A Generalized Normal Form and its Application to Sliding Mode Control
Hakan Fortell
Department of Electrical Engineering Linkoping University
S-581 83 Linkoping, Sweden Phone: +46 13 284028 E-mail: hafor@isy.liu.se
April 4, 1995
Abstract
In this report it is shown how a normal form, corresponding to that of ane state space systems, can be calculated for generalized state space descriptions. The calculations are performed by using Grobner bases.
This normal form is then used in the context of sliding mode control.
The result generalizes the previous results on sliding mode control for generalized state space representations.
Keywords: Normal Forms, Sliding Mode Control, Grobner Bases
1 Introduction
In 6] it is shown how a normal form can be calculated for generalized state space descriptions 4]. A thing to investigate is then if this normal form can be used in similar ways as the normal form of ane state space systems 7].
One such application is in the context of sliding mode control, see ,e.g., 10].
Constructing sliding mode controllers for generalized state space representations
has already been considered in 12] where the generalized controller canonical
form as described in 2] is used. The problem with this approach is that one
is restricted to zeroing a quite restricted class of elements. The order of the
di erential equation that must be solved to get the sliding control law is also of
maximal order of the input. The purpose of this report is to try to nd sliding
mode control laws for a larger class of manifolds and to see if one can nd
di erential equation for the input which is of an order lower than the maximal
one. The paper is organized as follows. The next section contains some of the
most basic concepts from commutative and di erential algebra. Section 3 then
shows how a generalized normal form can be computed and in section 4 this is
applied to sliding mode control. Finally in section 5 we give some concluding
remarks.
2 Basic Algebraic Concepts
In this section we present some of the most basic concepts from commutative and di erential algebra that will be used in later sections. The purpose is basically to display the notation.
2.1 Commutative Algebra
The theory in this section is collected from 1] and 5] where all proofs that are left out and also further references to commutative algebra can be found. We begin by making a few necessary denitions.
Denition 2.1
A commutative ring consists of a set
Rand two binary operations \ " and \+"
dened on
Rand such that:
1. (
a+
b) +
c=
a+ (
b+
c) and (
a b)
c=
a(
b c) for all
abc2R. 2.
a+
b=
b+
aand
a b=
b afor all
ab2R.
3.
a(
b+
c) =
a b+
a cfor all
abc2R.
4. There are 0
1
2Rsuch that
a+ 0 =
a1 =
afor all
a2R. 5. Given
a2Rthere is
b2Rsuch that
a+
b= 0.
We also need the following.
Denition 2.2
A eld
kis a commutative ring such that given
a2kthere is
c 2ksuch that
a c
= 1.
The commutative ring that we will mostly work with here is
kx1:::xn] which denotes the set of all polynomials in the variables
x1:::xnwith coecients taken from the eld
k. Usually we will assume that
kis the real numbers.
Denition 2.3
A subset
Iof a commutative ring
Ris an ideal if it satises 1. 0
2I.
2. If
ab2Ithen
a+
b2I. 3. If
a2Iand
b2Rthen
b a2I.
If we now set
f1:::fnto be polynomials in a ring
kx1:::xn] and introduce the notation
hf
1
:::f
ni
=
(
n X
i=1
pifi
:
pi2kx1:::xn]
i= 1
:::n)
(1) we can note that
hf1:::fniis an ideal and it is called the ideal generated by
f
1
:::f
n
.
A property that will be needed in the next section is to have some general- ization of linear dependence of polynomials. To nd such a generalization we
rst need some more denitions.
Denition 2.4
If the eld
kis a subset of the eld
Kthen
Kis said to be an extension eld of
k
. This is denoted by
K =k.
Denition 2.5
Let
Kbe an extension eld of
kand suppose that
K. The smallest eld containing
kis then called the eld obtained by adjoining
to
kand is denoted
k(
). Furthermore,
is said to be algebraic over
kif there exists a nonzero polynomial
f 2kx] such that
f(
) = 0. If
is not algebraic it is said to be transcendental over
k.
A denition can also be made in a similar manner for several elements in an extension eld.
Denition 2.6
Let
f1:::fmbe elements in an extension eld
Kof
k. The
fiare said to be algebraically dependent over
kif there exists a nonzero
P 2 kx1:::xm] such that
P(
f1:::fm) = 0. Otherwise the
fiare said to be algebraically independent.
Since algebraic dependence can be viewed as a generalization of linear depen- dence in linear algebra we can go on to dene a basis for these extension elds.
Denition 2.7
Let
Kbe an extension eld of
kand let
Ube a subset of
Ksuch that all elements of
Uare algebraically independent. The maximum number of elements that such a
Ucan contain is called the transcendence degree of
K =kand is denoted by trdeg
K =k. The elements of such a
Uis called the transcendence basis for
K =k. An important consequence of this denition is then the theorem below.
Theorem 2.1
If
kis a eld and the variables
xi i= 1
:::nare algebraically independent over
kthen trdeg
k(
x1:::xn)
=k=
n.
Proof.
We only have to note that the
xi i= 1
:::nis a transcendence basis for
k
(
x1:::xn)
=k.
2The following corollary will also be useful.
Corollary 2.1 The elements
y1:::yN 2 k(
x1:::xn) are algebraically de- pendent if
N >n.
Proof.
Otherwise
y1:::yNwould be a transcendence basis for
k(
x1:::xn)
=k.
2
Since algebraic dependence will be an important concept later on we give a theorem related to this subject.
Theorem 2.2
Suppose
kis a eld of characteristic zero and that
f1:::fN 2 k(
x1:::xn).
Then if
N>n
,
f1:::fNare algebraically dependent.
N
=
n,
f1:::fNare algebraically dependent i the Jacobian matrix
J
(
f) =
@fi
@xj
ij
(2)
is identically singular.
N <n
,
f1:::fNare algebraically independent i the Jacobian matrix (2) has rank
N.
2.2 Grobner bases
The reason why we introduce the concept of Grobner bases is that we wish to have some way of eliminating variables from a set of polynomial equation. To be able to choose which variables to eliminate we must introduce an ordering of variables.
Denition 2.8
Let
x2kx1:::xn] with
=
f1:::ng2Nndenote
x11 xnn. A term ordering
<is an ordering on
Nnsuch that for all
2Nn0
<and
< )+
<+
(3) There are of course several ways of dening these term orderings but we will only consider one of them.
Denition 2.9
The lexicographic term ordering is dened by
< () 9j
:
j < j 8i<j:
i< i(4) With a term ordering will now the di erent terms in a polynomial be ordered.
In particular will one of them be ordered rst.
Denition 2.10
If
f=
Pcx 2kx1:::xn] then the degree of
fis dened as
deg
f= max
f:
c6= 0
g(5)
The leading term of
fis
LT
f=
cdeg fxdeg
f(6)
and correspondingly the leading monomial of
fis LM
f=
xdeg
fand the leading coecient is LC
f=
cdeg f.
This denition can now be extended to cover ideals.
Denition 2.11
Let
Ibe an ideal in
kx1:::xn]. Then
LM
I=
hLM
f:
f 2Iideg
I=
fdeg
f:
2Ig(7)
Using this last denition we can now go on to dene Grobner bases.
Denition 2.12
A set
G Ifor an ideal
Iis a Grobner base with respect to a given term ordering i LM
I=
hLM
Gi.
At a rst glance it not at all clear how the concept of Grobner bases can be useful but it can be shown to have several nice properties. One of these properties is that
hGi
=
I(8)
which says that the set of solutions to the generating set of
Iis the same for the equations described by the Grobner base. To be able to describe the algorithm we need to introduce a few more concepts.
Denition 2.13
Let
fg2kx1:::xn].
fis said to be reduced w.r.t.
gif there is no term in
fthat is divisible by the leading term of
g. A subset of
kx1:::xn] in which all elements are reduced w.r.t. each other is said to be auto-reduced
If some polynomials are not reduced we can make them reduced.
Theorem 2.3
Let
F=
ff1:::fmgkx1:::xn] and
p2kx1:::xn]. Then it is possible to nd polynomials
g1:::gmr2kx1:::xn] such that
r
=
p;Xmi=1
gifi
(9)
and
ris reduced w.r.t. all the
fi.
The denition below is related to this theorem.
Denition 2.14
If all
rwhich satises the conditions of Theorem 2.3 are equal then
preduces to
rmodulo
Fwhich we write
p;!F r.
The algorithm for calculating Grobner bases uses the so called S-polynomials.
Denition 2.15
Let
f1f22kx1:::xn]. Then the S-polynomial of
f1and
f2is
S
(
f1f2) =
h1f1;h2f2(10) where
h
1
= LC
f2lcm(LM
f1LM
f2)
LM
f1 h2= LC
f1lcm(LM
f1LM
f2)
LM
f2(11)
and lcm(
) denotes the least common monomial.
The Grobner base algorithm can now be given as follows given some set of
polynomials
F.
Algorithm 2.1
1. Make
Fauto-reduced.
2. Add to
Fthe S-polynomial of two elements in
F.
3. If all S-polynomials reduce to zero modulo
Fthen
Fis a Grobner base.
Otherwise return to step 1.
The following theorem ensures that the result of this algorithm is a Grobner base.
Theorem 2.4
The set
G=
fg1:::gmgis a Grobner base for
hGii
S
(
gigj)
;!G0 (12)
A property of Grobner bases that we have mentioned earlier is that we can eliminate variables from a set of polynomial equations. How this elimination is performed is given in the following theorem.
Theorem 2.5
Let
Ibe an ideal in
kx1:::xn] and partition
x1:::xninto two disjoint sets
A
and
B. Then if
Gis a Grobner base for
Iwith a lexicographic term ordering where
A < Bthen
kA]
\Gis a Grobner base for
kA]
\Iw.r.t. the term ordering given by
<on
kA].
What this theorem says is that with a proper choice of term ordering we can
nd out if there are elements in
Iwhich are polynomial in the variables
Aonly.
If this is the case we also get a generating set for the corresponding ideal. We give a simple example.
Example 2.1
Consider the following set if polynomial equations
x
2
+
y+
z= 1
x
+
y2+
z= 1
x
+
y+
z2= 1 (13)
Calculating a Grobner base for the polynomials
x
2
+
y+
z;1
x+
y2+
z;1
x+
y+
z2;1
(14) with the term ordering
x>y>zgives the result
x
+
y+
z2;1
y2+
z;y;z22
yz2+
z4;z2 z6;4
z4+ 4
z3;z2(15) From this it is clear that
z 6
;
4
z4+ 4
z3;z22hx2+
y+
z;1
x+
y2+
z;1
x+
y+
z2;1
i(16)
and we have eliminated
xand
yfrom the set of equations (13).
2This example indicates that Grobner bases can be interpreted as a generalization
of Gaussian elimination. The algorithm for calculating Grobner bases has been
implemented in every major computer algebra language such as
Mapleand
Mathematica.
2.3 Dierential algebra
The theory presented in this section is gathered from 4] and the proofs can be found either there or in the books 9, 8].
The main di erence between di erential and commutative algebra is that we add di erentiation w.r.t. time,
@=@t, to the set of allowed operations.
Denition 2.16
A di erential eld
kis a eld satisfying the conditions
8 a2k
@a
@t
= _
a2k(17)
8 ab2k
@
@t
(
a+
b) = _
a+ _
b(18)
8 ab2k
@
@t
(
ab) = _
ab+
ab_ (19) Here and in the sequel will the \dot" notation be used instead of
@=@tto de- note di erentiation with respect to time. We also make a few more denitions regarding di erential elds which are the di erential variants of some of the commutative denitions in the previous section.
Denition 2.17
If the di erential eld
kis a subset of the di erential eld
Kthen
Kis said to be a di erential extension eld of
k. This is denoted by
K =k. If
Kis a di erential extension eld of
kand
Kthen the smallest di erential eld containing
k
is called the di erential eld obtained by adjoining
to
kand is denoted
khi
.
We now consider a denition of dynamics based on di erential algebra.
Denition 2.18
A dynamic system is a nitely generated di erential algebraic extension
D=khuiwhere
u=
fu1:::umgcan be viewed as inputs and the outputs
yis chosen as some nite set
y=
fy1:::ypgin
D.
Using this denition a more general form of state space description than (22) can be made.
Theorem 2.6
Let
K =khuibe a dynamic system according to denition 2.18. Then there exists a generalized state
x=
fx1:::xngfor some nite integer
nand polynomials
Aj
( ) and
Bi( ) such that
Aj
x
_
jxuu_
:::u(j)= 0
j= 1
:::n(20)
Bi
yixuu
_
:::u(i)= 0
i= 1
:::p(21) A theorem which is important when considering di erentially algebraic exten- sions is given below.
Theorem 2.7
Let
L=Kbe a nitely generated di erentially generated algebraic extension such
that
Kcontains non constant elements. Then there exists an element
2 Lsuch that
L=
Khi. The element
is called a di erentially primitive element.
3 A Generalized Normal Form
Here we will study how the normal form of ane state space descriptions can generalized to the more general dynamics of Denition 2.18.
3.1 The normal form of ane state space systems
The theory in this section is collected from 7]. Consider the SISO nonlinear system which is described by the following state equations
_
x
=
f(
x) +
g(
x)
uy
=
h(
x) (22)
where
x2Rn. The system (22) is said to have relative degree
rnif (
i)
LgLkfh(
x)
0
k= 0
:::r;2
(
ii)
LgLrf;1h(
x)
60 (23)
at some point
x0. Here
Lfand
Lgdenote the Lie derivatives in the directions
f
(
x) and
g(
x) respectively. Using the notation
dh
(
x) =
@h
(
x)
@x
1
@h
(
x)
@x
2
:::
@h
(
x)
@xn
(24) the following lemma can be shown to hold.
Lemma 3.1 The row vectors
dh
(
x0)
dLfh(
x0)
:::dLrf;1h(
x0) (25) are linearly independent.
We can now perform the change of variables locally around
x0where
1
(
x) =
h(
x)
2
(
x) =
Lfh(
x) ...
r
(
x) =
Lrf;1h(
x)
(26)
If
r<nthen
i(
x)
i=
r+1
:::ncan be found so that the change of variables becomes invertible. In particular these functions can be chosen so that
Lgi
(
x) = 0
i=
r+ 1
:::n(27) We see that nding
i(
x)
i=
r+ 1
:::nsuch that (27) is satised involves solving
n;rpartial di erential equations which is of course dicult in general.
However, if we do not bother with the condition (27) the invertible transforma-
tion is most often very easy to nd. With this transformation the new state
equations becomes _
z
1
=
z2_
z
2
=
z3_ ...
zr;1
=
zr_
zr
=
Lrfh(
;1(
z)) +
LgLrf;1h(
;1(
z))
u_
zr+1
=
qr+1(
z) +
pr+1(
z)
u_ ...
zn
=
qn(
z) +
pn(
z)
uy
=
z1(28)
This is known as the normal form of the system. We note that with the condition (27) we get
pi(
x)
0
i=
r+ 1
:::n.
3.2 A generalization of the normal form
In this section we repeat the results of 6] where all proofs that are left out and further constructive details can be found. We consider a dynamics
D=khuiwhich has di erential transcendence degree 1 and trdeg
D=khui=
n. In generalized state space form we thus get
A
1
;x
_
1xuu_
:::u(1)= 0 ...
An;x
_
nxuu_
:::u(n)= 0
B
;
yxuu
_
:::u()= 0
(29)
where we we for simplicity say that the system has only one output. Since we want to be able to have a unique description of the output we assume that the condition
@B
(
yx)
@y
6
= 0 (30)
is fullled. Our goal is then to try to nd a generalization of the normal form (28). The rst thing we must do is then to generalize the Lie derivative and the relative degree. The rst thing we note is that if the output polynomial explicitly depends on
uand its derivatives we have a case which do not correspond to the ane case. The natural thing is then to say that such a system has relative degree zero and that (29) is already in normal form. If this is not the case we must study the following polynomial
dB
(
yx)
dt
=
Xni=1
@B
(
yx)
@xi x
_
i+
@B(
yx)
@y
_
y
(31)
The correspondence to the Lie derivative (23) would then be if all _
xiand the
ycould be eliminated from (31). Obviously this can be done with Grobner bases
as given in the denition below.
Denition 3.1
Suppose that we have a system (29) where the output polynomial does not depend explicitly on
uand that
G1is a Grobner base for the polynomials
fA
1
( )
::: An( )
B(
yx)
dB(
yx)
=dtg(32) under the term ordering
xn< <x1<y
_
<u<u_
< <u(max)<y<x_
n < <x_
1(33) where
max= max
f1:::ng. The generalized Lie derivative of
yw.r.t.
A
=
fA1( )
::: An( )
g(34) is then dened as
LAB
(
yx) =
G1\kx1:::xnuu_
:::u(max)y_ ] (35) with
@LAB
(
yx)
@y
_
6= 0 (36)
The iterated generalized Lie derivative
LiAB(
yx) with
i>1 is dened as
LiAB
(
yx) =
Gi\kx1:::xnuu_
:::u(max)y(i)] (37) with
@LiAB
(
yx)
@y
(i) 6
= 0 (38)
Here
Giis the Grobner base of
A
1
( )
::: An( )
B(
yx)
LAB(
yx)
::: LiA;1B(
yx)
dLiA;1B(
yx)
=dtw.r.t. the term ordering (39)
xn< <x1<y(i)<u<u
_
< <u(max)< <y_
<y<x_
n< <x_
1(40) Here and in the sequel will
kdenote the eld of real numbers. With this deni- tion we can dene a relative degree for systems of the form (29).
Denition 3.2
A system (29) has generalized relative degree
r1 if
LiAB(
yx)
i= 1
:::rconsist of only one polynomial for each
iand it holds that
@LiAB
(
yx)
@u
(j)
0
8j0
i= 1
:::r;1 (41) and
@LrAB
(
yx)
@u
(j) 6
= 0
for some
j0 (42)
In order to justify the names generalized Lie derivative and generalized relative
degree we must now show that the usual denitions of these concepts follow as
special cases.
Theorem 3.1
Suppose that we have a system (22) with
f(
x),
g(
x) and
h(
x) polynomials in the ring
kx1:::xn]. Then this system has relative degree
ri it has generalized relative degree
r. The Lie derivatives
Lifh(
x)
i= 0
:::r;1 are given as the solutions to
LiA
(
y;h(
x)) = 0
i= 0
:::r;1 (43) w.r.t.
y(i) i= 1
:::r;1. Furthermore,
Lrfh(
x)+
LgLrf;1h(
x)
uis the solution to
LrA
(
y;h(
x)) = 0 (44)
w.r.t.
y(r).
The question is now how these Lie derivatives can be used for nding a normal form for our system (29). Given a relative degree
r1 we have a set of
polynomials
np
1
(
yx)
p2( _
yx)
::: pr(
y(r;1)x)
o(45) and in correspondence with the normal form (28) we want to perform a trans- formation such that
z
1
=
y z2= _
y::: zr=
y(r;1)(46) The
y(i) i= 0
:::r;1 can now be shown to have the following property.
Theorem 3.2
Suppose that the dynamic system (29) has generalized relative degree
r1.
Then
r nand the output derivatives
y(i) i= 0
:::r;1 are algebraically independent over
khui.
Remark 3.1 This theorem has a nice interpretation in the ane polynomial case namely that the proof of Lemma 3.1 follows directly from Theorem 3.2 and Theorem 2.2.
Since trdeg
D=khui=
nwe know that there exists
zi 2 D=khui i=
r+ 1
:::nsuch that with
zi=
y(i;1) i= 1
:::rthe set
fz1::: zngbecomes a transcendence basis for
D=khui. The question is how the
zr+1:::zncan be chosen. The lemma below gives a partial answer to this question.
Lemma 3.2 Suppose that the dynamic system (29) has generalized relative de- gree
rn. Then there exists
n;rdierent integers
ir+1:::inbetween 1 and
n
such that with
zj
=
y
(j;1) j
= 1
:::rxij j
=
r+ 1
:::n(47) the set
fz1::: zngbecomes a transcendence basis for
D=khui.
A constructive note that must be made now is that in Lemma 3.2 it is not shown how to choose the
xiwhich are algebraically independent over the eld
khui
(
z1:::zj)
j=
r:::r;1. These
xican be found with Grobner bases in a
way which is described in 6]. The main tool for performing the transformation
the system (29) will of course also be Grobner bases and the generalized normal
form becomes
z_
1=
z2_
z
2
=
z3_ ...
zr;1
=
zrqr;z
_
rzuu_
:::u(r)= 0
qr+1;z
_
r+1zuu_
:::u(r+1)= 0 ...
qn;z
_
nzuu_
:::u(n)= 0
y
=
z1(48)
Due to the fact that if we would have performed the calculations above for an ane system (22) we would get (28) we can call this a generalized normal form for (29).
4 Sliding Mode Control for Generalized State Space Descriptions
In 12] a method for designing sliding mode controllers for generalized state space descriptions (29) is given. However, this method has several drawbacks.
One is that one is restricted to zeroing the di erentially primitive element and another is that one must solve a high order di erential equation to nd the sliding control
u. We will now study if the generalized normal form dened in the previous section can be used for reducing these problems.
4.1 Sliding Mode Control using the Dierentially Primi- tive Element
Here we will repeat the results of 12]. First of all it is shown in 2] how Theorem 2.7 can be used for getting a more suitable form for the dynamics.
Theorem 4.1
Suppose that
is a primitive element of the dynamics
D=khui. Then
x1=
:::xn
=
(n)qualies a generalized state vector and the corresponding state space description becomes
_
x
1
=
x2_
x
2
=
x3_ ...
xn;1
=
xnC
;x
_
nx1:::xnuu_
:::u()= 0
(49)
This is known as the generalized controller canonical form.
What we want to achieve is to keep the primitive element
zero. As before we introduce an auxiliary output dened as