Levant 1993
Levant 1993
Levant 1993
To cite this article: ARIE LEVANT (1993): Sliding order and sliding accuracy in sliding mode control, International Journal of
Control, 58:6, 1247-1263
To link to this article: http://dx.doi.org/10.1080/00207179308923053
INT.
J. CONTROL,
1993,
VOL.
ARIE LEVANTH
The synthesis of a control algorithm that stirs a nonlinear system to a given
manifold and keeps it within this constraint is considered. Usually, what is
called sliding mode is employed in such synthesis. This sliding mode is
characterized, in practice, by a high-frequency switching of the control. It turns
out that the deviation of the system from its prescribed constraints (sliding
accuracy) is proportional to the switching time delay. A new class of sliding
modes and algorithms is presented and the concept of sliding mode order is
introduced. These algorithms feature a bounded control continuously depending on time, with discontinuities only in the control derivative. It is also shown
that the sliding accuracy is proportional to the square of the switching time
delay.
1.
Introduction
Consider a smooth dynamic system described by
i
= f(t,
x, u)
(1)
Real sliding
We call every motion that takes place strictly on the constraint manifold
a = 0 an ideal sliding. We also informally call every motion in a small
neighbourhood of the manifold a real sliding (Utkin 1977, 1981). The common
sliding mode exists due to an infinite frequency of the control switching.
However, because of switching imperfections this frequency is finite. The sliding
Received 8 August 1991.
t Mechanical Engineering Department, Ben-Gurion University of the Negev, Beer Sheva,
Israel.
Formerly, L. V. Levantovsky.
0020-7179/93 $10.00
1248
A. Levant
]'
e1 ",; Cly(e)l'
t E [TI> T z].
In the particular case when y(e) is the smallest time interval of control
smoothness the words 'with respect to y' may be omitted.
o
for
Some examples are now in order. (1) High gain feedback systems (Young et
al. 1977, Saksena et al. 1984): such systems constitute real sliding algorithms of
the first order with respect to :', where k is a large gain. (2) Regular sliding
mode: this system constitutes an ideal sliding mode (theoretically) and also
constitutes real sliding of the first order, provided switching time delay is
accounted for.
Let (t,x(t, s) be a real sliding family with e->O, t belongs to a bounded
interval. Let a( t, x) be a smooth constraint function and r > 0 be the real
sliding order with respect to ,(e), where '(Ii) > 0 is the smallest time interval of
smoothness of the piecewise smooth function x(t, s).
Proposition 1: Let 1= [r] be the maximum integer number not exceeding r. If
the lth derivative a(l) = (d/dt)la(t, x(t, s) is uniformly bounded in e for the
steady-state part of x(t, s), then there exist positive constants CI> C z, ... , C1_ I ,
such that for the steady-state process the following inequalities hold
CZ,I-Z, ... ,
Proposition 2: Let {j > 0 and p > 0 be an integer, and let {s.} be a sequence
with e, -> 0, Assume that for every e, there exists a time interval on which the
steady-state process is smooth, and such that for everyone of these intervals the
Iial p ) II"" 15 on
1249
It follows from these propositions that in order to get the rth order of real
sliding with a discrete switching it is required to satisfy the equalities a = a =
. .. = aU-I) =
in the ideal sliding. It is obvious that in a regular variable
structure system only the first order of the real sliding may be achieved with
discrete switching.
To prove these propositions we need the following simple lemma.
Lemma 1: There is a constant T:> 0, such that for every real function w(r) E C'l
defined on an interval of length t: there exists an internal point t I on the interval
such that
Ideal sliding
Consider the differential equation
y = v(y)
where y E R'", v: R'" -> R" is a locally bounded measurable (Lebesgue) vector
function. This equation is understood in the Filippov's sense (Filippov 1960,
1963, 1985); that is, the equation is replaced by an equivalent differential
inclusion
YE
V(y)
In the particular case when the vector-field v is continuous almost everywhere, the set-valued function V(Yo) is the convex closure of the set of all
possible limits of v(yO') as YO'-> Yo, where {yO'} are continuity points of v. The
solution of the equation is defined as the absolutely continuous function y(t),
satisfying the differential inclusion almost everywhere.
Definition 4: Let r be a smooth manifold. The set T itself is called
first-order sliding point set. The second-order sliding point set is defined as
set of the points y E r, where V(y) lies in the tangential space Tyr to
manifold T at the point y.
the
the
the
0
Definition 5: It is said that there exists a first (or second) order sliding mode on
the manifold F in the vicinity of a first (second) order sliding point Yo if, in this
vicinity of the point Yo, the first (second) order sliding set is an integral set, i.e.
0
consists of the Filippov's sense solutions.
The above definitions are easily extended to include non-autonomous differential equations by introduction of the fictitious equation i = 1. A sliding
mode is considered to be stable if the corresponding integral sliding set is stable.
Regular sliding modes satisfy conditions that the set of possible velocities V
A. Levant
1250
does not lie in Tyr, but V(y) n Tyr*- 0, and there exists a state trajectory on r
with velocity vector lying in Tyr. Such modes are the main operation modes in
the variable structure systems (Emelyanov et ai. 1970, Utkin 1977,. 1981, 1983,
Itkis 1976, Ryan and Corless 1984, DeCarlo et ai. 1988) and according to the
above definitions they are of the first order. When a switching error is present
the trajectory leaves the manifold with a certain angle. On the other hand, in
the case of second-order sliding, all the possible velocities lie in the tangential
space ~.(n and even when a switching error is present, the state trajectory is
tangential to the manifold at the time of leaving.
In the case when the manifold T is given by the constraint w(x) = 0, the
second-order sliding set is given by equations
Downloaded by [McGill University Library] at 15:14 17 October 2012
= 0, "Iv E
V(y) w~v
=0
(1)
u=U(t,x,~)
(2)
g = 1J!(t, x,
(3)
~)
Equations (2) and (3) are called the first (second) order sliding
the constraint a = 0 if a stable sliding mode of the first (second)
manifold a = 0 is achieved, and with every initial condition (to, xo)
transformed to the sliding mode in a finite time.
0
Sliding algorithms, which are used in the variable structure systems, are of
the first order and characterized by a piecewise continuous function U and
1Ji = O. The second-order sliding algorithms (Levantovsky 1985, 1986, 1987,
Emelyanov et 01. 1986 a, b, c, 1990) are given by a continuous function U and a
bounded discontinuous function 1J!. As a result of this the sliding problem is
solved by means of a continuous control.
It follows from the definition that in second-order sliding with a = 0, the
system is described by the equation
i
= f(l,
x, ueq(t, x
(4)
where u eq is the equivalent control (Utkin 1977) that is evaluated from the
1251
equation
a
a;(t, x) +
a~(t,
x)f(t, x, u eq )
=0
4.
= ~ (.) + ~
at
ax
such that
()f(t, x, u)
= Lua(t, x) = a;(t, x) +
a~(t,
x)f(t, x, u)
(3) There are positive constants ao, K m , K M , uo, Uo < 1, such that if
la(t,x)1 < ao then
0<
aa K
; -,,;;
M
au
A. Levant
1252
= a(t,
x) + bi), x)u
where x E R", Under conventional assumptions the task of keeping the constraint a = 0 is reduced to the task stated above. A new control }.t and a
constraint function rJ> are to be defined in this case by the transformation
u = }.tk<P(x), rJ> = a(t, x)/<P(x), where
<P(x)
[x'Dx + h)l/2
k , h > 0 are constants, D is a non-negative definite symmetric matrix (Levantovsky 1985, Emelyanov et al. 1986 a, b, c).
In the simple case when
= A(t)x + b(t)u,
= (c(t),
.r )
~(t)
c, ~,
~, A,
..4.,
b,
Ii
={
-kawith Ikal,,;;; 1
forms a real sliding algorithm of the first order with respect to k -I when k
It is easy to show that a is also of the order of :',
The algorithm
=-
sign a
--> 00.
(5)
is the ordinary sliding algorithm on a, i.e. it is of the first order. If the values of
a are measured at discrete times to, t1> t2, ... , t; - t;_1 = tH > 0 we get the
first-order real sliding algorithm u(t) = - sign a(t;), where t;";;; t < 1;+1'
The AI'-algorithm (Emelyanov and Korovin 1981, Emelyanov 1984)
it
-u with lui> 1
{ _ o-sign a with lui,,;;; 1
(6)
constitutes, with CY--> 00, a first-order real sliding algorithm with respect to cy- I .
Under the above assumptions there exists a unique function ueq(t, x)
satisfying a( t, x, ueq(t, x)) = 0 in the linear region. The second-order sliding set
may therefore be given by a=O, u = ueq(t,x) and is not empty. The next
theorem is easy to prove.
Theorem 1: Assume conditions (1), (3), (4) are satisfied and let the control
algorithm be
it = 1/1(t, x, u)
(7)
1253
where 1jJ is a bounded measurable (Lebesgue) function. Assume also that for
every second-order sliding point M and for every neighbourhood V(M)
J.t(V(M)
where J.t is a Lebesgue measure. Then there is a second-order sliding mode on the
constraint a = 0 and the motion in this mode is described by (4)
x
Proof:
x=f(t,x,u)
u
e(t, x, u) C R
It follows from the conditions of the theorem that for every second-order sliding
point (t, x , u) the set e includes the interval [- ColKm, ColKml. So every
trajectory (t,x(t), u(t)) that satisfies (1) with il E [-CoIKm, ColKml is a solution
of (1) and (7) if it lies on the second-order sliding set a = a = O.
It follows from conditions (3) and (4) and from the implicit function
theorem, that lill "" ColKm. Therefore, we can choose
il
= ileq =
LI/u eq ( t, x)
The theorem follows now from the fact that the second-order sliding set a
u = u eq is invariant with respect to (1) and (8).
(8)
= 0,
0
-u with
il =
lui> 1
lui"" 1
oo > 0, lui"" 1
(9)
c.j; }
(10)
1254
A. Levant
Figure 1.
state path of the algorithm (9) converges to the manifold in a finite time and
makes an infinite number of rotations (Fig. 2). The result is achieved by the
switching of the parameter (Y in (9). The actual calculation of the derivative a
may cause a serious obstacle in applications. Let us use the first difference in (9)
instead of a itself. Suppose that the measurements of a are being made at times
10, II' 12, ... , I; - 1;-1 = 'l: > 0. Define
0,
i =
tia(I;) =
{ a(lj, X(I; - a(lj-b X(I;-I,
i "" 1
Figure 2.
1255
-u(t;),
lu(t;)1 > 1
lu(t;)I,;;; 1
lu(t;)I,;;; 1
(11)
Theorem 3: Under the conditions of Theorem 2 with ,--- 0 the algorithm (11) is
a second-order real sliding algorithm. There are positive constants ao, al such that
in the steady-state mode inequalities [o] ,;;; ao?, 101,;;; a(f hold.
It follows from Proposition 2 that this is the best possible accuracy which
may be achieved by means of a control with its derivative being switched.
Consider the so-called 'drift algorithm' (Emelyanov et at. 1986 c)
-u(t;),
li
lu(t;)1 > 1
lu(t;)l,;;; 1
(12)
0, lu(t;)I,;;; 1
Here, IXM > IXm > O. After substitution of ~a for 0, the first-order sliding mode
on the constraint 0 = 0 would be achieved. This implies a = constant. But, since
the artificial switching time delay is introduced by (12), we ensure a real sliding
on 0 with most of the time being spent in the set oo < O. The algorithm does
not force the trajectory to reach the linearity region Io] < ao but forces this
trajectory to stay there once it reaches this set. Inequalities like the ones
appearing in Theorem 3 are ensured within a finite time by the algorithm. The
accuracy of the real sliding on 0 = 0 increases with decreasing r; in fact, it is
proportional to r. Hence, the duration of the transient process is proportional to
t- 1 Such an algorithm does not satisfy the definition of a real sliding algorithm.
Let
t;+1 - t;
';+1
,(a(t;, i
= 0,
1,2, ...
(13)
'M
(14)
and let
,(S)
vISI P
'M'
= vlsl
'm
'M
vlsl
<
P
"" 'M
vlSI P <
,;;;
'm
'm
'm ---
The specific restrictions on 'm' IXm' IXM' V may be found in the paper by
Emelyanov et al. (1986 c). The characteristic forms of the transient process in
the cases of the twisting and drift algorithms are shown in Figs 3 and 4,
respectively. The drift algorithm has no overshoot if parameters are chosen
properly (Emelyanov et al. 1986 c).
A. Levant
1256
Figure 3.
-0.0013 L -
-::-=::-:-::--
"'-
05506
Figure 4.
J
1
where rr > 0, and the continuous function g(a) is smooth everywhere except on
a = o. It is assumed here that all the solutions of the equation a = g( a) vanish
in a finite time, and the function g'(a)g(a) is bounded. For example
g = -Asign alal Y where A> 0,05";; Y < 1, may be used.
Theorem 5: With initial conditions within the reduced linearity region and for cy
sufficiently large, the algorithm (15) constitutes a second-order sliding algorithm
on the constraint a = o.
1257
The substitution of fla(t;) - l:g(a(t;) for a - g(a) turns the algorithm (15)
into a real sliding algorithm. The order of real sliding depends on the function g
but does not exceed 2. In Fig. 5 the characteristic form of the transient process
is shown with g = -AlaI 2/ 3 sign a. Near the transfer point t* with t e: t*
a(t) = 1/27Alt - t*1 3
All the above examples of the sliding algorithms use the derivatives of a
calculated with respect to the system. The following is an example that does not
utilize this property (Emelyanov et al. 1990).
u =
Ul
ill
={
Uz
={
(16)
Uz
-u,
lui> 1
(17)
-a'signa, lui'" 1
-Alaolpsigna
lal > ao
- AI alP sign a
Ial '" ao
(18)
where a', A> 0, P E (0, 1), and the initial values UI(tO) are to be chosen from the
condition
lui
IUt(to)
p(AKmF> (KMa'
(19)
I
C o)(2KMF-
(20)
Theorem 5: Assume that conditions (19), (20) are satisfied and 0 < P'" 1/2.
Then the algorithm (16), (17), (18) is a second-order sliding algorithm.
It may be shown that with P = 1, a' and A/a' sufficiently large there is a stable
second-order sliding mode. In this case lal + lal tends to zero with exponential
upper and lower bounds (Levantovsky 1986). Under the conditions of the
-00004::---
---==-
-=-==04075
Figure 5.
1258
A. Levant
theorem, discrete measurements turn these algorithms into real sliding algorithms. We remark here that, with p = 1, no real sliding algorithm will be
obtained, since the rate of the convergence is only exponential. Such an
algorithm may be efficiently implemented because of its simplicity. The bound
on the convergence decrement may be assigned any desirable value.
Remarks:
(1) Utilizing (13) and (14), all the above-listed algorithms with discrete
variable measurements become stable with respect to small model disturbances
and measurement errors. It may be shown that the accuracy of such modified
algorithms without measurement errors is the same as the accuracy of the initial
algorithms.
(2) It may be shown that for all except for the last algorithms, smoothness of
in (1) and of a is not necessary. It is sufficient for f to be only a locally
Lipschitzian function and the same is true for the partial derivatives of a( t, x).
(3) In the case where X is a Banach space the conclusions of the above
theorems still hold provided the statements are changed as follows: a and 0
vanish in a finite time or lal :;;; Ct r, 101 :;;; Cz1: after a finite time.
5.
dt Z
Let C
= LuLua,
=~
Lua
LuLua
dt
~Lua, it
au
(21)
<
; :;;;
K :;;; K M
(22)
aE
[-Co, Co]
+ [K m, KM]it
(23)
The real trajectory starting at the axis 0 on the plane a, 0 lies between the
extreme solutions of (23).
.
Figure 6 shows the set of the solutions of (23) and the solution of (21) which
are associated with the twisting algorithm (9). An infinite number encircling the
origin occurs here within a finite time. The total time duration of the origin
encircling is proportional to the total variation of a, which is estimated by a
geometric series.
The trajectories of the drift algorithm (12) are shown in Fig. 7. A real sliding
along the axis 0 = 0 may be seen here. Due to the switching of 0 the motion
with ao<O prevails. The duration of any cycle M oM tM zM 3 is proportional to
1:, where 1: is the time interval between the measurements of a, and the drift
1259
Figure 6.
Figure 7.
r.
M oM3 is proportional to
So, the time that the drift takes to reach a = 0 is
proportional to r- 1 with r = constant. With the variable r (14) the convergence
time turns out to be bounded (Emelyanov et al. 1986 c).
In Fig. 8 the trajectories of the algorithm (15) are shown. On the line
0= g(a) a common sliding mode (of the first order) arises. Considering now the
existence of the sliding mode on the manifold 0 - g( a) = 0 we have
~
dt
(0 - g(a
= ij
=c-
g~(a)o = LuLua
g~(a)g(a)
g~g(a) + ~Lua,
au
- Kasign(o - g(a
A. Levant
1260
Figure 8.
The last term must dominate the first two terms but it can do so only when
g;,(a)g(a) is bounded. That puts a restriction on g(a).
Consider now the algorithm (16), (17) and (18). In this case
(j E
Mathematical modelling
We consider now the following example
(24)
= 3u
<P(x)
- cos30tsinu - u 2/4 }
[xi + x~ + x~ +
1]1(2
(25)
1261
Figure 9.
Let a = X3/<P(X) be the constraint function. It is easy to show here that all the
assumptions are satisfied.
Consider the algorithm (16), (17) and (18). Here (17) may be simplified
because of the inequality au > 0 which holds for the entire state space with
lui;;" 1. Let
Ul -
lalPsign a
f -u
Ul
= 1 -8:igna,
lui> 1
lui.;; 1
Here, Euler's method has been employed for the numerical solution with the
step of 10- 4 . Figures 10 and 11 depict a(t) and u(t) with p = 05. In Fig. 12 the
graphs !J.(t, u(t)) and !J.eq(t, x(t)) are shown, where !J. = !J.eq is found from the
03208
Figure 10.
A. Levant
1262
---::;-==-_---:
-10028~...L.
03208
Figure 11.
Figure 12.
1263
ACKNOWLEDGMENT
REFERENCES