Sunteți pe pagina 1din 6

..

IEEE tNDlA ANNUAL CONFERENCE 2004, lNDlCON 2005 199

On Balancing Control Strategies for a Reaction


Wheel Pendulum
.
Bonagiri Bapiraju, K.N.Srinivas, Prem. Kumar P. and Laxmidhar Behera

Abmr-IW paper invpstigatw various bdanciup control


techniques for a reaction wheel pendulum, a prototype of a
nonlinear underactuated system. Using classical linearization
techniques three different h ear models of RWP system nre
derived m u d the upright unstable equilibrium plot. In
the Brst model the exact nonlinear dynaplics of R W using
EL formulntlotu is linearized using lbylors r r l o s 6irst order
spproximatlon. Second llnEarlzatloa technique u 8 ~ 1collacated
linearization whicb converts the system dynamics to two w o i n t
subsystems having nonlinear and linear dynamlcs. Ia the third
model the state tradormatioa is used in such a way that
system dynamlm changes from hurtb order to third order. AU
the algorithms have been simulated and experimented in real
time. Firrplly a fuzzy lqlc controller hm been implemented for
1
-dag the d o n wheel pendulum Fig. 1. The Reaction Wheel Pendulum

I. INTRODUCTION
Reaction Wheel Pendulum was introduced by Spong el [ll] used the neural approach for partial, full and approxi-
al. [l] where a dual mode control consisting of swing up mate input-ouput linearization. Pfeiffer [12] showed that the
and balancing control was implemented. As swing up control feedback linearization becomes robust when combined with
takes the pendulum very near to the upright position, the sliding mode control. For unkown nonaffine systems fuzzy
balancing control is aktivated. In a sense, this scheme involves based adaptive linearizing control has been discussed in [13].
switching from swing up to balancing control to stabilize In this paper, two control techniques are proposed to
the pendulum at the upright position. Saber [2] used a state stabilize the RWP at the upright position. The first one is
transformation that transformed the pendulum dynamics into linepiization technique and the second one is fuzzy logic
a cascade nonlinear system in strict feedback form and then controller. In linearization technique, three different models
he applied the standard backstepping p d u r e to globally of the plant are taken. The linearid models were found to
stabilize the top equilibrium point. A forwarding technique be controllable and stablized using pole placement.
[3] has been used to globally stabilize the pendulum at the All the controllers are implemented in real-time on a mecha-
upright position. Although 1-r two algorithms [2], 131 result tronics control kit.
in asymptotic stabilization for the pendulum, the demand on The paper is organized in the following manner. In section
transient torque and speed are too high to implement in real- II the dynamics of RWP are given. Control algorithms using
time. linearized techniques are presented in section Ill along with
Although there exists various nonlinear control strategies simulation and experimental results. Implementation of fuzzy
such as feed-backlinearization [4], [5], sliding mode control logic controller is described in section IV.Concluding remarks
[63 and back-stepping approaches [7], these techniques do not are given in section V.
directly apply to reaction wheel pendulum. In this paper, the
methodologies for balancing control have been investigated 11. THEREACTIONWHEELPENDULUM
from implementation point of view.
Reaction Whed Pendulum ( R W )is one of the simplest
Scott A. Bortoff [SI used spline functions for approximate
nonlinear underactuated systems. It is a pendulum with a
state-feedback linearization of a nonlinear system dynamics
rotating disk at the unpivoted end which can be driven
expressed in affw structure. Pseudolinearization scheme 191
has been discussed for systems,having input output behaviour by a DC motor. Thus the wheel is actuated while the
pendulum is itself unactuated. The coupling torque gener-
independent of operating point and static state feedback laws
ated by the angular acceleration of2the disk can be used
have been constructed. Rios-Bolivar [IO] used the backstep
to actively control the system. The schematic diagram of
ping approach for adaptive input-output linearization. Shouling
a Reaction Wheel Pendulum is shown in figure (1). The
The authors arc with the Department of Electrical Engineering, Indian system dynamics of a RWP can be derived using Euler-
institute of Technology, h p w , 208 016. UP, INDIA Lagrangian(EL) equations of motions. Following notations
I _ - _ _ I
I. I -. -
0-7803-8909-3/04/$20.00 02004 IEEE
m IEEE INDIA ANNUAL CONFERENCE 2Mw, INDIMN Z W

are used for RWP system parameters and dynamic variables: control kit so that simulation results can be compared with
mi : mass of thependuhm the experimental results.
m2 : massof thewhecl
11 : length of the pendulum 111. LINEARtZATlON TECHNIQUES USING CLASSICAL
2,1 : distance from the pivot to APPROACH
the center of the pendulum
11 : moment of inertia of the pendulum. . One of the most basic techniques in studying the stability of
I2 : moment of inertia of the wheel nonlinear systems around an equilibrium point is to linearize
q1 : pendulum angle with venical axis thesystem dynamics mound that equilibrium point. In this
92 : angle of the wheel technique three controllers are derived from three different
P : motor toque input applied on the disk models of the same pendulum system. In the first model,
The system dynamics as derived using EL formulations can we used the actual model (derived from Euler-Lagrangian
be given as [I41 formulation) [I] itself to linearize. In the second one the
Collocated linearization proposed by Spong [ 151 is used ,and
(m& + m& + I1 + 12)h- Fgsin(q1) = 0. in third one, the transformation proposed by Saber [2] is used.
1261 +I242 = 7 (I)
where A. Lineatiration of actual model
The system dynamic equation (3) can be rearranged in the
A = m12,1 +7nZll (2) following form :
Equation (1) can,& viewed in the following s k d a r d matrix -
m22 m12
* .ql = figsin(q1) - - . (74
form I. det(M) .: det(M)

W!J)ii+ dQ)= I6 .
I
(3)
where new notations can be encrypted as:
where d e t ( M ) = mslm22 - m21m12 = (mll:, + +
q = [ :i ] is the vector of generalised coordinates,
The linearized model around the origin is given by
m&
11)12 # 0. Consider the state vector X = [ q l , ~ l , q ~ , q ~ ] ~ .

u = [9] is t h i vector of torques at the joints, 0 1 0 0

-
- [ .mz1 m12
mi2 1 is the inertia matrix, r
-*o i
is the gravity matrix

Equation (1) can be rewritten using standasd notations intro-


duced in equation (4) in the following manner:
(4)

= AX+&
f

l&l (8)
Thedeteqinant of controllability matrix comes out to be
m11& + m 2 i 2 + d91) = 0 (34 &$f
+ O, that means the linearized system is control-
m2141 +mz242 = T (Sb) lable. tncc a full state feedback controller

Here the simplicity of the model, when compared to a general 7 = - K T X = - [ K1 K2 K3 K4 I[ q1 41 ~2 Q IT


underactuated system with two degrees of freedom.is evident with an appropriate choice of gain vector K, it is able to
from two facts viz, * stabilize the system in neighbourhood of the origin X =
in the above model the centrifugal Coriolis matrix (see [O,O,O,OIT. Here the controller attempts to bring the wheel
_..
(3)) is null and angle also to zero which we really do not need. Instead
the inertia matrix, M is constant and symmetric positive a state feedback controller can be designed with states
definite (q1, q1, q 2 ) , where the controller only stabilizes the above three
In this paper, following parameter values have been considered states. The state feedback gain vector has been computed
for simulations as [-356.215, -35.857, -0.0437971 for balancing in upright

[ 0.004596 2.495 x ] ,
position and the simulation and experimental results are given
in fig(2). Even if a full state feedback controller is designed,
M(q

9(d =
2.495x
[
2.495 x
0.3549sin(ql)
I
,
(6) ] the mode corresponding to 42 should be made very slow, which
is apparently same as not feeding back 92. Moreover in real-
0 time system motor speed will not become zero. This is due to
The parameter values in (6) has been computed using es- the disturbances acting on the system, connecting cable tension
timated parameters of the Quamer supplied mechatronics etc.
B. h'ncarbtion of tMnsfomed model 1 transformations
The system obtained from Collocated linearization [15] is 21 = - m 4 1 + m 1 2 4 2
used to design the linear regulator, Collocated linearization rz = G1 ._
separates the system dynamics to two disjoint subsystems. 23 = 92 (13)
One subsystem has nonniiear dynamics and the another one
has linear dynamics. The nonlinear subsystem-dynamics is the system equations ( 5 ) can be written as
. .
linearized and linear regulator is designed. The transformation.' i t = fhgsin(z2) . .
is as follows:
i2 = (21 - mlZz3)/mll
mlql +mz* = 21 I
is = U (14)
m l l h i-mh = 22 (9)
.where U is the virtual input.
Using the above transformation, the model (1) is transforked The model is linearized around the origin as:
to

$1 = [
= AX+Bu
0
+
0 i0
fig 0
0

(15)
Using equation (1 l), control torque has been calculated.
From the state transformation (13) and the transformed model
where U is the virtual input and the control torque is given by (14), it is clear that state variable q2 is not involved .and
the order of the system dynamics is reduced by one. The
7 = mz,na;l'figsin(g1) + (-~z";Tnt2 -I- ")U (11) system dynamics is described by state variables (z1,22,z3)
From (lo), its clear that first two equations form a non- and the linearized system (15) can be shown to be controllable.
linear dynamic subsystem and the latter form a linear double The state feedback controller stabilizes the system at X =
integrator. subsystem. The nonlinear dynamics is linearized [ZI,9 , 1 3 1 = [0, 0, 01 which corresponds to the equilibrium
around &e unstable equlibrium point and the linear model so point [q1,q1,Q] = (O,O,O] (from (13)). The balancing at
obtained is the upright position has been achieved by taking the state

;[ f
,feedback gain vwtor as [-7.661 x IO5,-3.7102 x lOf114.5]
21
and cbrresponding Simulation and experimental results are
shown in fig(4).

D. Simulation and Experimental Results


In this section, resultsand concluding remarks of balancing
the RWP using lineezation technique are presented. From
simulation results, it has been found that the state feedback
controller for linearized system performs well as long as the
pendulum angle 141I 5 .Ogradian. Both simdation !nd exper-
The linearized system is controllable hence a fuIl state feed- iments has been performed with initial state vector taken as
back controller stabilizes the pendulum around origin X = [ q l ,41, q2, 421 = [0.08,0,0,0] and torque saturated to f0.0473
\ z l , z 2 , @ , ~=
2 ][O,O,O,
~ IllT,which refer to the top equilib- Nm. Sampling interval is .taken as 5ms in both simulation
rium point that is [ q l ,ql ,~ 2&]
, = [O,0, 0,0] (from(9)). It is and experiment. Simulation and experimental rksults have been
obvious from the linearized model (12) that the state variable shown in figures (2), (3) and (4). Following observations are
~a can not be ignored to design the state feedback cotroller. So made:
the mode Eorresppnding to q 2 has been 'made very slow. The 1). The pendulum angle exactly converges :to zero in sim-
simulation and experimental results are given in fig(3) with ulation while this angle has small offset in steady state
state feedback gain vector chosen as [-8.119x lo', -0.773 x ' during ac'tual experiment. This is because of measure-
lo', 202.55,14.58]. Even though the gain vector is high, the ment offset and torque acting due to connecting wire
actual torque as computed in (11) is very small which is of tension.
the order of mN-m. 2) Even though wheel angular velocity follows same pat-
tern ,it does not converges to zero in real-time due .to
unmodelied dynamics.
C. Model tram-fomatbn and lirteurizalion. 3) The contol input does not reduce to zero in real:tike.'
This is due to measurement offset, wire tension, friction.
In this method the state transformation suggested by Saber 4) Though the control input is smooth in simulation, it is
in 121 is used. The transformed system dynamics is of reduced .fluctuating in actual experiment. Inorder to stabilize +e,
order and is in strict feedback form. Selecting follpwing state RWP in upright position the con~o1,sequencEmust be,.
I E ~ EINDIA ANNUAL CONFERENCEU)O~.'~ND~CON
zoo4

applied in such away so that the motor toque is both in


clock-wise and anti clock-wise directions. The velocity
has been estimated from position measurement and this
may introduce high frequency noise in control action,

(b) pndulum angle, 91

. W "
*1

(d) w h d vciocity. Qz

(d) wheel velocity, g


it .
' O
r..

I
I . 111
II

(f) toque input, T

ir
t '
'. . ,

Pi. 3. M&od-2 : Rcsponsc b a d on Transformed model 1 (left column of


figures ghows the simulation results, the corresponding experimental mults
tvc shown imright column)

.*I

(f) torqu hPUS 7

fixed one. The rulebase is designed based on the response of


state feedback regulator depicted in section @-A). Each of
the compcinents of the fuzzy controller given in figure ( 5 ) for
stabilizing the wheel pendulum at the unstable top equilibrium
paint locally are described.
Let t ( t ) denote the desired angularbpositionof the pendulum.
Tho goal is to balance the pendulum in the upright position,
i.e. r(t) = 0.
IV. FUZZYLOGICCONTROLLER
A fuzzy logic controller for stabilizing the pendulum at Fuzzi6lcation Module The main function of this module
the top unstable quilibrium point has been designed. The is to fuzzify the input crisp @oint-wise) value into a fuzzy
motive behind this work is to reassert the appiicability of set described by a membership function and a linguistic
the fuzzy logic cdntrol in real-time applications. Rule base, is variable in order to make it compatible with the fuzzy
designed based on the response of a linear regulator dcscribtd set representation of state variabie in the rule-antecedent.
in section (m-A). In the figure (3,architecture of a general The state variables ( q l l g 1 , q 2 ) have been taken as input
(nonadaptive) fuzzy logic controller has been given. The furzy to the fuzzification module (FM):The input variables are
logic conmller is noncldaptive as the rule base designed is a 'normalized to common universe of discourse of [-1,1]
INDIAN I"OF TECHNOWY, KlURAGPUR 721302, DECEMBER M22.2W 203

-1;
1- i
r
-0
L i ' ..............."
:i
,

m
a
'.....,,..,.,,,...

-
IN
f

N ~ i ~ d i s input
c d crisp vdw -
Pig.6. Membership definition

The normalization coefficients ( q l m o z ,qlmar,q2-=)


have been found from the response of the regulator
presented in section (IU-A). The coefficients have been
selected based on the region of convergence from section

-1 -1
(III-A). Once the normalised input crisp value is available
* it can be fuzzified.For fuzzification, a miangular member-
ship function has been taken. The universe of discourse
(d) whcct vclocity, Q ,
has been described by 7 linguistic variables as shown in
the figure 6.where the linguistic variables are denoted as
below
LI -.-,; LP :
HP
MP
Large Positive
: High Positive
: Medium Positive
1'8 .:j ; ...................................
SS : Small value
.. .. MN : MediumNegative
.. HN : High Negative
I i.:

-
(e) torque input,
U

T
"
3

(f) toque input, T


LN : Large Negative
and p denotes the membership value.
Rule base is solely designed based on the response
of the regulator with feedback variables ( q 1 , q 1 , q z ) . It
can be easily calculated that for 3 input variables and
7 linguistic variables total number of rules come out to
be 73 = 343. So 343 rules constitutes the rulebase. The
practical response of the Linear regulator was considered
for constructing the Rule base.
Inference Mechanism For a given input set of values
number rules fired may be more than one (in the present
problem at the most 8 rules can he). Individual rule
based inference mechanism has been used. In this
mechanism the idea is to compute the overall value
I -.
of the control output variable based on the individual
contributions of each rule in the rule base. Each such
I individual contribution represents the value of the control
output variable as computed by the single rule only. In
Fig. 5. Puay logic controller architcstui
this problem the following inference has been used
Truth value of a nth rule fired =
m i 4 C l ~ ~ l n o r mP)(,i l n o r m ) ,P(42*an;))..
Defuzzification Module This Module @M) converts the
set of fuzzified control output values into a single point-
wise values. The control output has been calculated using
. the Center Of Gravity (COG) as: , I

Each output variable corresponding to each rule is a fuzzy


204 MDlA ANNUAL CONFERENCE Z W , INDICON 2ow

singleton. These variables have been heuristically tuned. (E)assumes that the platform ofR W is stationary. But in real
Finally the control output value is denormalised. time it is wobbling. This adds one more degree of freedom.
This can be mother important explanation for control action
to appear fluctuating.
A. Simulation und Experimental Results
In this section results of balancing the RWP using fuzzy v. CONCLUSiONS
logic controlIer are presented. Both simulation zgd experimen- ' Three classical linearization techniques and a fuzzy logic
tal results are shown in fig(7). The initial vector is taken as controller have been implemented to balance a reaction wheel

"
[ql,41 ,a,6 1 = [0.08,0,0,0].The camments made in section pendulum in upright position. The experimental setup is a
( In-D) arc equally applicabk for this controller also. Mechatronics Kit provided by Quancer. The angular measure-
ments are made using relative optical encoders. The optical
encoders have a resolution of 4096 divisons per 2x radians.
The actuating torque is given by a 24V brushless DC mo-
tor. The controller is implemented using TI'S C6711 Digital
Signal Processor. The mother board of the conwller i s TI'S
C6711DSK. Since the maximum actuating torque of the qtup
is satuarated at the value 0.04437 Nm,the active region of all
the controllers designed is very narrow, 1q1I 5 .09 radian only.
Any linearized controller that would accomodate a wide range

j ;I
of operation will demand a very high torque at the control
actuation and hence such schemes can not be implemented on
a physical set up that we have taken. The best way to expand
wide range of operation is to opt for dual mode controllers
where a swing up conuollerpreceeds the balancing controller.
,~.
;; . ... "~ - Results related to dual mode control will be reported in another
-
~ ~~

;; . ... . "~ ~ ~~

paper. The fuzzy logic controller proposed here is for large


..
..
*.
number of parameters. It will be prudent to use optimazation
In! - technique such as GA to get good optimal behaviour of FLC.
.... . . i

.........:..........
-
i.
::
.-
.***-......... I
1 ;
REFERENCES
I
(11 Mark W; Spong. Nonlinear control of the inertia wheel pcnddum.
A#o&&, 37:184S-1851, 2001.
(c) whttl vtlofity, 8 (d) wheel vdocity, Q (21 h ovati spbor. Global stabilizafim of flal UnderaCtWd syslun the
inertia wheel pendulum. In proceedings of&h cortfcrrncc on dccirwn
and con", Orlando. 2M1. IEEE.
131 Romeo M g a Laurent Raly and Georgia Kaliora. Stabilization of
noalincar systems via forwsrding mod{&V}. IEEE Trnnrtrcionr on
Auton~#ic con", 46(9):1461-1466, !kpte&r Ulol.
IS] M. Vidymgar. Nunfincar Sy&m ARalysis. Rcntice Hall, Englewood
CWs..Npw J a y 07632,.1993.
is] * A l b Isidori. Nonlincar~Co~ml . System.
. Springer-Vcrlng London,
1995,
[6] Jean-Jacques E. Slotine and Weiping Lee. Applied Nodincar Contmf.
F'rcntic8 Hid. Eaglcwood Cliffs. New Jcrsey 07632, 1991.
I* I [7] P. Motovic M.Kristic, I. Kancltakopoulos. Nonlintor Ana' Adoptive
Control Design. John Wilcy & Sons, Inc, 1995.
(e) torque input, r ( f ) torque input, T [a] Scott A. Bortoff. Approximate sm feedback linearization using spline
functions. Aulomaticu. vol 33.1997 pp.1449-1458.
[9] Douglas-A. Lawrence. A gcncral &roach to inputoutput pseudoh-
pig, 7, Method-4 : Raponst baied on Pupy logic controUa(left column of wization for nonlinear systems. IEEE Pmccdings 011 Decision and
figum shows the simulation results, the companding upgimentd results ~ 0 t U r 0 l . 34th CO@WUX, 1995. pp.613-618.
arc shown in right urlumn) [lo] Zinokr A S R~os-BoUW M,S i r a - h w k H. &put control
via adaptive input output linearization: a backstepping approach. P m .
34th IEEE CDC,New Orkanr..Vol. 2,1995, pp.1579-1584.
ill] Unkhauca R Shouling He,Relf K. A dappmach for control of
nonlinear system with feedback linearization. N e d N e m r l u IEEE
In simulation the wheel velocity under equilibrium condition h. . pp. 1409-1421.
v01.9,1W8.
fig(7c) is not zero. Because the parameter q2 has not been [ 121 Pcmand0.L B Ffciffcr C .F,Edgar J.P. Robust feedback LidZarion and
optimized during the design of the controller. Even though fuzy control. ~meedingsof Amer. Cow, Con. San Diego, CA.,1999
pp. 1508-1514.
the controller WO& well in simulation, the control action in [ U ] Fqulloy'L Bo@ezzoula R, GsIichtt S. F w y adaplive linearizhg control
experimentation is highly oscillatory. Since in all experiments, . of uon-ahe rystcmk FLIZZ IEEE 2003. The ]EEE Inlemd~rkal
the velocity has bccn wtimated from position measurement * Cot$ennce on F u q Sysfem, USA.,2W3,pp. 543-548.
[14] I. Fantoni and R. Luzano. NbnJinrar Conlmf f i r Undcmcruared
using optimal encder, it is possible for measurement noise Mechpnud System. Springer,'London, u)M.
being amplified appreciably, This may be one of the reasons [IS] M. W. Spong. The control of undpacuretcd mechanical system. In
for the control action to osciIlate. The model derivedin section Fim Inrcmrional Conjcrence.on Mecartunus, Muicn city. 1994.

S-ar putea să vă placă și