Proceedings of the American Control Conference

Chicago, Illinois June 2000

On the Application of Multi-Parameter Extremum Seeking

Control
Gregory C. Walsh
Department of Mechanical Engineering
University of Maryland
College Park, MD 20742

Abstract In addition to handling several unknown parameters

simultaneously, the new algorithm presented in this
In many control applications the best set point for op- paper places few conditions on the system, filters or
eration is not know a priori, and yet, controller design search functions and hence leaves much room for inno-
tools tend to focus on regulation to known set points vation, such as using lead compensation on the estima-
and reference trajectories. This paper presents and an- tion loop. Two additions used in the flywheel appli-
alytically verifies techniques for on-line optimization of cation but not within the scope of the analysis include
set points. The main contributions of this paper in- the automatic scaling of S and choosing estimation gain
clude a new control law for multiple parameter set- k and search size 6 and separately for each parameter
points and a proof of exponential stability for the av- channel.
eraged system. Methodologies for implementing this
control law in a discrete time setting are discussed. Consider a general MIS0 nonlinear model
The application of this new algorithm to a magneti-
cally suspended flywheel is detailed. The extremum
seeking controller is used to minimize vibrations and
the results of twenty runs compared with the expected where z ( t ) E !Rnis the state, u ( t ) E !Rm is the input,
physical model of the process. The fit is poor enough y(t) E !R is the cost output, and f,h are smooth. Fur-
to justify the use of the on-line optimizer proposed. ther we have a control law u = u(z,O) parameterized
by the vector 8 E !RP which renders the system glob-
ally (or locally) exponentially stable with equilibrium
point z = l ( 8 ) . The parameter E is a measure of the
1 Introduction
speed of convergence. As E + 0, the system becomes
singularly perturbed and Tikhonov’s theorem [3] may
In this paper we provide a general framework for multi- be employed.
parameter extremum seeking control in an effort to ex-
tend the pioneering work in [5,6]. There are essentially The output y(t) reports to cost of state z ( t ) and we
three time scales in an extremum seeking control sys- assume the existence of a global optimum 8 * . With-
tem: the plant,the search, and the estimation. The out loss of generality we assume the optimal point is
searching algorithm selects equilibrium points for the a minimum. Define e(t) = 8 ( t ) - 8*. The cost func-
plant and does so slowly enough the plant may be ap- tion g(e) = h o Z(e + 8.) - h o Z(8*)is assumed to satisfy
proximated by a static map using singular perturbation globally (locally) convexity assumptions. That is, there
theory. The algorithm parameter S controls the size of exists five positive numbers 0 1 , 0 2 , 0 3 , 0 4 , cr5 and a dis-
the space searched, and for sufficiently small 6 this area tance r such that
will be locally convex. The estimation gain k is chosen
small in order to average the search results, creating
the third and slowest time scale, estimation.
(3)
Multiple parameters are addressed simultaneously by
frequency division in continuous time or code division for all e such that llell < T.
in discrete time. That is, a set of orthogonal periodic
functions are selected and each injected into a separate Remarks:
parameter channel. After passing through the plant,
the resulting mix is then demodulated and relevant All of these conditions are satisfied locally for suf-
gradient information extracted and used in a steepest ficiently smooth g with isolated equilibrium with
decent estimator. non-zero second derivatives.

0-7803-5519-9/00 $10.00 0 2000 AACC 41 1

 0 These conditions may be relaxed for slower than 2 Analysis
exponential convergence rates. Convexity is
needed, however. For small 6 we consider the singularly perturbed [3]
solution of the system. That is, we set

s(t) = l(e(t) + O* + 6Se(t))

Plant
h(z) = d e ( t ) ) (5)

I + - I " to find the differential equations

1 rI ' - z 9 -xlLow
I lnlrpnlor
-Faller
i P&S

Ch"l
1_ _ _ _ _ _ _ - - - - - - _ _ - - I
I
I
I
1
The time periodic portion of equation 6 appears in the
input stage scaled by 6, and is consequently small. Ex-
pand the map g with respect to small 6 and apply
:J Sd(t)ST(t)dt= I p , the p x p identity matrix.
io = a050 +Bg(e) +
Figure 1: The block diagram of the extremum seeking
controller.
+ lT B ( d v g ( e ) S e ( t >+ o(62))

We now add the estimator dynamics to the control sys-

tem. The equations of motion become
. ,
EX = f(z,u ( z ,e + e* + 6 S e ( t ) ) ) = AZ +B 6 2 v g ( q +~o(63)
ZO = aOzO + b o h ( z )
= -kCZ (7)
i = A2 + d B S d ( t ) ( ~ z o h ( z ) ) + t?

d = -kCz (4) where O(6j) are terms dependent on z and e and the
parameter 6 in a j or higher order manner.
where k is the estimation gain, S e ( t ) Sd(t), are vectors
of T periodic encoding and decoding functions, and 6 Proposition 1 T h e uverage, singularly perturbed
a small positive number. The filter Co(ao, bo, C O , 1) is equations 7 are globally exponentially stable for small
high pass, and C ( A ,B , C, 0) the concatenation of p in- positive 6 and k.
dependent low pass filters. Hence z = [zl;z 2 ; . . . ; zp]
and A = diag(a1 ,a2, . . . ,a,}. The matrices B and C
employ bi and ci respectively to implement the filters. The dynamics of the high pass filter CO, when aver-
The encoding S e ( t )and decoding functions Sd(t) serve aged, are decoupled for small 6. They depend on the
to separate searches among different parameters. These dynamics of E and Zi but the relationship is not recip-
functions are assumed to be normal in the sense that rocal. We will show the averaged dynamics of E and Zi
:J S d ( t ) S T ( t ) d t = I,, the p x p identity matrix. Such are stable, as the high-pass filter dynamics follows suit.
functions can be constructed using sinusoids in contin-
uous time and orthogonal codes in discrete time. Each filter Ci is stable, so there exists a positive definite
matrices Pt which solve aTPt + P t a i = - I . We also
Remark: The purpose of the high-pass filter CO is to know that since ai has no poles at the origin, ai1 is well
remove the average value of g ( e ) from the output signal; defined. For convenience we assume the passband gain
the purpose of the low-pass filters Ci is to average the of the filter is 1, hence h(0) = -ciu;lbi = 1. Consider
parameter gradient information. The filters allow for the following Lyapunov function.
higher gains in the continuous system estimators, and
Y
although not necessary for the analysis, are included
for completeness. The high-pass filter is removed by
V(Z,E ) = g(E) +k (ZTPtZi + VgiciaLlZi) (8)
i=l
the averaging arguments along with the average value
of g ( e ) ; the low-pass filters prove more complicated and First we verify that the function is positive definite
must be carried though the entire Lyapunov argument. decresent. Certainly if = 0 and Z = 0 then V ( z ,e) =

412
0. First examine whether V(E,E)>_ 0. Note that we for all i so again by the quadratic formula we may con-
can write k E:=, Vgiciai'zi is equal to clude that

= kVg(e)diag{ciai'}z
5 klldiadc1ai' 1II IlVgll llzll (9)
+k2ETQz
and further with 70 equal to the minimum singular
value of all P t ,
P
zTPfzZi = zTdiag(PfZ)z
i=l

2 II~1I2Yo (10) As matrix Q multiplied by a second order gain k while

the negative definite part is first order in k, for small k
we have we have % quadratically negative definite. The aver-
aged system is exponentially stable for a region deter-
mined by the validity of the a-bounds on g(e), so for
global stability, T = CO.

For sufficiently small k

3 Application
(k7d2 < a1k7o (12)
which implies, via the quadratic formula The algorithm was simulated on several test cases be-
fore its application to the flywheel. The interior of
V(Z,z) > f (ai llE1I2 + k7011zl12) (13) the estimator block is shown in figure 2. For every
frequency used, two parameters can be identified as
The Lyapunov function is decrescent because g(e) < one might suspect from adaptive control [8]. In this
a211e1I2 and the remaining terms are quadratic in na- example, the signals were 90 degrees phase separated
ture. and hence orthogonal. Simulation results are shown
in Figure 3, with optimal set point arbitrarily chose
The derivative % is equal to to be (-2,3). The plant in the simulation was an ill-
conditioned, randomly chosen quadratic static map.
P
- k c (-Vgicizi + 2 (Z'Ptaizi + b2$Pij:biVgi)) +
i=l
P
k (Vgicia,la& + Vgi62ciai1biVgi) -
i=l
P
k V2gijc,kjzjciai1& (14)
i,j=l
Figure 2: T h e Simulink block diagram for a single esti-
Note the first term is cancelled, and a quadratic term mation channel.
for the error is introduced after applying
The energy storage flywheel [9, 41 built by the author'
62ciai1bi = -S2 consists of two magnetic bearings which together sus-
(aTPij:+Pij:ai) = I . (15) pend a cylindrical shell consisting of a fiber-wound
epoxy composite as seen in figure 4. The shell serves
The resulting equations reduce to as the flywheel, designed without a hub so that stress
concentrations are avoided. A brushless DC motor,
sandwiched between the two bearings, spins the fly-
wheel up or down, adding or removing kinetic energy
from the system. The control system has, as a conse-
quence, x and y-axis inputs and outputs for both the
where the i , j t h matrix element of Q is Qij = upper and lower bearings for a total of four coupled in-
V2gjicTcjay1. For sufficiently small 6 we have puts and outputs with the the z-axis direction passively
'The author worked on a team consisting of Ronald Zmood,
James Kirk, Kevin Kefauver, and Louis Hromada. The project
was funded by an SBIR through NASA Glenn.

413
 Figure 5: Two-plane point mass model for mass unbal-
ance.

Figure 3: The estimate time history. methods require the estimation of the mass unbalance,
either as an axis mismatch with four parameters or as
suspended and rotation about the z-axis is controlled two point masses with four parameters. The second ap-
by the brushless DC motor. proach was chosen even though it is speed limited due
to the finite bandwidth of the bearing actuators. The
second approach yields direct insight t o mass removal
needed to balance the wheel.

The linearized equations of motion for the plant are

- - - K z ( t ) + w * P-dx (4 + Bu(t)
d2z(t)
dt2 dt

Figure 4: A photograph of the energy storage flywheel

prototype, disassembled.

Flywheels are used for attitude control in low-earth or-

bit (LEO) satellites. The attitude control flywheels,
with extensive modification, can be used also to store
energy in lieu of chemical batteries. Vibration prob-
lems due to mass unbalance in the wheel are partic-
ularly acute because a flywheel must be spun at high
B = [ l/m
0
0
h
J,I
0
l/m
-h/Jxz
0
l/m
0
0
-h/Jzx
0
l/m
h/Jxx
0
speeds in order to be used in for energy storage. Mass
imbalances are due to voids and other imperfections in
the fiber wound expoxy-composite rotor and create a
mismatch between the geometric and inertial z-axis of
rotation. The literature [2, 71 traditionally models the
c =
1 0 0
0 1 - h
[ 0l 10 h
0 ]-;
h
0

(20)
imbalance in a rigid rotor as two point masses, the first
located on the top rim, the second on the bottom rim where z ( t ) = [xcm,ycm, $ ( t ) ,$ ( t ) ] ,u ( t ) is the vector of
as shown in figure 5. The net effect of the point masses coil input currents, y(t) is the vector of distance mea-
is to introduce a force rotating with the flywheel [l]. surements, and w is the operating speed of the wheel,
Both the amount of mass and their angular positions measured in rad/s. For the flywheel under considera-
are unknown. tion, the rotational inertia about the x and y body-axe
of the flywheel is 1.563e-4 Nm s2, the rotational inertia
The effect may be compensated for either by letting the about the L axis is 1.141e-4 Nm s2. The dimension-
flywheel spin about its inertial axis or forcing it to spin less shape factor p is approximately 1. The mass of
about its geometric axis SBLC94.art,KTF97.art. Both the wheel is 340 grams, the height from the center of

414
mass is 3 cm. A linearized coil model gives the actuator
parameters kl and kz. 1.8-

1.6 -
The linearized model of the plant is a function of the
z-axis rotational angular velocity. At operating speeds 1.4 -
below 10,000RPM, the flywheel has four unstable poles
mirrored by four stable poles. At 10,000 RPM, four
5 l.* -
/’’
poles reach and thereafter remain confined to the imag-
inary axis, as shown in Figure 6. When the lineariza-
tion of the flywheel has imaginary open loop poles, it
is considered high-speed.

Figure 7: Magnitude of the resulting correcting term as

a function of flywheel operating speed for both
the upper and lower bearings. The solid line
indicates a least square fit.

References
[l] P. Bovik and C. Hogfors. Autobalancing of ro-
tors. The Journal of Sound and Vibration, 111(3):429-
440, October 1986.
Figure 6: The roots locus of the flywheel open-loop poles
as a function of operating speed. [2] L. Everett. Two-plane balancing of a rotor sys-
tem without phase response measurements. Transac-
The multi-parameter optimizer was run for 20 differ- tions of the ASME, 109:162-167, April 1987.
ent operating speeds of the flywheel. The separation [3] H. Khalil. Nonlinear Systems. Prentice-Hall Inc.,
of time scales used in the analysis was well supported second edition, 1996.
in implementation: the feedback law for the the plant
was run at 20kHz and the estimation algorithm at 10 [4] J. Kirk, G. Walsh, and L. Hromada. The open
core composite flywheel. In Proceedings of the Thirty
Hz, both on a single TMS320C31 DSP. Using orthogo-
Second Intersociety Energy Conversion Engineering
nal codes instead sinusoids greatly simplified the digital
Conference, pages 1748-1753, 1997.
implementation of the search algorithm. The two-plane
model of the mass unbalance suggests a quadratic re- [5] M. Kristic and H.H. Wang. Design and stability
lationship between correcting force magnitude and op- analysis of extremum seeking feedback for general non-
erating speed. The results, shown in Figure 7, shows linear systems. In Proceedings of the IEEE Conference
poor fit to a least squares quadratic although the gen- on Decision and Control, 1997.
eral trend is correct. The unbalance model used is to [6] M. Kristic and H.H. Wang. Stability of extremum
simple to capture the true behavior of the system; this seeking feedback for general nonlinear dynamic sys-
inaccuracy is compensated for by the four parameter tems. Automatica, 2000. to appear.
extremum seeking controller.
[7] S. Rao. Mechanical Vibrations. Addison-Wesley,
Reading, Massachusetts, 1995.
[8] S.S. Sastry and M. Bodson. Adaptive Control:
4 Conclusion
Stability, Convergence and Robustness. Prentice Hall,
1989.
This paper contributes a new control law for multiple
parameter set-points and a proof of exponential stabil- [9] G. Walsh, K. Kefauver, L. Hromada, J. Kirk, and
ity for the averaged system. Results of the application R. Zmood. A 50 wh open core high-speed flywheel. In
of this new algorithm to a magnetically suspended fly- Proceedings of the Thirty Second Intersociety Energy
wheel are discussed. The extremum seeking controller Conversion Engineering Conference, 1999.
demonstrates the need for on-line estimation of the vi-
bration parameters when the physical models used t o
describe the disturbance process are uncertain.

415