WA9 1O:OO -

The use of symbolic computation in nonlinear control: Is it viable?

Bram de Jager
Department of Mechanical Engineering,WH 2.137
Eindhoven University of Technology,P.O. Box 5 13,5600 MB &&oven, The Netherlands
Email:[email protected] Fax:+3140 447355

Abdract a solution of the problems, the algorithms used, and an investigation of the
To aid in the Mllysis and design of nonhear control systemsthe No&,-oN functionality needed. The g e n d structureof the imp1ementation in MAPLE Of
prlage,an raonym f o r m p l r m ( r o l h m b w d e v e @ e d . Thispaper thealgaithmsneededtosohretheseproblemsis~tedinSeclion5,wherethe
addressestheusefulngsoftheNO~ONpuckrgeu,aselfcontrinedtoolinthe NO&ON Package is described. Section 6 Presents some textbook e m P k
symbolic analysis and design of nonlinear control systems, u, a replacement using the NO&ON package, for each of the problem areas. The next section
d o r complement for numerical tools, and as a clubdtute for a sly tical contains a discusdon of the larger scale problem. Section 8 present%some
@pa and pencil) wak. The symbolic computation
- p
- r-o m is U& conclusions and gives directions for future research.
(u1a crmputing sub- for NON~XON. 2. The Control Roblemr
To -the viability of thistool,w d algorithmsavailablefor analysisand
From severa1areas in nonlinear control,where. symbolic computation is likely
design are implemented in N O d k 3 ~ and. applied to examples of textbook to be of some profit, we discuss the computation of the normal form, the
problems with published solutions and also to a larger scnle problem. The zero dynamics, and the input-output and state space exact lineariation. To
examples in the paper address the following symbolic computational arm
be able to solve some of these problems, modifications of the system to be
for nonlinear dynamic systems: (1) the normal form, (2) the zero dynamics, controlled arenecessary. Because, in our setup, the system itself isnot allowed
0 )the input-outputexaa linearizing statefeedback. (4) the state space. exact to be changed, the only possible modificationsare the judicious manipulation
linearizing statefeedback. including the eearch for output functions that give of control signals, is,signals that act on the system and can be influenced
the system a i l l or maximal *relative degree. 'Ihe systemsare not required
to have a well deiined relative degree. from the outside, and the manipulation of the information flowing out of the
system. 'he generation of control signals will be done by a control law, where
me textbook problems show that the N O ~ C O Npackage can be successfully information of the system is used to generate the control input.
used. 'Ihe larger scale problem is. however, too complex to be solved with the
21. Normplform
cumnt versions of N O I ~ C O N and W.Some recommendationsto improve
To reveal the structure inherent in the system, but disguised by the general
the implementation and to make more auitable for use with N&ON form of the mathematical model, it is necessary to uncover the sflucture.
are given. The conclusion is that symbolic computation is a viable approach This structure can then be used to advantage in the analysis and design of
for straightforward textbook problems, but not up enough to tackle control systems. It CM be revealed by transformation to standard, canonical
larger problems. However, we envision substantial hture progress. or, aa we call it here, normal forms. For this purpose a change of coordinates
is employed. The new coordinates are related to the system output and its
1. Inboduction derivatives. Depending on the relative degree ofthe system, Le., the number of
The use of symbolic computation programs for control purposes is investi- times the outputs have to be differentiated before the input explicitly appears,
gated by several " aSome control problems are handled by , "ER the system can be decomposed in a set of series connections of integrators,
~e [l]. The use of is discussed in [21 for a specific class of control a feedback over these sets, and, if the relative degree is less then the system
problema Zeitz et aL. 131. use the program MACNON, based on IMACSYMA, d e r , a remaining part that can be made "unobservable" at the output of the
to analyze obsevability and teachability. and to design observers and con- system by feedback, the ZBY) dynamics This will be made more precise in the
mllers for nonlinear sy-a Their package is mainly used for teaching, next section.
and there are no published results for larger scale problems. Blankemhip [4]
22 Zerodynamics
also uaed W ukuto solve some control problems with his implementation
CONDENS. but is now switching to h%umMAncA. and he provides a control T h e m dynamics of a nonlinear system can be characterizedasthe remaining
toolbox for this platform. Some M " M A T I C A notebooks, eg.COSXPAK, dynamics of the system if the output is required to be 0 for all times. It is the
are developed to mimic W tool boxes, primarily with the aim to get dynamics of the system an the largest unobservable submanifold that can be
a more powerful v i d z a t i a n and a possible integration of symbolic capa- obtained by judiciously manipulating the input to the system by s control law.
bilitieu, although those capabilities are not fully exploited now. The use of The characterization of a system by properties of its zero dynamics is of
W for a e v d control problems is reported in [51. Some problems re importance for some design goals, sg.,if the zero dynamics is unstable,
pated in this paper. eg.,with solving partial differentialequations,are parUy certain types of control laws are unable to stabilize the system, so these
reaolved in 161. They deaaibe a W package. here called NOX~WON(a control laws should be avoided.
mcceasor of the ZERODYN package presented and used in 15.71). that can 23. Input-wtput exact linearization
unnpute, e.&, the zero dynamics and provide solutions to exad linearization ?he input-output exact linearization problem is of long standing interest in
probha contro1theory. In essence,it is the problem of modifying a nonlinear dynamic
In the p m paper we illustrate the use of this package by applying it to system such that, after the modification, it behaves like a linear one, in the
m e textbook and a larger scale problem. Contrary to [7], that focusses on input-output sense, Le., the goal is to get a linear (dynamical) relation between
the camputation of the zao dynamics, and to [SI. that treats the state space the new input and the output of the plant.
a& liwarizacion problem, both with the assumption that the relative degree In a more complete control system design, the input-output exact linearization
is well &lined, here a more thorough treatment and a larger scale example is is o h only asubordinate goal, to make it possibleto use other design methods
presented, that more closely resanbles problems arising fmn practice. for attaining additional goals.
The main goals and contributionsof this paper are 24. State apace exact linearization
a proof of the viability of symbolic canputation for problems in the It is also possible to consider a more ambitious goal, where the behavior
analysis llnd design of nonlinear control systems be4ww the new input and the (transformed) state of the plant is required to
to show the Eharaderistica of a prototype implementation for the compu- be linear. 'Ihisexact (or state space exact) linearization problem is also of long
tation of normcll forms, the zeo dynamics, input-output and state space -ding in- in control theay. In essence, it is the problem of modifying
exact linerrization a nonlinear dynamic system such that, aftet the modification, it behaves like
to give some examplea (also a larger scale or semi-industrial one) of the a h e a r one, so powerful design methods for linear systems can again be
use of thin implementation and to document some applications employed.
to discuss directions for future research in symbolic computation and in
the analysis and synthesis for nonlinear control systems 3. The Mathematical Formulation of the Problems
to familiuize a lrrger audience in the control community with the use of In the presentation of the mathematics. we closely follow the WO& of
symboliccampumion. Isidori 191. We start with a nonlinear model of a plant and assume that it
cnn be described adequately by a set of nonlinear differential equations, a n e
The paper is as follows. F i section 2 presents and makea m e
m a r k s on the controlproblems that are treated in this study. Then, Section 3 in the input U , and without direct feed-through from input to output
discusses the marhematical details of the problems. Section 4 follows with =m+ gQu, y = W) (1)

0191-2216/93/$3.00 0 1993 IEEE 276

-~
n
~~~~ ~
 The nonllnear dynamics obtrined when the output y = he)is restrictedto 0
with &ate vector x E R",eont4intrg all necessay infomation of the pbnt,
m p t v e d o r u E Rn,andoutputvectory E P.Thenumberofinputsis
U@ to the number of outputs, La,the phut is quare. lhis aclaunption is for
amealmce only and mrLu a oimplified -on poodble PMS of the
-
i.e.v 0, is
e.
by mit&k initid conditions for i.e., m 0. and a suhrblc oontrol v in (S),

-
4 do.tl). tlm = V0.
theory can rbo be duivai if Ihe number of inputs is huger than the number
of ouipuk I b v e c t o r fieldf h "0th.g h u m columru gr of moothvedor It b invisible U the output, .nd is called the ZCID dyflunicsof the ay-.
fiddr. m d h is acokmn ofm rarlrr-valued moothfundomhi. because the dynamics is rehted to the zaos for lineu syyatem~, and at0
Thenonlinearsy-
atx-xOif
...
(1) is said to have a vector dative dclpra {rl, ,rm) beaume it is relazed to the zero output. Wben the involulivity conditions f a
MIMO ay-s rre not fulfilled. the zero dynamics will also depend on U.
1. Lg,f,,Sh) = O f o r k = 1 ,..., rj-2(i,j = 1 ,..., m)and.llxina W e a n now &ate our problems more formally.
The probkm of t"ing the ay- (1) to the m d h amounts
dghborhd o f f , to leaing up the trinstormrtion, compthg tfie i w a a ~tmnsfmation md
2 rhcfolbving m xmmurixisnonaingulnr at f qxning tht sy#fzIn in theaandml p l s and in thepcltc aslociated with the
zero dynamics. Of amme, this can be done if the rdadve degree is well
d&ed only. Ifthereis no reWie degree we h v e to reaorlto another method
to derive the equations f a the zero dynamics and to solve the input-output
aud linearizationprobkm (seethe following sections).
7he zem dyasdcs problem is to obtain the dynamics of the system when the
output y is required to be 0 f a all 1, by a propet choice of initial slate x(0) and
amtrolinput u(l). Herewe have to employ an appropriatestatic state-feedback
Here L,hi@) meana the ALh s u d v e Lie derivative of the scalar functim and we proper initial conditions More specific: we are looking for the locrrLly
hi@) in the direction of the vector fieldf, cg.&lie)= (alri@y&y@). The maximal output m h g submanifold, md its aesociated dynamics
matrixA i s m c t i m e s called the decouphgmatnx. Fm systems with a well defined relative de- the zero dynamics follows
Thc modification we allow is &tefdhck, la,a fe+dbrl bagd on the quite easily from the normal form. F a systems without a relative degree the
explicit knowledge of the value of the strte vector 41). The type of control situation is more complicated. "his problem has been treated in [9, Section
h w uwd hae is &c@d to &atefeedback. Heam, the value of the 6.11, and the solution does not requirethe s y m to haven well defined relative
input V e d a .(I) dcpcDdi aa the #&e &I)lad a new m&wm input v& Our aim is to i m p h e n t the dution algorithm and compute the zero
Nr). This dependence is of the fom dynamicaforrealsyrtrms.
'Ihe Inputsutput exacf linepIzaton problan: under which conditions is it
U = aQ + P O V (2) possible to transform the system (1) to a linear one by stabfeedback (2)?
The linearity property should be established beween the new input v and the
where the componcmta ai and &, are anooth functions output y. pormally, we are looking for a neighborhood U o f f and a static
For linear syaems a linear &go of diz = Tx with a nonsingular arcrto-feedbacksuch that f a all k 2 0 and all 1 S i,j < m the expression
matrix T is u a l Q adequate. For nonlinear sy-s it is more appropriate to k ~ # + ~ , h j @ ) is independent of x on U.
allow for anonlinear change of coordinates
For systems with a well defined relative degree, the input-output linearizing
z = (t,rl) = ow. feedback follows quite eSeily fium h e normal farm. For systems without a
0) relatiie degree the situation is more compliceted. This has also been solved,
I is r e q u i d that the kcobiana#/& of the trrsnafmation veuor CP in, at see, e.g..,[9, Section 5.41. Our goal is to test the conditions under which the
least locally, i m d b l e for CP to q d f y as a change of coordinates, beaus problem an be solved and to derive explicit expressions a and /3 for the
then Q, CM be used a two-way mapping. feedback (2).
When the sydem has a well detined dative &gree wecan use such a am- The &ate pace exlct Z n e d ~ h pn d h can be dated as: under which
dinate trursfarmatian to transform (I), under mme involutivity condition8 for moditions isitpc#sibletotMsfamthesrstem(l)toalineumdcon~~le
MIMOsyaemn, to t h e n a m a l h one by -feedback (2) and a c h i p of coordinated @)? The linearity
propetty sbwld bec&&liahed b e t w w tbenewinput vand the tmnsfmed
r*.terThispmblan haa becn solved. see, ag. [ 9 ] , ~ 1 d ow goal is toteat the
conditionsand to derive explicit equations f a the feedback and the change of
d i n a k a for specific planta The solution is only valid for systems with a
well ddned relative degree,and requires the dstenca of (synthetk) outputs
for which the sydan haa a ItU mla relative degree, r = n.
When a fullordanhtive degree canna beobtained. it k sanetimeconvenient
to Mive for a maximal relative d e p . 'Ihen the corresponding input-output
linearizingstate-feedback realizesa minimal dimension of the zeeo dynamics.

4. T h e Sdutim of ihe Pmblanr

Al. Normalforms
An algorithm to compute the n a a l form is inherent to the definition of the
normal form in the previous section, md needs no further discussion.
where A symboliccomputationpmgrunshould provide facilitiesfor computing(Lie)
derivatives, testing involutivenew, computing solutions of seta of nonlinear
dgdmic quatiom tbr the invemc mapping, and of wle of parlid differential
equations to transform to a simple n m a l form. Besides, more mundane
facilitieslike symbolic subaitution are needed.
4.2 Zcrodynmia
When the system has a well defined relative degree, the zero dynamics follows
fran t h e n a a l farm, by substitution of the output nutling input U* and using
e.
the propw that the states that do not belong to the zero dynamics, can be
set to 0.
F a systems without a relative degree the zero dynamics can be computed
by using the Zero ~ m l cAfgorMm. s The way this algorithm works is
by considering a sequence of nested nubmanifolds Mi. with Mi 3 Mi+! and
MO = h-'(O).ic.thefitstsubmanifoldistheinverseimageofthepointy = 0.
When m e conditions are fulfilled this sequence convergea to the locally
m h d output zaoing Wbmanifold 2 ' in some hti@o&ood o f f and
then exia a mapping U* auCh t h a t f ~ = /@) + gCt)u*Q is tangent to z'.
me pair < z * , f ) i s a m the zero dynamics ofthe sygcn. when the mappin
H(x) is defined in aneighbomood (Io f P by 2' n U = {x E: U:HW) = Of
the hput U* CYI be computed an the solution of GHQ)+ L,HCx)u' = 0. F a
derrila ofthis alprilhm. lrnd Uic conditionsto be imposed to guarantee
convmgence ofthe sequence to z',we refer to 191.

277
 thefe Cnirt mhtionr A&), i = 1, ...,m, f a the following partial differential
alurtiolu
LUG&@)I 0, f a 0 S k S ri - 2md 0 S j 5 m. (a)

ALO r b , r i = n, *the of in- {rl, ...,

it} is the W i e
-vcdor. ’ I h e m nmdions 1 em be amputed. b a d on a conamdive
proofoflbeoran4.1.
U- the fiurtianr 4,the change of coordin~eaand stabfeedback U =
a@) + p@)v that mive the &ate re exact linearivltion p b l e m can be
or”.
An “lydr of thir algaithm &owr tht a symbolic computation progrrun
Quld be &k to compute the Lie daivltive md Lie produd, perfam matrix
Veda mulllplicriion. compute a matrix inverse, UC.lhese computatim are
rrtthrdy w.The mein problem io in the canputation of the functions Ai,
wl#cpr*.l diiFaarcial equation8have to be incegnted. We will discuss this
in scaion 5.
5. Solvtkn with Symbdk Computation
SoMio118for tbe problems dgeribed in the previous d o n , i.e.to compute
1. theItmnAlfam
2 hstcrodynamica
3. the input-output exact linearization
4. the alate preexact liDariwion
me hchded in NdCON. Fa Ppblems (1) ancl (4)the systan should have a
well ddimd dative degrre. for (2) and (3) this is not n-. Implementa-
tionr of algorithmr f a Mme o h problems, ag.. to solve partial differential
equati008, n inchded in this package b. Ihe crtrudure of the implementa-
tianisrtacsedinFig l.

fa a and P in aneighborboodo f f . Fa fiuihs dehila of thir algorithm m

refer to [91.
An 8nalyais of the smctum Algorithm &ow8 that a symbolic atmpuuion
progran A d be abk to compute the Lie derhrrsive.perform marix-vastw
d murix-nutrix muliplidon, check if anne eJrprdon ir epwl to 0.
perfam rrat te8a computethekcanelof anupping, r lhaccanplwau
are r e w e l y atrai@fo”l Hae a h , the mrinpmbhn irto 6nd aohtians
foradaofnanlinereqwtiana
4.4.8taocprrartli.crintk.
’Ihecmditiom~so~gtheeactlinesriptionpmblemusgiv~.To~
the ml&m m a c oanprcly, we define the a x d l e d didbutionr

Gi = apan(.afgj: 0 5 k 5 i, 1 5 j 5 m ) f a 0 S i S R- 1.
Ib de4ine the adpint d.we need the Lie product

figure 1: StNcture of N o d C O N

-se the N&ON package is ampk documented in 1101, there is no need

f a @tion hac F a 8 more d b l e accuunt, aee [6]. Hae, we only
foau on m e apecific problems that appeared during the implementation
ofthe alpritlunr presented in Section 4. lheae problems p-t the major
bocllcacEtr fa the aymbolic rolution of our d&gn problana
hthepreviouraection wen&edthatamainproblem wastocomputesolutions
fasets of nonlineu equations In MASLE a facility is available to solve
of rrmlinar equations, namely the solve function, but this facility is not
. did not try to impmve thia
dwayr d c i a ~ IWe
h the previau. aection we a b remarked that another main problem was
to campute the findom 4..nd that the integration of partial differential
~ s m r r 6 n a l in thiscomputation. In MAPLE ahnostno facilitiesare
dI&kto robe putial diff-al equationa To improve this the following
muleW.rdlorccr.
”be M aldilTera~tialequrtiau we have to solve are from the “completely
m k ” type, in otha words, baaed on Frobarius’ ’Iheorem. we know
1&r I #ohtion f a tbe p4.l d i f f d a l equations exista To compute the
roLtimr Frobmius’ lbeoran i t d f is of no help. lhis problem was solved
by computing the mlution8 with M algorithm baaed on a conmuctive proof
of Fmbedur’ lbawan. ’Iheprocedure is M followa
Ihe mlution of the p 4 . l diffeencial equation can be condructed dnwn the normform
solutions of related ads of a d i diffeamtial cqurtiona Because IbfMU zero dynamics is
prwidg~e&cifitierforthistypeofprobkms,thedsolveoo"adthe [eta[l]dot = eta[2], eta[Z]dot = grav sin(eta[ll)l
probkm aemsmlved, Howeva.thedaolvecrwn"disnosvcypmvcrlu2 output nulling input2 is
and is o b unable to p " t a solution, althou& this solution is knownto ex- [ sin(eta[l]) eta[2] t cos(eta[ll) grav sln(eta[l
ist. Therefore the dsolveproccdure was exbded m an d hoc way. tohvldk transform
&ts of differential equations, Mrting h the
imp-)
"*"
a krga ckss of problems, by implementing a recumive p"to mlve transformation
(amuned to be the
equation, sub&ituting the mlution in the ranlining equations, and
[ x[ll, x[3], XL21, X[41 I
inverse transformation
soon. Noeffort wan spend in wing to detect .(block) trianp,ubr dependency (x[2]=eta[11, x [4l=eta[2], x [ll=zeta[ll, x [3l=zeta
m d ue in the set of diffe~entialequations, that would be a mare rigorous
r
option. Despite this exkmion. the computation of the fundions k is o h lhmtheaecondcesefory=&.
U M J C C ~ ~erpecially
, for more complicated crydema, o N&W cannot normform
finishthecanputationa'LhisputofN&~No~thenforebe~~derrd zero dynamics is
[eta[l]dot = eta[21, eta[2]dot = 3/5 grav sin(eta
ssexpdmentd ItscanrunlilrelythltuMIher~boliccanputationprogrun output nulling input is [ 1/5 grav sln(eta[ll)
will improve this situation. transform
For m e otha minor pmblans, that wexe m c o u n during ~ the imphen- transformation
Won of the algorithms, work around8 are applied. We discuss m e of these [ x[21, x[41, x[ll. x[31 1
work arounds. The problw wanthat someofthedrndard W functionsin inverse transformation
the linalgpackageareonly aJitableforrationdpo~miab.Thiswastoo (x[4]=zeta[2], x[3I=eta[21, x~2l=zeta[ll, x[ll=et
limited far o w purposer w o r e , the r a n k and implicitly the g a u s s e -
1im and gauss J ordprocedureawereextended. 80 alarger c h of problems Ihermlts coincide with the onesgiven above.
could be handled. ' b e modifidon consisted simply in moving b e c k &
for the type ofthe entrier of the m&ix for which the rank, Gaurs elimina- h p k 6.2. l h e model of the syat#n is
tion or GUM J0rd.n form &odd be canputd. and adding an a d d i t i d all
to the s i m p l i f y function to amme that the dezsction of pco apremions
wasgrrvantesd f o r . k r g e r c ~ a p r o b ~ r I h e n m , ~ o n a e x t r a n k .
extgausselim,andextgauss~orduethaeforeerctendedinrnthar
ad hoc way. and this part of N&ON should be conSidered 8s e x p e d m d
*.
6. Texibookprobkms The output functions do not give the model a full order relative d e p aa
To ilhuante the -of N ~ o weNcoasida severalexampler 'Ihe fin#three ~ b y t h e f o l l o w i n g b g o f a ~ ~ m w h e r r t h e r e l dfunction
eg
e~ramplesdo have a well defined relative degree. Ihe fint example computes ir used to compute the nktive d-
mezao6/MmicaofaSISOsydemwirhtwo~forinputandadpue'Ihe * reldeg
second example is far the zao dynamica crmputation of a MIMO ry@em vector relative degree = [ 1, 1 I
total relative degree = 2
with two inputs and two outputs. The third example is f a the input-outpu matrix Adeg = I x[31 1 1
exact linearization problan, again for a SISO sydem. For the fauth example [ 1 0 1
the zero dynamia is computed with the Zero Dynamica Algorithm. For the bdeg = [ X U ] + x[ll x[41. xL21 exp(x[31) 1
61th the input-output linearizing -feedback iscomputed with theSlmcture
Algorithm. 'he kat two examples are based on modelsthat do nothave a well To peanit a linearizing s t a t e f e e d b e we have to find functionsI for which
defined reZative degree Both of these examplea are MIMO with two inputs
the model hM a full ader relative degree Ihenext k g of a NOtdCON d o n
and two outputs. All examplea are contrived on@. lhe first example is taken
ahowsthat this is p d b k Ihe fundon output f u n c l a u d to compute the
~om[11,Example12.43],andthenextfocu6om[9,Examplen7.4.1,4.8.1,I'sandthefunction s t a t e l i n t o c a n p u t e t h e ~ z i n g ~ f c e d ~ .
6.1.2.5.4.1resp.1. outputfunc
output functions lambda for full (vector) relative degree
&ample 6.1. The model of this system. a robot with two revolute joints, [ xr31, XI41 - x[21 - x[ll I
nee Fig. 2, Mbe derived with the metbod of Lagrange from the following * statelin
expfwaions for the kine& and potentid energy 7' and'b [I 11 the exact linearizing feedback: (U)
3
2~ = m l / t B : + m 2 (@f+&bI + & ) ' + ~ 1 1 2 ( b i +&)cos&) [ v[l] - x[21 exp(x[31) -
2 x[31 x[21 2 x[31 , -
- x[3]v[ll - v[llexp(x[31) - v[21 + 3x[3]x[2]exp(x[3])
-V = { ( m l + mzh- el + m2l2 cos(& + &>I. 2 4
t 2 X[Zl x[31 + 2 X[31
lhe inputs ul and u2 are the joint toques and the outputs y~ and yz the joint 3 2
positions. lhe &ate x of the model camenponds with the degFeee-of-freedom t 2 exp(xt31) x[31 t x[21 exp(x[31) - xi21 exp(x[31)
andtheirderivativesinthefolbwhgwayf=[& & & &I. -
2
x[21 exp(xl31) xI31 - x[ll - x[21 + x[41 - x[lJ x [ 4 1 I

h 'his cannot be verified a g a published mb,because thir model is used

in p] for other purposes. Nevertheless, due to intemal vaification in fhe
NodCW package, the m k ir cared.

Erpmplc 6.3. The model of the system is

Figure 2: Robot with two revolutejoints

nK a h is to derive the dynamica of the model when it ir conetroiipt& "e For this model we cannot find a Anction I that gives the model a fill order
aretwocaaes,thetimtonewithy-~~ = 01conetrainedtooanduz =O,the relativedegree,beau&tbeconditiau,oflheomn4.1 areviokted.Accading
second one withy = yz = & CmMIUned to O n d u1 = 0. A d g to [Ill to [A the function A@) .IXI giver themodel a r e w e degree of 3. the hgea
t h e m dynunica arcgiven by (with a U m o d e l p u u n ~ to 1) that irpossibk. The funclion output f u n c that isnormally u d to compute
functions that give the model a fclll ada relative degree is ala0 abk, but
f2=m & = g d n q reap. &=@/S)griaq. only for S E 0 sytfana, to compute functions that will give a maximal rehtiie
ThefollowinglogofaNor&"?adon howsthattheredtaof [lllcanbe degree. l l m f o W i g bg of a hlrvlE d o n bows thir
reproduced. ~ o f u n c t l o n s a r c u dnormformtocanputethezaodynam-
: outputfunc *
ics (expressed in thenew coordinrtg) and t r a n s fo m fathetransformation
the output function gives relative degree of [ 3 ]
output lambda for maximal relative degree [ x[l] ]
cothenewcoordin~and~iavaae.Ihebgiorthefirstcrse,whaey= 01.
 This mlt is in agreement with the result mentioned above. The normal form
of this model with 1 = X I Moutput y is given by 19. p. 2121 (U)
fl = 52
4 = 53
The folbwig kg of a - session.with aane edits, showsthat NO*W
can compuw the input-output linearizing statefeedback. From the functions
appliedby N d C m weanly need inout 1i n that implementsthe Structure
Algdthm.
inoutlin
23 = v the exact linearizing feedback and linearized
system dynamics: (U, fl, 91)
9 - 2 1 +<. 3
[ V[ 11 tx [lI-x [31, VI2 I -2x [l]x [2]-2x [l I -x [l I x [31 1
This is computed with the N o r d c function
~~ normfo r m . 2 3
[ x[2ltx[ll , -2x[1]x~2]-2x[11 , 0, x[l]-x[31,
normform 2
the normal form: (fn, gn) XI51 tx 131 +x[llx[2l-x [2lx [31 1
2 2 [ 0 0 1
[ zeta[2], zeta[3], zeta[ll , zeta[3] t zeta[l] ] [ 0 1 1
[ O I [ 1 0 1
[ O I [ 1 0 1
[ I 1 [ x[21 0 1
[ O I the explicit feedback:(alpha, beta)
the zero dynamics: (zerodyn) [eta[lldot = 01 3
the zeroing input: (uzero) [ 0 1 [ X[l] - X[31, - 2 X[l] XI21 - 2 x[ll - x[ll x[31 1
[ 1 0 1
This result (of the form t = f . ( z ) + g,,(z)v) does not agree with the iesult given [ O 1 1
h e . The differmce is in theequstion fort3 that ShOuM b e 3 = + v. 'lle 4
tam 5: h.8 been inadvmtly dropped in [9]. The computed statefeedback agrees with the redts given above.
7. Vcbidc dmulstion pmblcrn
.4. lhemodel of the s y m is
Brsmpk 6 This larger eale p m b h is ddved 5om UI inverse simulation p b l e m in
muki-body dynamicd. In this pmbkm a vehicle is required to perfam a
Mndardized manoeuvre. To simulate. the manoeuvre the required inputs to
the systan must be know. Normally, the manoeuvre is such that a suitable
selected output of the syatem can be set to 0. In mod ~ ~ theae8 outputs
8 do
not fully detemine the beluvia of the sydan and the remaining M o m
represenfs exactly the mdynmnics. If this dynamica is stable, then the
This model h.8 no well dehed relative degree at XO = 0, because simulation is ala0 stnble,if it is -le, then additional measures are needed
toperfmastablesimulation.

A= :[. ]:
As an example we use a simple two dimensionalmodel of a vehicle, a M
moving in the horizontal plane, with W o n and tire forces acting on the
e

frame end-points. It looks like the model of a biie. because it only oonsidas
the center line of a mota vehicle, see Fig. 3.
is singular f a x 2 = 0. The zero dynamics is given in [9] by& = -x3 (ii the
original amrdinates) and the zeroing input by

'Ihenext edited bg of a m session showsthat the ZBP dynamics can be

computed by N-ON. From the fimdions supplied by NO&ON we only
need ext normfo m that implementsthe Zero Dynamics Algorithm.
extnormform
1st step. output zeroing submanifold:
[x[21 = 0, x[ll= 01
1st step, constraints on the neighborhood: none
2nd step, output zeroing submanifold:
[X[21 = 0, x[ll = 0, x[41 = 01
2nd step, constraints on the neighborhood: none
3rd step, output zeroing submanifold:
[x[21 = 0, X[l] = 0, x[41 = 0, x[51 = 01
3rd step, constraints on the neighborhood: none
3rd step. matrix L-gE has full rank m
the zero dynamics: (zerodyn) Figure 3: Model of a frame moving in the horizontal plane with faces acting
IX [1Idot=O,x [ 2 1 dot=O, x [3ld0t=-x [3], x [4]dot=O,x [ 51 dot=O ) on the h e endpoints
the zeroing input: (uzero) [ 0, -
x[3] ]
The required manoeuvre is steady state tuming: the longitudinal speed of the
Only the non trivial equations of the zero dynamics are of importrurce. The center of mass M of the vehicle is wncrtant. and a point P on the axis of the
resultscoincide with the ones given above. frame should describe a circle of specified radius. The gecific problem is for
which distances p. from P to M. the m d y n a m i c s is stable or unstable.
a m p k 65. The model of the s y m is The equations of motion of the model are
-
w,,, c - ~ fsin@ + e) F, sin e + F,,cos e
my,,, = +Ff ws(8 + e) + F r cos e +Ft,sin 0
J8=aFfws6-bFr
with inputs Ft,( W o n face) and 6 (Steering angle). The Ume degreea-of-
freedom&,.yn.and 8,arereSpectivelythecoordinatesof M and the orientation
This model has no well defiaed dative degree, because of the frame with respect to a fixed reference h e cxo,
5'). indicated by
the aqmmipt '.The steering angle 8 and the drift angles af and a, are
A= [: E] given with reaped to a body fixed reference frame. The Erame has mass m and
moment of inertia J with respect to M.The forces acting on the frame are the
traction force. Ft, and the lateral tire faces Ff. F,. The lest two forces can
is singular for dlx. Accading to 191 the neccsmry feedback f a input-output be srpresad in the drift angle and the normal tire force F,,by the 90 called
linearizatim is U = a@) + /3Qv with Msaic f m u h
W".
a) = D(F.) sine Udan(Bv)) (7)
v = (1 - E ) t + (Em)arctan(Bt), t = a + Sh

280

ri -
~.
 At the moment, the computations cannot be perfamed for complicated sys
tar& due to the d a e d capability to solve sczs of nonlinear equations and
becum the pdbilitieu to solve seta of nonlinear differential equations are
limited, ‘Ihrrforc.the designas of c i ” l systems cannot yet routinely com-
pute solutions f a bproblems, using tools that are based on symbolic
computationp r o m s .
Toremedy this,we recommendto extend the capabilities of symbolic compu-
tation prognms f a solving large and intriute seis of nonlinear equations and
f a solving sets ofdiffeenti.1 equations. Another possibility is to use another
and dpiUun. cg..for the Mte space e m linearidon problem. that completely
w a d s the inteption step.
Futureresesrch willhaveto aim at
devising new or modifying existing algorithms (0 be mare e63cient in
spaceandtime
where the aup“ipt * indicateathecoo&mtes with respect to thebody fixed implanentingtherlgaithmsinamoreefficientway, especially with regard
ref- h e (Z’ ,:Z ).Eqecially the functionaldependency of the hta;rl to compute memory requirements
tire forces on the de-of-ftcedom and the steeringangle make this system solving more wall and larger scale problems, to further guide in the
quite nonbear and difficult to Maze. selection of pressing lines of reeevch
By choosing the uutputa &I To come back to the title of the paper. at the moment we cannot deny for
mre the viability of this apptoech without being unjust to the developera of
y1 =t,,,Me+j,,,sine-vd symbolic computation programs and without being shut of of a promising
y2 = csl+ p COS el2 + cvn +pain el2- e, approach that Mawnphnent a cambined analylical @irper and pencil) and
numdcs approach. Both Utitudes to symbolic computation. Le, to refinii
and to embrace are (not) viable, depending on the expectations of the near
with R d the derited radius of the C k C h Md v d the deihd bll#lUdind rpced. future, and the viability question is still open. It is hoped fa, and more a
theerequirrdm~~cm#pondswithazaooutpltIhtouautyzdspends leu exp&ed, that withii this deedc the efficiency and capacity of general
m thedid8nsCp fwm F t o M . To mlve tbe probkm. ilra canprde the zero symbolic computation programs arc &M&. and that the capabilities of
dynamics and t h a ~detamie its ntability as a function of p. datively cheap computers will erulble the solution of the benchmark and
The firer pmblan encountered in computing the ZCIO dynamics is the non larger problems at reasonable costa
& n e chuaaa of tbe eqlutions with fwpecl to the m g mgk 6. To
ov-me this problem an additional integrator M be added to the system
and I can be ryprded u anew input. Howeva, this cumwthe system to Acknowledgement
have a singular decoupling ma?rix A. which makea the analysis more difRcult
The extnormform fmdion. b ~ e on d the Zao m m i c s Algorithm, can Ham VM Eapen implemented the enalysis and design algorithms in MAPLE
be used but this leds to immnounlable problems in the computalion due and is the main author of N d O N . Ted van de Broek supplied the vehi-
to huge memory need& soolution is lo add an integrator to the other cle simulation problem and cmesponding model equations. The author is
hput Fe d80, Md &g ph M a new @Ut, m&ing the dCCalpkg matrix indebted for both contributions.
mgulu by rbo “dekying” the input Fe. All this dding of intssr;rtOn must
be pafamed by hmd through madificationsoff and g. beuuae the Q ” i c
EwansionA&dtbm is not implemented. References
’Iheadditim oftwohtcgr&a pamits the use of the studad ’on [l] T. Umeno, K. Abe, S. Yamashita, and 0.Saito, “A software package
tothe m a l fam with the fmction normf o m . &hough fora largersym. for control design based on the algebraic theory using symbolic manipulation
B e to the larger n u m b of again p b h s uiaed with the memory hnguage REDUCE:’ in Ikx.TENCON 87.voL 3, (Seoul,Korea),pp. 1102-
needed to @cm the oompltrtion.To simplify the jmbkm. Motha “W- 1106. IEEE,Aug. 1987.
[Z] H.A. Baker, Y.W. KO.and P. ” e n d , m e application ofa compute%
-
urating” hction was wed in the magic famuk (7 ‘ ) by uaing instead
of sin(uban(-)).Now, the tnulsfamalion, urd even a solution f a the partial algebra system to the analysis of a claar of nonlinear systems,” in Nwlinear
di-al equation LfhQ 0. Mbe computed, but the invuac t M s f o r Cmrml Sybtans W g n : sdence papers of the FAC Symposium, (Capri,
mat~oncannot, beuuaeof UIPlifici@llimit in the WE of objects permitted by Italy), pp. 131-136. IFAC, Oxfad: Pergamon Press,1989.
the softwue, puttin8 Muntimely end to the computation. [3] R. RothfuO, J. Schaf€ner, and M. Zeitz. “RechneqestUtae A ~ l y ~ e
Anothe cutempf to solve the problem of non rWne input is not to a c e the und Syntheae nichtlineam Systeme.” in Nicht- Regclung: Merho-
modelbut to simplipV it mthe input 6 appears linear in the equations. To this den, W h e . Anwendungen, voL 1026 of VDIBmkhte, pp. 267-291,
ad,themaunption that I is mallhas to be adopted. ?hen,by using a Taylor Diis~eldorf:VDI-Valag, 1993.
series expansion, they- of equations caa be written lineer in b if the [4] 0.ALhriff and G.L.Blankenship. “Computa awn algorithms for non-
depcndarcy of I of t h e w ktaltite force isdroppd. so I hould idso be linear control,” in A d w i n d Computing Cancepts and Techniquesin Control
mull rclrtive to cy. m unrdidc asamption. lhis time the tmsfomution figjneaing (M. I. Denham and A. J. Laub, eda). voL47 of NATOASISm’es
to the m a l fam M be computed, but the oonlinur yaem, needed to E pp. 53-80, Berlin: Springer-Ver@. 1988.
compute the inverse transform&ion. could not be solved,this time not due to [SI B. de Jager, “Nonlinear control system analysis m d design with Maple:’
manory c a n d t s but due to Imit.tioas of the solve functim of MAPLE. in Mkid Mgeace, Expat Sy&“ and Symbolic Computing (E. N.
conchding: several rttanps to mlve this p b l e m were in vain due to HouStis and J. R. Rice. eda), Selected and wised papem from the IMACS
limitsonthemurimrlobjeurize 13th Wald C O ~ ~ S S(Dublin,, Irelend, July 1991). pp. 155-164, IMACS,
h e t e r d ~ nN: o ~ U l - H ~ l h 1992.
d.
limitations of the solve function
limits of the milabk physical (or vhtual) memory. [6] H. van Essen and B. de Jager, “Analysis and design of nonlinear control
sy-s with the symbolic computatbn system Maple.” in h. of the second
The lnat problem is relatively eury to mlve. but becuiae thememory require- Eumpesn Cmbd Cod (J. Nieuwenhuis, C. Raagman. and H. Trentehan,
ments are likely to be qxmentid or double exponential with respect to the edr). vol. 4, (Groningen. IXe Netharlands), pp. 2081-2085, June 1993.
problem size, this is not a dKlp way out. Ihe fiw two problems hould be [71 B. de Jager. “Symbolic calculation of zao dynamics fanonlinear control
resolved by tbe Mrrpre developas. systems,” in Ikx.of the 1991 h t m a t . Symp. on Symbolic and Algebraic
All in rU,the symbolic armputcltion softwareyet is not mature enough to solve Computalicn, ISSAC’91 (S. M. Watt, 4.) (Bonn.
. Gennany). pp. 321-322,
largm rrle pmbluns. and in this case we have to remt to a p u d y numeric ACM Press, New York,July 1991.
apprach to detaminc the M i ofthe zao dynamics. [SI B. de Jager, ‘Symbolics f a control: Maple used in solving the exact lin-
We propole to use this @kan &I a benchmut.a ydrtick. for the viability earizationpmblem.”inStudiainCmputcrAlgelmJ%wMu~, ptooeednm
of symbolic canputrtioa programs. ofthe 1992 SCAFISeminar, voL Il of The SCAFIhpea (Amsterdam, The
Nechalands). pp. 125-134, CAN, 1992
8. Condudon and DiruriOn [9] A. ldai.NonLinearConOSysfms:AnhlnniuUio~z Berlin:Springer-
The computation of the namd form, of the m dynamics, of the solution Verhg, 2nd ed., 1989.
of the input-output and the d.te lplct exact kearizalion pmblems con be [lo] H.vanEssen,”Symbolsspe;rkbuderthannmnbas:Analysisand design
automatedby using s y e l i c computalian, ag.by Uangthe N ~ O pack- N of nonlinear cantrol systems with the symbolic computation s y m MAPLE,”
age ’Lhir meansthat it i s much easier now tou~cantrolksb a d on the Maercr’s thcais, Eindhoven Univadty of Technoloa, Dept. of Mechanical
linearization apprmch, that can hlly take into account the nonlinedies in Engineering,June 1992. Report WFW 92.061.
ml systems. An enhancement of the p e r f o r m ~ ~ofe some cantrol syatems, [ll] H.Nijmeijer and A. J. van der SchaFl, Nonlinear Qvnamical Confro1
for a large set of operation conditions, can therefore be expected. Sy-a New Yorl;:Springer-Verlag, 1990.

281