameters
vicinity D of the origin is provided. Obviously, the trans-
i .
formation G: t; ; _ ; . . . ; r1 7!t; r ; r1 _ ; . . . ;
r1
transfers (30), (11) into the same inclusion, but
with the new measurement interval . Thus, G D is an
Lemma 7: Let t satisfy the condition that the inte-
gral j tjdt over a time interval is less than some
attracting set corresponding to that new value of the
xed K > 0. Then for any 0 < Si < Si0 ; i 0; . . . ; n,
measurement interval. &
each trajectory of (31), (32) starting from the region
ji j Si does not leave the region ji j Si0 during this
time interval if is suciently small.
Proof of Proposition 3: As is seen from the proof of
Theorems 3 and 4 the only restriction on the choice of
1 is formulated in Lemma 3, and is chosen with re- Lemma 8: For each set of numbers Si > 0; i 0; . . . ; n,
spect to Lemma 3 and is suciently enlarged after- there exist such numbers Si > Si ; ki > 0 and T >
wards. Thus, at rst the existence of the virtual sliding 0; "M 0 that for any t 2 "; "; " "M , any trajec-
modes
i;r 0, i 1; . . . ; r 1 is provided and than tory of (31), (32) starting from the region ji j Si
is taken so large that the real motion will take place in enters within the time T, and without leaving the region
arbitrarily small vicinity of these modes. Thus, the ji j Si , the region ji j ki "ni1 and stays there
parameters may be chosen suciently large in the or- forever.
der 1 ; . . . ; r1 , so that the convergence from any
xed compact region of initial conditions be arbitrarily Mark that _ 0 plays rule of the disturbance for the
fast, and at the same time the overregulation be arbi- (n 1)th-order system (32). With n 0 (31), (32) is
trarily small. & reduced to _ 0 2
0 sign 0 t L; L, and
Lemmas 7 and 8 are obviously true with n 0;
0 > L.
The lemmas are proved by induction. Let their state-
Appendix 2. Proofs of Theorems 57 ments be true for the system of the order n 1 with
some choice of parameters
i , i 0; . . . ; n 1. Prove
Consider for simplicity dierentiator (16). The proof
their statements for the nth order system (31), (32)
for dierentiator (15) is very similar. Introduce
with suciently large
0 , and
i
i1 ; i 1; . . . ; n.
functions 0 z0 f0 t, 1 z1 f_0 t; . . . ; n
n
zn f0 t, f t f0 t. Then any solution of (16)
satises the following dierential inclusion understood Proof of Lemma 7: Choose some SMi ; Si < Si0 < SMi ,
in the Filippov sense i 0; . . . ; n. Then
940 A. Levant
j_ 0 j
0 ktj j0 kn=n1 j1 j Theorems 5 and 6 are simple consequences of
Lemma 8 and the homogeneity of the system. To
n=n1
0 jtjn=n1
0 SM0 SM1 prove Theorem 7 it is sucient to consider
Thus, according to the Holder inequality _ 0
0 j0 tj jn=n1 sign 0 t 1 34
n=n1
n=n1 instead of (31) with tj t < tj1 tj . The resulting
j_ 0 jdt
0 1=n1 jtdt
0 SM0 SM1 hybrid system (34), (32) is invariant with respect to the
transformation
Hence, j0 j S00 with small . On the other hand _ 0
serves as the input disturbance for the (n 1)th order t; ; i 7!t; ; ni1 i 35
system (32) and satises the conditions of Lemma 7, On the other hand, it may be considered as system (31),
thus due to the induction assumption ji j Si0 ; (32), jj ", with arbitrarily small " for any xed region
i 1; . . . ; n with small . & of the initial values when is suciently small. Applying
successively Lemma 8 and the homogeneity reasoning
Lemma 9: If for some Si < Si0 ; i 0; . . . ; n; 0 and with transformation (35) achieve Theorem 7. &
T > 0 any trajectory of (31), (32) starting from the re-
gion ji j Si0 enters within the time T the region
ji j Si and stays there forever, then the system (31), Remark: It is easy to see that the above proof may
(32) is nite-time stable with 0. be transformed in order to obtain a constructive upper
estimation of the convergence time. Also the accuracy
Lemma 9 is a simple consequence of the invariance may be estimated by means of the suggested inductive
of (31), (32) with respect to transformation (33). The approach.
convergence time is estimated as a sum of a geometric
series.
References
Proof of Lemma 8: Consider rst the case "M 0; i.e. Atassi, A. N., and Khalil, H. K., 2000, Separation results
0. Choose some larger region ji j Si0 ; Si < Si0 , for the stabilization of nonlinear systems using dierent
high-gain observer designs. Systems and Control Letters,
i 0; . . . ; n, and let Si > Si0 ; i 0; . . . ; n 1, be some
39, 183191.
upper bounds chosen with respect to Lemma 8 for (32) Bartolini, G., Ferrara, A., Levant, A., and Usai, E.,
and that region. It is easy to check that for any q > 1 1999 a, On second order sliding mode controllers. In K. D.
with suciently large
0 the trajectory enters the re- Young and U. Ozguner (Eds) Variable Structure Systems,
gion j0 j qS1 =
0 n1=n in arbitrarily small time. Sliding Mode and Nonlinear Control (Lecture Notes in
Control and Information Science, 247) (London: Springer-
During that time _ 0 does not change its sign. There- Verlag), pp. 329350.
fore, j_ 0 jdt S00 for suciently large
0 . Thus, the Bartolini, G., Ferrara, A., and Usai, E., 1998, Chattering
disturbance _ 0 entering subsystem (32) satises Lem- avoidance by second-order sliding mode control. IEEE
ma 7 and the inequalities ji j Si0 are kept with i 1. Transactions on Automatic Control, 43, 241246.
As follows from (31), from that moment on j_ 0 j 3 S1 Bartolini, G., Ferrara, A., and Usai, E., 1999 b, On multi-
input second order sliding mode control of nonlinear sys-
is kept with a properly chosen q.
tems with uncertainty. Proceedings of the 38th Conference on
Dierentiating (31) with 0 achieve Decision and Control, Phoenix, AZ, USA, December 1999.
Bartolini, G., Pisano, A., and Usai, E., 2000, First and
0
0 j0 j1=n1 _ 0 _ 1 second derivative estimation by sliding mode technique.
where according to (32) the inequality j_ 1 j 4
1 S1 S2 Journal of Signal Processing, 4, 167176.
Dabroom, A., and Khalil, H. K., 1997, Numerical dieren-
holds. Thus, with suciently large
0 within arbitrarily tiation using high-gain observers. Proceedings of the 37th
small time the inequality j_ 0 j
1
0 4
1 S1 Conference on Decision and Control (CDC97), San Diego,
1=n
S2 qS1 =
0 is established. Its right-hand side may CA, USA, December 1997, pp. 47904795.
be made arbitrarily small with large
0 . Thus, due to Elmali, H., and Olgac, N., 1992, Robust output tracking
Lemma 8 for (n 1)-order system (32) and to Lemma control of nonlinear MIMO systems via sliding mode tech-
nique. Automatica, 28, 145151.
9, the statement of Lemma 8 is proved with "M 0. Emelyanov, S. V., Korovin, S. K., and Levantovsky, L.
Let now "M > 0. As follows from the continuous V., 1986, Higher order sliding regimes in the binary control
dependence of the solutions of a dierential inclusion systems. Soviet Physics, Doklady, 31, 291293.
on the right-hand side (Filippov 1988) with suciently Filippov, A. F., 1988, Dierential Equations with
small "M all trajectories concentrate within a nite time Discontinuous Right-Hand Side (Dordrecht, The
Netherlands: Kluwer).
in a small vicinity of the origin. The asymptotic features Fridman, L. M., 1990, Singular extension of the denition of
of that vicinity with "M changing follow now from the discontinuous systems and stability. Dierential Equations,
homogeneity of the system. & 10, 13071312.
Higher-order sliding modes 941
Fridman, L., and Levant, A., 1996 , Sliding modes of higher Levant, A., 2001 b, Higher order sliding modes and arbitrary-
order as a natural phenomenon in control theory. In F. order exact robust dierentiation. Proceedings of the
Garofalo and L. Glielmo (Eds) Robust Control via European Control Conference, Porto, Portugal, pp. 996
Variable Structure and Lyapunov Techniques (Lecture 1001.
Notes in Control and Information Science, 217) (London: Levant, A., 2002, Universal SISO output-feedback controller.
Springer-Verlag), pp. 107133. Proceedings of IFAC2002, Barcelona, Spain, July 2126.
Fridman, L., and Levant, A., 2002, Higher order sliding Levantovsky, L. V., 1985, Second order sliding algorithms.
modes. In W. Perruquetti and J. P. Barbot (Eds) Sliding Their realization. In Dynamics of Heterogeneous Systems
Mode Control in Engineering (New York: Marcel Dekker), (Moscow: Institute for System Studies), pp. 3243 (in
pp. 53101. Russian).
Furuta, K., and Pan, Y., 2000, Variable structure control Man, Z., Paplinski, A. P., and Wu, H. R., 1994, A robust
with sliding sector. Automatica, 36, 211228. MIMO terminal sliding mode control for rigid robotic
Golembo, B. Z., Emelyanov, S. V., Utkin, V. I., and manipulators. IEEE Transactions on Automatic Control,
Shubladze, A. M., 1976, Application of piecewise-contin- 39, 24642468.
uous dynamic systems to ltering problems, Automation and Murray, R., and Sastry, S., 1993, Nonholonomic motion
Remote Control, 37, Part 1, 369377. planning: steering using sinusoids. IEEE Transactions on
Isidori, A., 1989, Nonlinear Control Systems, 2nd edn (New Automatic Control, 38, 700716.
York: Springer Verlag). Rios-Bol Ivar, M., Zinober, A. S. I., and Sira-Ram Irez, H.,
Isidori, A., Teelc, A. R., and Praly, L., 2000, A note on the 1997, Dynamical adaptive sliding mode output tracking
problem of semiglobal practical stabilization of uncertain control of a class of nonlinear systems. International
nonlinear systems via dynamic output feedback. Systems Journal of Robust and Nonlinear Control, 7, 387405.
and Control Letters, 39, 165171.
Rosier, L., 1992, Homogeneous Lyapunov function for
Krupp, D., Shkolnikov, I. A., and Shtessel, Y. B., 2001, 2-
homogeneous continuous vector eld. Systems Control
sliding mode control for nonlinear plants with parametric
Letters, 19, 467473.
and dynamic uncertainties. Proceedings of AIAA Guidance,
Sira-Ram Irez, H., 1993, On the dynamical sliding mode con-
Navigation, and Control Conference, Denver, CO, USA,
AIAA paper No. 20003965. trol of nonlinear systems. International Journal of Control,
Kokotovic, P., and Arcak, M., 2001, Constructive nonlinear 57, 10391061.
control: a historical perspective. Automatica, 37, 637662. Slotine, J.-J. E., and Li, W., 1991, Applied Nonlinear Control
Landau, I. D., Lozano, R., and MSaad, M., 1998, Adaptive (London: Prentice-Hall).
Control (London: Springer-Verlag). Spurgeon, S. K., and Lu, X. Y., 1997, Output tracking using
Levant, A. (Levantovsky, L.V.), 1993, Sliding order and slid- dynamic sliding mode techniques. International Journal of
ing accuracy in sliding mode control. International Journal Robust and Nonlinear Control, Special Issue on New Trends
of Control, 58, 12471263. in Sliding Mode Control, 7, 407427.
Levant, A., 1998 a, Robust exact dierentiation via sliding Su, W.-C., Drakunov, S. V., and O zgu ner, U ., 1994,
mode technique. Automatica, 34, 379384. Implementation of variable structure control for sampled-
Levant, A., 1998 b, Arbitrary-order sliding modes with nite data systems. Proceedings of IEEE Workshop on Robust
time convergence. Proceedings of the 6th IEEE Control via Variable Structure and Lyapunov Techniques,
Mediterranean Conference on Control and Systems, June 9 Benevento, Italy, pp. 166173.
11, Alghero, Sardinia, Italy. Utkin, V. I., 1992, Sliding Modes in Optimization and Control
Levant, A., 1999, Controlling output variables via higher Problems (New York: Springer Verlag).
order sliding modes. Proceedings of the European Control Wu, Y., Yu, X., and Man, Z., 1998, Terminal sliding mode
Conference, September, Karlsruhe, Germany. control design for uncertain dynamic systems. Systems and
Levant, A., 2001 a, Universal SISO sliding-mode controllers Control Letters, 34, 281287.
with nite-time convergence. IEEE Transactions on Yu, X., and Xu, J. X., 1996, Nonlinear derivative estimator.
Automatic Control, 46, 14471451. Electronic Letters, 32, 14451447.