Proceedings of the ThPI20.

18
46th IEEE Conference on Decision and Control
New Orleans, LA, USA, Dec. 12-14, 2007

Stabilization of an Electrostatic MEMS Including Uncontrollable

Linearization
Guchuan Zhu Jean Lévine, Laurent Praly
Department of Electrical Engineering Centre Automatique et Systèmes,
École Polytechnique de Montréal École des Mines de Paris
C.P. 6079, Succursale centre-ville 35 rue Saint Honoré
Montréal, QC, Canada H3C 3A7 77305 Fontainebleau cedex, France
guchuan.zhu@polymtl.ca {jean.levine,laurent.praly}@ensmp.fr

Abstract— Electrostatic micro-actuators are not linearly con- Micro-actuator is a key component in such applications
trollable in a set containing the origin due to a quadratic term as adaptive optics in which a deformable mirror is actuated
of electrical variable appearing as the input to the mechan- by an underlying two-dimensional array of such devices,
ical subsystem. Consequently, such systems are not feedback
linearizable and thus not differentially flat on this set and the forming the desired configuration of reflecting surface for
application of techniques based on feedback linearization leads correcting optical aberrations [19]. Potentially, a deformable
usually to an unbounded control. This work aims at developing mirror in adaptive optics systems can contain several hundred
control schemes which should be bounded everywhere in the thousands of micro-actuators. Therefore the control system
whole operational range. As there are no existing general is of great importance in order to obtain the enhanced
frameworks for tackling the control design for the system under
consideration, the approach of Lyapunov design combined with performance required in this application.
backstepping is used. The obtained control scheme is proved to It is straightforward to show that due to the quadratic
stabilize the system at the above mentioned uncontrollable set. term q 2 appearing in the mechanical subsystem, where q
Furthermore, we address the output feedback control using a represents the charge on the device (cf. (1) in Section II),
reduced order observer and certainty-equivalence implementa- the Jacobian linearization of such a system is not controllable
tion. The closed-loop stability is demonstrated by both stability
analysis and numerical simulation. at points where q = 0. Consequently, controls derived from
feedback linearization will usually explode as the trajectory
I. I NTRODUCTION of the system is approaching these points. We remark that
the only uncontrollable equilibrium point for System (1) is
This paper addresses the problem of the control of a one the origin, corresponding to the zero-voltage position. As
degree of freedom (1DOF) parallel-plate electrostatic actua- the zero-voltage position is often taken as the initial position
tor driven by a voltage source. The schematic representation of the device, one needs to operate frequently around this
of such a device is given in Fig. 1, where m is the mass of singular point. It happens also in the application of adaptive
the movable upper electrode, b is the damping coefficient, optics that a big amount of devices need only to deflect
k is the elastic constant, A is the area of electrodes, G is slightly from their initial position for producing desired con-
the air gap, G0 is the zero-voltage gap, x is the normalized figurations for certain patterns. Hence, a precise manipulation
deflection, and R is the loop resistance. This is one of the for small deflections while assuring the overall performance
most popular devices in micro-electromechanical systems in the full operational range is a realistic requirement. Note
(MEMS), such as micro-mirrors, optical gratings, variable that System (1) is globally asymptotically stable at the origin
capacitors, and accelerometers. This simple MEMS is often if the actuation voltage is set to zero. Therefore a simple
modeled as a rigid body, though it may be considered as a solution for avoiding the singularity is to remove the control
reduced-order model of an infinite dimensional micro-device, signal if one wants to bring the device to this position.
e.g. micro-beam or micro-plate [16]. However, with this control the accuracy and the resolution
of the system around the origin will be compromised. In
addition, the amplitude of control signal can still be very
high over a region near the origin, which constitutes a serious
obstacle to experimental implementations.
Control of linearly uncontrollable systems has attracted
many attentions in recent years and one can find in the
literature existing frameworks for tackling this problem for
systems with special structures, in particular systems includ-
ing odd order power integrators [2], [20], [9], [3]. However,
Fig. 1. 1DOF parallel-plate electrostatic actuator.
for systems with even order power integrators, as the one
This work was supported in part by the Natural Science and Engineering considered in this work, no simple solution is known. In this
Research Council of Canada (NSERC). work, we seek control schemes that should be bounded in the

1-4244-1498-9/07/$25.00 ©2007 IEEE. 2433

46th IEEE CDC, New Orleans, USA, Dec. 12-14, 2007 ThPI20.18

whole operational range while being capable of enhancing x2 = v − v̄, x3 = q − q̄, and u = us − ūs , the system (1)
the performance of the system described in (1). becomes
Note that to avoid the singularity due to the uncontrollable
linearization, one can use such techniques as input-output ẋ1 = x2 , (2a)
linearization [12], which is an implementation of the charge 2q̄ 1
ẋ2 = −x1 − 2ζx2 + x3 + x23 , (2b)
feedback introduced in [15]. However, the performance of 3 3
the system using this type of control is mostly dominated 1 2
ẋ3 = (q̄x1 + x1 x3 + (x̄ − 1)x3 ) + u, (2c)
by the mechanical subsystem and might be poor if, e.g., the r 3r
natural damping of the device is too low or too high. A which is defined on the state space
remarkable work on avoiding this singularity while providing
X̄ = (x1 , x2 , x3 ) ∈ R3 | x1 ≤ 1 − δ − x̄ .
© ª
an enhanced performance is the passivity-based control [10],
[11]. In this work, we present an alternative solution based
The set-point control of System (1) is then transformed to
on backstepping and Lyapunov design. We will show that our
the stabilization of System (2) at the origin.
design will result in a continuous or smooth control and the
Note that (2c) can be written as
system considered can be stabilized by an output feedback
control. ẋ3 = ū (3)
It is interesting to note that the dynamics of some magnetic
levitation systems have similar properties as the one studied where ū is a new control defined by
in this paper [7], [8], [13]. A more popular example is 1 2
the magnetic levitation of a steel ball (see, e.g., §8.3 of ū = (q̄x1 + x1 x3 + (x̄ − 1)x3 ) + u. (4)
r 3r
[5]). In these systems the electrical variable (the current or
the flux) acts as the input to the mechanical subsystem in The contact of the two plates happens when x1 = 1−δ−x̄,
quadratic form. The approach proposed in this paper can at the boundary of X̄ defined as
then be applied to these systems. Since electrostatic and © ª
∂ X̄ = (x1 , x2 , x3 ) ∈ X̄ | x1 = 1 − δ − x̄ .
electromagnetic forces are among the most popular actuation
mechanisms in electromechanical systems, we might expect After contact, if one continues to charge the device, the dy-
that our work is of practical interest. namics of the mechanical subsystem might collapse. When it
happens, the mechanical subsystem will no long be governed
II. S YSTEM D ESCRIPTION
by (2a)-(2b) but by
According to [14], [21], the dynamical model of 1DOF
parallel-plate electrostatic actuator in a normalized coordi- ẋ1 = 0, (5a)
nate is given by : ẋ2 = 0. (5b)
ẋ = v, (1a) Hence, the system exhibits switching behavior on ∂ X̄ . Note
1 2 that at any equilibrium q 2 = q̄ 2 = 3x̄ ≤ 3(1 − δ), therefore
v̇ = −2ζv − x + q , (1b)
3 q can be used as a switching signal, as proposed in [10]. To
1 2 completely characterize the contact dynamics, we need the
q̇ = − q(1 − x) + us , (1c)
r 3r following assumptions:
where x is the deflection, v is the deflection speed, q is Assumption 1: [10] The velocity of the moveable elec-
the charge, us is the actuation voltage (the control signal), trode before and after contact satisfy the relation x2 (t+c ) =
ζ > 0 is the damping ratio, and r is the resistance in the −µx2 (t−c ) where 0 ≤ µ ≤ 1, and x (t
2 c
−
) and x +
2 c ) are
(t
actuation circuit loop. All variables appearing in (1) are the velocities of the moveable electrode just before and after
defined in normalized coordinates and are dimensionless. contact respectively.
Physically, the movement of the actuator is limited by the Assumption 2: [10] When the system is restricted on ∂ X̄ ,
fixed electrode. As in practice an insulating layer is often the mechanical dynamics are governed by (5a)-(5b) if q 2 ≥
added on the fixed electrode in order to prevent the device 3(1 − δ) and switch back to (2a)-(2b) if q 2 < 3(1 − δ).
from shorting, the maximum displacement of the movable Assumption 3: The charge on the device just before and
plate is x = 1 − δ, where δ is the normalized thickness of after contact remains unchanged.
the insulating
© layer. System (1) is thusª defined on the state Note that Assumption 1 is an intuitive consequence of
space X = (x, v, q) ∈ R3 | x ≤ 1 − δ . Newton’s law of motion. When the movable plate hits the
To address the problem of stabilizing the system at the set- fixed one, it will change the moving direction. After contact,
points, we consider an equilibrium (x̄, v̄, q̄, ūs ). To determine the velocity would be reduced or become null since the
all equilibria of System (1) we note firstly that v̄ ≡ 0. kinetic energy might be partially or entirely absorbed at the
Secondly, since the electrostatic force is always attractive contact. Assumption 3 is also in accordance with the physical
regardless of the sign of the charge, there are no equilibria property of electrostatic actuators, because, as the current
for x < 0. Therefore, x̄ ∈ [0, 1 √
− δ] and at the equilibrium across the device is always finite, one can not add or remove
the charge is given by q̄ = ± 3x̄. Letting x1 = x − x̄, charges to or from the device instantaneously.

2434
46th IEEE CDC, New Orleans, USA, Dec. 12-14, 2007 ThPI20.18

−
III. S TABILIZATION BY S TATE F EEDBACK This implies V (t+ c ) ≤ V (tc ). Hence, V is a common
A. Stabilization Including the Uncontrollable Equilibrium Lyapunov function for the switched system. Finally, as the
largest invariant set in
Our objective is to find a controller being able to stabilize n o
System (2) at any equilibrium, including (x̄, v̄, q̄) = (0, 0, 0) E = (x1 , x2 , x3 ) ∈ X̄ |V̇ = 0
at which the system is not linearly controllable. Consider
now an energy-like Lyapunov function candidate: is the origin, we can conclude from LaSalle’s invariance
principle that System (2) is globally asymptotically stable
1 2 1 2
V1 = x + x . (6) (GAS) at the origin with the proposed control.
2 1 2 2 The actual control can be obtained by reversing (4), which
The time derivative of V1 along the solutions of the corre- is given by
sponding subsystem of (2) is ³ r ´
u = − k2 (x3 − x3d ) − x2 rq̄x2 + (x3 + x3d )
2q̄ 1 2
V̇1 = −2ζx22 + x2 x3 + x2 x23 . (7) 3r 3
3 3 + ẋ3d − (q̄x1 + x1 x3 + (x̄ − 1)x3 ) . (12)
Therefore a virtual control of the form 2 2
Note that as
x3d = k1 sgn(q̄) min(x2 , 0)2n (1 − δ − x1 − x̄), (8)
ẋ3d =k1 sgn(q̄) min(x2 , 0)2n−1 (2n(1 − δ − x1 − x̄)ẋ2
where k1 is a positive constant, n is any positive integer,
− min(x2 , 0)x2 ) (13)
and sgn : R → {−1, 1} with sgn(x) = 1 iff x ≥ 0, would
render V̇1 negative semidefinite for any q̄. The term sgn(q̄) is differentiable for all n ≥ 2, u given in (12) is smooth for
is for dealing with q̄ < 0. Note that x3d is identically zero all n ≥ 2. When n = 1, u is only continuous.
on ∂ X̄ . Obviously, this control will have the effect of adding
B. Stabilization Excluding the Uncontrollable Equilibrium
damping to the system when x2 < 0.
As the virtual control x3d is differentiable, we can proceed As the virtual control given in (8) does not add sufficient
with the backstepping design by augmenting V1 as: damping in the closing phase where x2 is mostly positive,
we present another one which will add damping for both the
1 2
V = V1 + (x3 − x3d ) . (9) opening and the closing phases. However, as we will see later
2 on, this control cannot guarantee the closed-loop stability at
The time derivative of V along the solutions of (2) yields the uncontrollable equilibrium (x̄, v̄, q̄) = (0, 0, 0).
2q̄ 1 Consider the Lyapunov function V1 given in (6). We chose
V̇ = − 2ζx22 + x2 x3 + x2 x23 a virtual control of the following form
3 3
+ (x3 − x3d ) (ẋ3 − ẋ3d ) x3d = −q̄k1 tanh(x2 ) tanh(1 − δ − x1 − x̄). (14)
2q̄ 2q̄
= − 2ζx22 + x2 x3d + x2 (x3 − x3d ) Obviously, x3d is smooth and vanishes at the contact. Note
3 3
1 1 ¡ 2 that since the value of x might be smaller than −1 − δ, we
+ x2 x3d + x2 x3 − x23d
2
¢
3 3 need to saturate the amplitude of 1 − δ − x = 1 − δ − x1 − x̄.
+ (x3 − x3d ) (ẋ3 − ẋ3d ) With this virtual control, the time derivative of V1 becomes
2q̄ 1 1
= − 2ζx22 + x2 x3d + x2 x23d + (x3 − x3d ) V̇1 = − 2ζx22 − q̄ 2 k1 x2 tanh(x2 ) tanh(1 − δ − x1 − x̄)
3 3
µ 3
× (2 − k1 tanh(x2 ) tanh(1 − δ − x1 − x̄)) , (15)
µ ¶¶
2q̄ 1
× ū − ẋ3d + x2 + (x3 + x3d ) .
3 3 which is negative semidefinite if 0 < k1 < 2.
Therefore a control of the form We can proceed once again with backstepping design and
prove by following the same analysis as in Section III-
ū = −k2 (x3 − x3d ) + ẋ3d − x2 γ(x3 , x3d ) (10) B that System (2) is globally asymptotically stable at any
where equilibrium points, except for the uncontrollable one, with
2q̄ 1 the control given in (12), where
γ(x3 , x3d ) = + (x3 + x3d ) (11)
3 3
ẋ3d = − q̄k1 1 − tanh2 (x2 ) tanh(1 − δ − x1 − x̄)ẋ2
¡¡ ¢
and k2 is a positive constant, would render V̇ negative
+ tanh(x2 ) tanh2 (1 − δ − x1 − x̄) − 1 x2 . (16)
¡ ¢ ¢
semidefinite. It is straightforward to verify that the time
derivative of V defined in (9) along the solutions of the con- We can now expect to get enhanced performance by
strained system defined on ∂ X̄ is also negative semidefinite. combining the two virtual controls given in (8) and (14) as:
Furthermore, the virtual control x3d ≡ 0 on ∂ X̄ . Therefore ½
k1 sgn(q̄) min(x2 , 0)2n (1 − δ − x1 − x̄), x̄ ≤ X̄,
we have by Assumption 1 and Assumption 3 that x3d =
−q̄k1 tanh(x2 ) tanh(1 − δ − x1 − x̄), x̄ > X̄,
x22 (t+ 2 − (17)
c ) ≤ x2 (tc ),
+ 2 − 2
where 0 < X̄ < 1 is a constant, while using the control given
x3 (t+ = x3 (t−
¡ ¢ ¡ ¢
c ) − x3d (tc ) c ) − x3d (tc ) . in (12). Since the system is stabilizable at any equilibrium

2435
46th IEEE CDC, New Orleans, USA, Dec. 12-14, 2007 ThPI20.18

with the first virtual control in (17), the choice of X̄ depends with k̄v = 2ζ + kv . The vector-value function f (ξ, ε) can
only on performance consideration. be derived from (2) with the control given above and is
smooth for any n ≥ 2. Thus one can show, by using the
IV. O UTPUT F EEDBACK C ONTROL standard approaches in the literature, e.g. [17], that the above
Usually, the charge on the device and the gap between closed-loop system is semi-global asymptotically stable at
the electrodes can be deduced from the input current, the the origin.
voltage across the device, and the capacitance (see, e.g., [1]). In fact, since the origin of the system ξ˙ = f (ξ, 0) is GAS,
However, direct sensing of velocity during normal operations by the converse Lyapunov theorem [6], there exists a positive
for micro-devices is extremely difficult, if not impossible. definite and proper Lyapunov function V0 (ξ) such that
We need therefore to construct a speed observer in order to
provide the estimate of v = x2 required for implementing the ∂V0
(ξ)f (ξ, 0) ≤ −α(|ξ|) (26)
closed-loop control described in the previous section. It can ∂ξ
be shown that System (1) with the deflection and the charge
as outputs admits the observer canonical form [4]. Therefore where α is some positive definite strictly increasing function
it is possible to find a full order observer with linear error on [0, ∞), also called class-K function.
dynamics. However, we need only to directly construct a Consider the following Lyapunov function candidate
reduced order speed observer. Furthermore, since x2 ≡ v,
we can estimate the speed in the original coordinates. 1
V (ξ, ε) = V0 (ξ) + ε2 . (27)
Consider the following dynamical system: 2
1 We need to prove that for any 0 < c1 < c2 there exists a
ż = −((2ζ + kv )kv + 1)x − (2ζ + kv )z + q 2 , (18)
3 kv∗ such that for all k̄v > kv∗ the time derivative of V along
with z(0) = −(2ζ + kv )x(0), where kv is a positive real solutions of the closed loop system is negative-definite for
number. Thus, by setting all initial conditions starting from the set

v̂ = z + kv x, (19) Sξε = {(ξ, ε) : c1 ≤ V (ξ, ε) ≤ c2 }. (28)

the dynamics of estimation error ε = v − v̂ will be given by:
Since Sξε is a compact set, then if the control is smooth,
ε̇ = −(2ζ + kv )ε, ε(0) = −x2 (0), (20) there exists a positive constant k3 such that, for all (ξ, ε) ∈
Sξε ,
which is globally exponentially stable at the origin with a ¯ ¯
¯ ∂V0
decay rate defined by kv . This implies that (18) and (19) |
¯
¯ ∂ξ (ξ)(f (ξ, ε) − f (ξ, 0))¯ ≤ k3 |ε .
¯ ¯
form an exponential observer.
We use a certainty-equivalence implementation of state-
Hence, the time derivative of V along the solutions of the
feedback design by replacing x2 (= v) by its estimate x̂2 (=
closed-loop system satisfies
v̂). The control law is now given by
ˆ3d − x̂2 γ(x3 , x̂3d ),
ū = −k2 (x3 − x̂3d ) + ẋ (21) ∂V0
V̇ ≤ −α(|ξ|) + (ξ)(f (ξ, ε) − f (ξ, 0)) − k̄v ε2
∂ξ
where ≤ −α(|ξ|) − k̄v ε2 + k3 |ε|. (29)
½
k1 sgn(q̄) min(x̂2 , 0)2n (1 − δ − x1 − x̄), x̄ ≤ X̄,
x̂3d =
−q̄k1 tanh(x̂2 ) tanh(1 − δ − x1 − x̄), x̄ > X̄, By applying Lemma 2.1 of [18], (29) implies that there
(22) should exist a positive constant kv∗ such that for all k̄v > kv∗
ˆ3d
and ẋ is computed by
ˆ3d =k1 sgn(q̄) min(x̂2 , 0)2n−1 (2n(1 − δ − x1 − x̄)(kv x̂2
ẋ V̇ < 0, ∀(ξ, ε) ∈ Sξε . (30)
+ ż) − min(x̂2 , 0)x̂2 ) (23)
It can be seen from (20) that the stability of the observer
if x̄ ≤ X̄ or by error dynamics is independent of ξ. Therefore, it can be
shown, by following the same lines as in [17], that for a
ˆ3d = − q̄k1 1 − tanh2 (x̂2 ) tanh(1 − δ − x1 − x̄)(kv x̂2
¡¡ ¢
ẋ sufficiently large k̄v every trajectory starting from Sξε under
+ż) + tanh(x̂2 ) tanh2 (1 − δ − x1 − x̄) − 1 x̂2
¡ ¢ ¢
the constraint ε(0) = −x2 (0) is ultimately bounded and will
(24) eventually enter into an arbitrarily small neighborhood of the
origin contained in the basin of attraction of the closed-loop
otherwise. In (22) and (23), n ≥ 2 is a positive integer.
system. This proves our assertion.
Denoting ξ = (x1 , x2 , x3 ), in coordinates (ξ, ε) the closed-
As System (2) is stable with the developed control at any
loop system can be expressed as
equilibrium point, including the one at which the system
ξ˙ = f (ξ, ε), is not linearly controllable, we have achieved the design
(25)
ε̇ = −k̄v ε, objective.

2436
46th IEEE CDC, New Orleans, USA, Dec. 12-14, 2007 ThPI20.18

V. S IMULATION R ESULTS 1

We have performed numerical simulations to verify the

Gap (1−x )
1
0.9
stability and the performance of the control scheme pro-
0.8
posed. As the performance of the state feedback control
is usually superior to the one of its certainty-equivalence 0.7
implementation, we present only the simulation results for 0 5 10 15 20
Time (t)
the latter. The simulated device is a under-damped system
(a)
with ζ = 0.1 and the resistance in the driving circuit loop is
supposed to be 1. The actuator is supposed to be driven by 2
x =0.05

Control Signal (u)

f
a bipolar voltage source. 1.5
x =0.1
f
x =0.15
We have found that the performance of the system at f
xf=0.2
different set-points is not significantly affected by the vari- 1

ation of the controller gain k2 and the observer gain kv . 0.5

Therefore once tuned their value is fixed. However, when 0 5 10 15 20
Time (t)
the observer-based implementation of the virtual control (8)
(b)
is used, the controller gain k1 should be carefully adjusted at
each set-point, in order to obtain satisfactory performance. Fig. 2. Stabilization of 0.95, 0.9, 0.85, and 0.8 gap position in closing
We have used a polynomial interpolation of well tuned values phase: (a) gap (1 − x1 ); (b) control signal u.
at several set-points to present k1 as a function of set-points.
For the observer-based implementation of the virtual control 1

(17), k1 is set to 1.9. In the simulation, we chose X̄ = 0.2 0.8

Gap (1−x )
1
as the threshold for selecting the virtual control in (22), 0.6

µ = 0.5 for determining the plate velocity at the contact, 0.4

and δ = 0.05. 0.2

Figure 2 shows the simulation results for the stabilization 0 5 10 15 20

Time (t)
of 0.95, 0.9, 0.85, and 0.8 gap position in the closing phase
with x1 (0) = 0, corresponding to the zero-voltage position. (a)
Figure 3 shows the simulation results for the stabilization of 2
x =0.2
Control Signal (u)

0.8, 0.6, 0.4, 0.2, and 0.1 gap position. We can see that the 1.5
f
x =0.4
f
controller works well at all the tested operational positions 1
x =0.6
f
xf=0.8
and remains stabilizing even in the presence of contact. The xf=0.9
0.5
control signals are also depicted. We remark that in the initial
0
phase, a hight control amplitude might be generated. This is 0 5 10 15 20
Time (t)
because x3d = 0 at the equilibrium, which might introduce
an important deficit in terms of charge regulation. However, (b)
in the rest of the operation no excessive control efforts are Fig. 3. Stabilization of 0.8, 0.6, 0.4, 0.2, and 0.1 gap position in closing
employed. We can also see that, as expected, the control phase: (a) gap (1 − x1 ); (b) control signal u.
signals are quite smooth.
Figure 4 shows the simulation results for the stabilization
of 0.85, 0.9, 0.95, and the full gap position in the opening
Gap (1−x1)

1
phase with x1 (0) = 0.2. Figure 5 shows the simulation
results for the stabilization of 0.2, 0.4, 0.6, and 0.8 gap 0.9
position whit x1 (0) = 0.95, corresponding to the position
where the device is completely closed. It can be seen that 0.8
0 5 10 15 20
the controller works well in all the considered set-points, Time (t)

including the zero-voltage position where the system is not (a)

linearly controllable.
2 x =0.15
Control Signal (u)

f
VI. C ONCLUSIONS x =0.1
f
1 x =0.05
f
This paper addressed the control of a parallel-plate elec- xf=0.0
trostatic micro-actuator, which is a basic element in many 0

MEMS-based applications. We have developed a state feed- −1

0 5 10 15 20
back control which is bounded and can stabilize the system Time (t)
at any set-point in the operational range, including the (b)
uncontrollable sets. To meet the practical operation envi-
ronments, we constructed an output feedback control law Fig. 4. Stabilization of 0.85, 0.9, 0.95, and the full gap position in opening
phase: (a) gap (1 − x1 ); (b) control signal u.
based on the certainty-equivalence implementation using a

2437
46th IEEE CDC, New Orleans, USA, Dec. 12-14, 2007 ThPI20.18

1 [5] M. Krstić, I. Kanellakopoulos, and P. Kokotović, Nonlinear and

Gap (1−x1)
0.8 Adaptative Control Design. New York: John Wiley & Sons Ltd,
1995.
0.6
[6] J. Kurzweil, “On the inversion of Lyapunovs second theorem on
0.4 stability of motion,” Amer. Math. Soc. Translations, vol. 24, pp. 19–77,
0.2 1956.
[7] J. Lévine, J. Lottin, and J. Ponsart, “A nonlinear approach to the
0 5 10 15 20
Time (t) control of magnetic bearings,” IEEE Trans. Contr. Syst. Technol.,
vol. 4, no. 5, pp. 524–544, 1996.
(a) [8] J. Lévine, L. Praly, and E. Sedda, “On the control of an electromag-
netic actuator of valve positionning on a camless engine,” in Proc.
AVEC 04, Arnhem, The Netherlands, August 2004.
Control Signal (u)

1
x =0.8
[9] W. Lin and C. Qian, “Adding one power integrator: a tool for global
f
0 x =0.6
stabilization of high-order lower triangular systems,” Syst. & control.
f
x =0.4
lett., vol. 39, no. 5, pp. 339–351, 2000.
−1 f
[10] D. H. S. Maithripala, J. M. Berg, and W. P. Dayawansa, “Control of
xf=0.2
an electrostatic MEMS using static and dynamic output feedback,”
−2 ASME Journal of Dynamic Systems, Measurement and Control, vol.
0 5 10 15 20
Time (t) 127, pp. 443–450, 2005.
(b) [11] D. H. S. Maithripala, B. D. Kawade, J. M. Berg, and W. P. Dayawansa,
“A general modelling and control framework for electrostatically ac-
Fig. 5. Stabilization of 0.2, 0.4, 0.6, and 0.8 gap position in opening phase: tuated mechanical systems,” Int. J. Robust Nonlinear Control, vol. 15,
(a) gap (1 − x1 ); (b) control signal u. pp. 839–857, 2005.
[12] D. Maithripala, J. Berg, and W. Dayawansa, “Capacitive stabilization
of an electrostatic actuator: An output feedback viewpoint,” in Proc. of
the 2003 American Control Conference, Denver, CO, June 4-6 2003,
reduced-order speed observer and demonstrated the stability pp. 4053–4058.
of the closed-loop system. The simulation results have shown [13] K. Peterson, J. Grizzle, and A. Stefanopoulou, “Nonlinear control for
magnetic levitation of automotive engine valves,” IEEE Trans. Contr.
that the proposed control scheme exhibited a satisfactory Syst. Technol., vol. 14, no. 2, pp. 346–354, 2006.
performance in different operational conditions. However, [14] J. Pont-Nin, A. Rodrı́guez, and L. Castañer, “Voltage and pull-in time
the presented control did not allow adding arbitrary damping. in current drive of electrostatic actuators,” J. Microelectromech. Syst.,
vol. 11, no. 3, pp. 196–205, 2002.
To obtain a higher performance, one may consider a hybrid [15] J. Seeger and S. Crary, “Stabilization of electrostatically actuated
solution. For example, one can use the present control in a mechanical devices,” in Tech. Dig. 9th Int. Conf. Solid-State Sensors
neighborhood of the linearly uncontrollable equilibrium and and Actuators (Transducers’ 97), June 1997, pp. 1133–1136.
[16] S. Senturia, Microsystem Design. Norwell, MA: Kluwer Academic
a tracking control scheme (e.g., [21]) in other operational Publishers, 2002.
points. One may also take advantage of passivity-based [17] A. Teel and L. Praly, “Global stabilizability and observability imply
design (see, e.g., [10]), which would allow adding arbitrary semi-global stabilizability by output feedback,” Syst. & control. lett.,
vol. 22, pp. 313–325, 1994.
damping. The proposed approach can be applied to systems [18] ——, “Tools for semi-global stabilization by partial state and output
with similar properties, such as magnetic levitation systems. feedback,” SIAM J. Control Optimization, vol. 33, no. 5, pp. 1443–
1488, 1995.
R EFERENCES [19] R. K. Tyson, Principles of Adaptive Optics, 2nd ed. San Diego, CA:
Academic Press, 1998.
[1] R. C. Anderson, B. Kawade, D. H. S. Maithripala, K. Ragulan, J. M. [20] M. Tzamtzi and J. Tsinias, “Explicit formulas of feedback stabilizers
Berg, and R. O. Gale, “Integrated charge sensors for feedback control for a class of triangular systems with uncontrollable linearization,”
of electrostatic mems,” in Proc. of the SPIE conference on Smart Syst. & control. lett., vol. 38, no. 2, pp. 115–126, 1999.
Structures and Materials 2005, San Diego, March 2005, pp. 42–53. [21] G. Zhu, J. Lévine, L. Praly, and Y.-A. Peter, “Flatness-based control of
[2] J.-M. Coron and L. Praly, “Adding an integrator to a stabilization electrostatically actuated MEMS with application to adaptive optics:
problem,” Syst. & control. lett., vol. 17, no. 2, pp. 98–104, 1991. A simulation study,” J. Microelectromech. Syst., vol. 15, no. 5, pp.
[3] D. Dac̆ić and P. Kokotović, “A scaled feedback stabilization of power 1165–1174, 2006.
integrator triangular systems,” Syst. & control. lett., vol. 54, no. 7, pp.
645–653, 2005.
[4] A. Isidori, Nonlinear Control Systems, 3rd ed. London: Springer-
Verlage, 1995.

2438