INTERNATIONAL JOURNAL OF ADAPTIVE CONTROL AND SIGNAL PROCESSING

Int. J. Adapt. Control Signal Process. 2005; 19:485–503

Published online 1 April 2005 in Wiley InterScience (www.interscience.wiley.com) DOI:10.1002/acs.868

Discrete time adaptive control for a MEMS gyroscope

Sungsu Park1,n,y and Roberto Horowitz2,z

1
Department of Aerospace Engineering, Sejong University, Seoul, Korea
2
Department of Mechanical Engineering, University of California, Berkeley, CA 94720, U.S.A.

SUMMARY
This paper presents a discrete time version of the observer-based adaptive control system for micro-electro-
mechanical systems gyroscopes, which can be implemented using digital processors. A stochastic analysis
of this control algorithm is developed and it shows that the estimates of the angular rate and the fabrication
imperfections are biased due to the signal discretization errors in the feedforward control path introduced
by the sampler and holder. Thus, a two-rate discrete time control is proposed as a compromise between the
measurement biases and the computational burden imposed on the controller. The convergence analysis of
this algorithm is also conducted and an analysis method is developed for determining the trade-off between
the controller sampling frequency and the magnitude of the angular rate estimate biased errors. All
convergence and stochastic properties of a continuous time adaptive control are preserved, and this
analysis is verified with computer simulations. Copyright # 2005 John Wiley & Sons, Ltd.

KEY WORDS: MEMS; gyroscope; adaptive control; discrete time control; hybrid implementation

1. INTRODUCTION

Gyroscopes are commonly used sensors for measuring angular velocity in many areas of
applications such as navigation, homing, and control stabilization. Although, conventional
rotating wheel, fibre optic and ring laser gyroscopes have dominated a wide range of
applications, they are too large and, most often too expensive to be used in most emerging
applications. Recent advances in micro-machining technology have made the design and
fabrication of micro-electro-mechanical systems (MEMS) gyroscopes possible. These devices
are several orders of magnitude smaller than conventional mechanical gyroscopes, and can be
fabricated in large quantities by batch processes. Thus, there is great potential to significantly
reduce their fabrication cost. The emergence of MEMS gyroscopes is opening up new market
opportunities and applications in the area of low-cost to medium performance inertial
devices [1].

n
Correspondence to: Sungsu Park, Department of Aerospace Engineering, Sejong University, 98 Gunja-dong,
Kwangjin-gu, Seoul, Korea.
y
E-mail: sungsu@sejong.ac.kr
z
E-mail: horowitz@me.berkeley.edu
Contract/grant sponsor: DARPA; contract/grant number: N66001-97-C-8643

Received 15 October 2003

Copyright # 2005 John Wiley & Sons, Ltd. Accepted 14 January 2005
486 S. PARK AND R. HOROWITZ

Most MEMS gyroscopes are laminar vibratory mechanical structures fabricated on

polysilicon or crystal silicon. Common fabrication steps include bulk micromachining, wafer-
to-wafer bonding, surface micromachining, and high aspect ratio micromachining. Each of
these fabrication steps involves multiple process steps such as deposition, etching and patterning
of materials. Generally, every fabrication step contributes to imperfections in the gyroscope [2].
Fabrication imperfections that produce asymmetric structures, mis-alignment of actuation
mechanism and deviations of the centre of mass from the geometric centre, result in undesirable,
systematic perturbations in the form of mechanical and electrostatic forces, which degrade the
performance of a gyroscope.
Currently, force-balancing feedback control schemes [1,3,4] have been widely used to cancel
the effect of off-diagonal terms in the stiffness matrix which is referred to as the quadrature
error, and also to increase the bandwidth and dynamic range of the gyroscope beyond the open-
loop mode of operation. However, they rely on the exact measurement of input/output phase
difference, and moreover are inherently sensitive to some types of fabrication imperfections
which can be modelled as cross-damping terms, which produce zero-rate output (ZRO).
Recently, a new gyroscope operation mode and a corresponding adaptive control algorithm
have been developed, which are well suited for the on-line compensation of imperfects and to
operate in varying environments that affect the behaviour of a MEMS gyroscope [5]. This
adaptive controlled gyroscope is self-calibrating, compensates for friction forces and fabrication
imperfections that normally cause quadrature errors, and produces an unbiased angular velocity
measurement that has no ZRO. The adaptive control algorithm presented in Reference [5] is a
continuous time controller. It is assumed that the control and parameter adaptation laws are
updated continuously in time. Although the implementation of such a controller is certainly
possible utilizing analog circuits, it is of practical interest to explore the implementation of the
adaptive control laws utilizing digital computers.
In this paper, we present a hybrid discrete/continuous time version of the observer-based
adaptive control system developed in Reference [5], which can be implemented using digital
processors. The control algorithm considered in this paper is not fully a discrete time controller,
since only the feedback control, parameter adaptation algorithms and feedforward control law
are implemented in discrete time, while the velocity observer is still implemented in continuous
time.
In the next section, the dynamics of MEMS gyroscopes is presented, by accounting for the
effect of fabrication imperfections. In Section 3, the observer-based adaptive control algorithm
presented in Reference [5] is reviewed. The hybrid adaptive control law is developed in Section 4.
In Section 5, the performance of the hybrid adaptive controlled gyroscope is analysed, including
magnitude of parameter estimation biases, convergence rate and resolution estimation. Finally,
computer simulations are performed in Section 6.

2. DYNAMICS OF MEMS GYROSCOPES

Common MEMS vibratory gyroscope configurations include a proof mass suspended by spring
suspensions, and electrostatic actuations and sensing mechanisms for forcing an oscillatory
motion and sensing the position and velocity of the proof mass. These mechanical components
can be modelled as a mass, spring and damper system. Figure 1 shows a simplified model of a
MEMS gyroscope having two degrees of freedom in the associated Cartesian reference frames.

Copyright # 2005 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2005; 19:485–503
 DISCRETE TIME ADAPTIVE CONTROL FOR A MEMS GYROSCOPE 487

e g
Ω

{g}
y

{e}
ê3

ê2
ê1

Figure 1. A model of a MEMS z-axis gyroscope.

Assuming that the motion of the proof mass is constrained to be only along the x–y plane by
making the spring stiffness in the z-direction much larger than in the x- and y-directions, the
measured angular rate is almost constant over a long enough time interval, and linear
accelerations are cancelled out, either as an offset from the output response or by applying
counter-control forces, then the equation of motion of a gyroscope is simplified as follows:
mx. þ d1 x’ þ ðk1
mðO2y þ O2z ÞÞx þ mOx Oy y ¼ tx þ 2mOz y’
ð1Þ
m.y þ d2 y’ þ ðk2
mðO2x þ O2z ÞÞy þ mOx Oy x ¼ ty
2mOz x’
where x and y are the co-ordinates of the proof mass relative to the gyro frame, d1;2 ; k1;2 are
damping and spring coefficients, Ox;y;z are the angular velocity components along each axis of
the gyro frame and tx;y are control forces. The two last terms in Equation (1), 2mOz x’ and
2mOz y’ ; are due to the Coriolis forces and are the terms which are used to measure the angular
rate Oz :
As seen in Equation (1), in an ideal gyroscope, only the component of the angular rate along
the z-axis, Oz ; causes a dynamic coupling between the x and y axes, under the assumption that
O2x;y Ox Oy 0: In practice, however, small fabrication imperfections always occur, and also
cause dynamic coupling between the x and y axes through the asymmetric spring and damping
terms. Taking into account fabrication imperfections, the dynamic equations (1) are modified as

Copyright # 2005 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2005; 19:485–503
488 S. PARK AND R. HOROWITZ

follows [2]:
mx. þ dxx x’ þ dxy y’ þ kxx x þ kxy y ¼ tx þ 2mOz y’
ð2Þ
m.y þ dxy x’ þ dyy y’ þ kxy x þ kyy y ¼ ty
2mOz x’
Equation (2) is the governing equation for a MEMS z-axis gyroscope. Fabrication imperfections
contribute mainly to the asymmetric spring and damping terms, kxy and dxy : Therefore these
terms are unknown, but can be assumed to be small. The x and y axes spring and damping terms
are mostly known, but have small unknown variations from their nominal values. The proof
mass can be determined very accurately. The components of angular rate along x and y axes are
absorbed as part of the spring terms as unknown variations. Note that the spring coefficients kxx
and kyy also include the electrostatic spring softness.
Based on m; q0 and o0 ; which are a reference mass, length and natural resonance frequency,
respectively, where m is a proof mass of the gyroscope, the non-dimensionalization of Equation
(2) can be done as follows:
x. þ dxx x’ þ dxy y’ þ o2x x þ oxy y ¼ tx þ 2Oz y’
ð3Þ
y. þ dxy x’ þ dyy y’ þ oxy x þ o2y y ¼ ty
2Oz x’
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
where ox ¼ kxx =ðmo20 Þ; oy ¼ kyy =ðmo20 Þ; oxy ¼ kxy =ðmo20 Þ; dxx dxx =ðmo0 Þ; dyy dyy =
ðmo0 Þ; dxy dxy =ðmo0 Þ; Oz Oz =o0 ; tx tx =ðmo20 q0 Þ and ty ty =ðmo20 q0 Þ:

3. ADAPTIVE MODE OF OPERATION

The aim of the adaptive mode of operation scheme presented in Reference [5] is to achieve (i)
on-line compensation of fabrication imperfections, (ii) closed-loop identification of the angular
rate, (iii) to attain a large bandwidth and dynamic range, and (iv) self-calibration operation. The
adaptive mode of operation operates based on the observer-based adaptive control which needs
only position measurements of the proof mass of the gyroscope. For convenience, governing
Equation (3) of the MEMS gyroscopes is re-written as matrix form
q. þ D’q þ Kq ¼ t
2O’q þ b ð4Þ
where b is Brownian noise and
" # " # " #
x tx 0
Oz
q¼ ; t¼ ; O¼
y ty Oz 0

" # " #
dxx dxy o2x oxy
D¼ ; K¼
dxy dyy oxy o2y
The adaptive control problems of the gyroscope is formalized as follows: given Equation (4)
with unknown constant parameters D; K and O; determine the control law t based on measuring
q; such that the dynamic range is constrained within a specified region and is estimated correctly.
Note that direct measurement of the velocity of the proof mass is avoided in the problem
formulation. This is because current velocity sensing circuitry technology produces a noise with

Copyright # 2005 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2005; 19:485–503
 DISCRETE TIME ADAPTIVE CONTROL FOR A MEMS GYROSCOPE 489

spectral power that is 3–4 orders of magnitude larger than the ideally expected value, as
compared with position sensing technology.
The observer-based adaptive control and parameter adaptation laws are introduced in
Reference [5] as follows:
# qm þ Rq
t ¼ t0 þ D’ # qm
# m þ 2O’
ð5Þ
’# ¼ 1g ðt qT þ q tT Þ
R 2 R 0 m m 0

’# ¼ 1g ðt q’ T þ q’ tT Þ
D 2 D 0 m m 0

’# ¼ g ðt q’ T
q’ tT Þ
O ð6Þ
O 0 m m 0

t0 ¼
gðq’# cp
q’ m Þ
# R
where R ¼ K
Km ; D; # are estimates of D; R and O; and g ¼ diagfg1 ; g2 g; g; gR ; gD and
# and O
gO are adaptation gains. Reference trajectory, qm ¼ ½xm ym T ; is updated by the following ideal
oscillator:
q. m þ Km qm ¼ 0 ð7Þ
where Km ¼ diagfo21 ; o22 g are the reference resonant modes of both axis. Note that the signal qm
and q’ m may be calculated and stored off-line, resulting in a significant reduction in the number
of on-line computations.
The velocity estimate q’# p is corrupted by a measurement noise and produced by the following
velocity observer:
q’# cp ¼ q# cv þ Lðq þ n
q# cp Þ
ð8Þ
q’# cv ¼
Km q# cp

where q# cp is the estimate of the position, q# cv is an additional state of the velocity observer, L is a
observer gain matrix given by L ¼ diagfL1 ; L2 g and n is a position measurement noise, which is
assumed to be uncorrelated with Brownian noise b: The estimated power spectral densities of the
position measurement noise ðSp Þ and Brownian noise ðSb Þ are given by [6]
 
2C0 þ Cp 2 4kB Td
Sp ¼ 4kB TRw ; Sb ¼ ð9Þ
2V0 ðdC=dyÞ m2
where kB ; Cp ; C0 ; Rw ; T; d and m are, respectively, Boltzmann’s constant, the device’s parasite
capacitance, nominal sensing capacitance, wiring resistance, absolute temperature, damping
coefficient and mass. Both are assumed zero-mean white noises.
We present the following two theorems without proof. The proofs can be found in
Reference [5].

Theorem 1 (Stability)
Given the observer (8), the adaptive control (5) and parameter adaptation laws (6), it is always
possible to choose a velocity observer gain L; which makes the trajectory error, ep ¼ q
qm ;

Copyright # 2005 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2005; 19:485–503
490 S. PARK AND R. HOROWITZ

trajectory estimation errors, q* p ¼ q# p
q and q* v ¼ q# v
q’ ; and their time derivatives converge

locally, uniformly and exponentially to zero.

Theorem 2 (Persistent excitation condition)

With control law (5), parameter adaptation laws (6), and the observer (8), if the gyroscope is
controlled to follow the mode-unmatched reference model, i.e. o1 =o2 ; the persistent excitation
condition is satisfied and all unknown gyroscope parameters, including the angular rate, are
estimated correctly.

Theorems 1 and 2 show that the motion of a mode-unmatched gyroscope, in which the
resonance frequency of the x-axis is different from that of the y-axis, has sufficient persistence of
excitation to permit the identification of all major fabrication imperfections as well as ‘input’
angular rate. This means that adaptive controlled gyroscope has no ZRO and is self-calibrating.
The main advantages of the adaptive mode of operation include self-calibration, large
robustness to parameter variations, and no zero-rate output. Moreover, the adaptive controller
design is also easy to implement in high Q systems. Thus, the noise properties associated with a
high Q system can be fully utilized. Another advantage of the adaptive mode of operation is that
it is easy to adjust the trade-off between bandwidth and resolution by simply adjusting the
angular rate adaptation gain.

4. HYBRID ADAPTIVE CONTROL LAW

4.1. Discrete time adaptive control

We now consider the implementation of the adaptive control algorithm described by Equations
(5)–(6) utilizing a digital computer. The control law and parameter adaptation algorithm will be
implemented in discrete time, while the observer (8) will be implemented in continuous time.
Thus, the adaptive algorithm can no longer be analysed as a continuous time system, but rather
as a hybrid system which includes both discrete time and continuous time algorithms. For
convenience, let us define an adaptation gain matrix G; a signal regressor Wðqm ; q’ m Þ; and an
unknown parameter vector y as follows:
 
G ¼ diag gR ; 12gR ; gR ; gD ; 12gD ; gD ; 12gO
" #
xm y m 0 ’
x m y’ m 0
2’
y m
W T ðqm ; q’ m Þ ¼
0 xm ym 0 x’ m y’ m 2x’ m

yT ¼ ½rxx rxy ryy dxx dxy dyy Oz 

where rij ; dij and Oz are, respectively, elements of R; D and O: Also, let us define the sampling
index k and the sampling time Dt; such that tk ¼ kDt: We define the hold operator H such that,
H½sðkÞ denotes
H½sðkÞ ¼ sðtk Þ for tk 4t5tkþ1
Using Wm ðtÞ as a short hand notation for Wðqm ; q’ m Þ; let us define Wm ðkÞ as
Wm ðkÞ ¼ Wm ðtk Þ ¼ Wðqm ðtk Þ; q’ m ðtk ÞÞ

Copyright # 2005 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2005; 19:485–503
 DISCRETE TIME ADAPTIVE CONTROL FOR A MEMS GYROSCOPE 491

Now, define the discrete time adaptive control law as

#
tðtÞ ¼
gH½ðq’# c ðkÞ
q’ m ðkÞÞ þ dqðkÞ þ H½W T ðkÞyðkÞ ð10Þ
p m

where dqðkÞ is a quantization noise of the velocity estimate. Because of finite word length error
or digital computational error, the actual signal Wm ðkÞ must also be modelled as Wa ðkÞ ¼
Wm ðkÞ þ dWm ðkÞ; where dWm ðkÞ is a finite word length error. However, since the signal Wm ðkÞ is
known in advance and can be generated by the computer, its finite word length error can be
made arbitrary small and will be neglected here. Note that the signal Wm ðkÞ may be calculated
and stored in an off-line fashion, resulting in a significant reduction in the number of on-line
computations.
The parameter estimate sequence in Equation (10) is updated by the following approximated
continuous time parameter adaptation law:
Z tkþ1 Z tkþ1
’yðtÞ
# dt ¼
GH½Wm ðkÞgðq’# cp ðkÞ
q’ m ðkÞ þ dqðkÞÞ dt ð11Þ
tk tk

Thus,
# þ 1Þ ¼ yðkÞ
yðk #
GWm ðkÞgðq’# cp ðkÞ
q’ m ðkÞ þ dqðkÞÞDt ð12Þ
while the observer signals q’# cp ðtÞ and q# cv ðtÞ are still updated by the continuous time observer (8)
and sampled at every Dt:

4.2. Stability analysis

In order to derive the closed loop error equations, we apply the control law (10) and the
adaptation algorithm (12) to the system equation (4) and observer (8). Define the signal
discretization error function as
DWm ðtÞ ¼ H½Wm ðkÞ
Wm ðtÞ ð13Þ
then the trajectory error equations are
e. p þ ðD þ 2OÞ’ep þ Km ep ¼
gH½’ep ðkÞ þ q’* p ðkÞ þ dq ðkÞ

#
þ H½WmT ðkÞyðkÞ
WmT ðtÞy
Rep þ b ð14Þ
where
#
H½WmT ðkÞyðkÞ *

WmT ðtÞy ¼ H½WmT ðkÞyðkÞ þ ðH½WmT ðkÞ
WmT ðtÞÞy

*
¼ H½WmT ðkÞyðkÞ þ DWmT ðtÞy
where y* ¼ y#
y: The trajectory estimation error equations are
q’* p ¼ q* v
L*qp þ Ln

q’* v ¼
Km q* p þ ðD þ 2OÞ’ep

þ gH½’ep ðkÞ þ q’* p ðkÞ þ dqðkÞ

*

H½WmT ðkÞyðkÞ
DWmT ðtÞy þ Rep
b ð15Þ

Copyright # 2005 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2005; 19:485–503
492 S. PARK AND R. HOROWITZ

where ep ¼ q
qm ; q* p ¼ q# cp
q; and q* v ¼ q# cv
q’ : Equations (14) and (15) can be described

compactly by the sum of a linear known and a linear unknown term as follows:

x’ c ¼ Ac xc þ Acu xc

Bc gH½’ep ðkÞ þ q* v ðkÞ
L*qp ðkÞ þ Lns ðkÞ þ dqðkÞ

*
þ Bc H½WmT ðkÞyðkÞ þ Bc DWmT ðtÞy þ Bc b þ Cc n ð16Þ

where ns ðkÞ is the sampled measurement noise, where ns ðkÞ  ð0; Sp =DtÞ; xc ¼ ½ep e’ p q* p q* v T
and
2 3 2 3 2 3 2 3
0 I 0 0 0 0 0 0 0 0
6 7 6 7 6 7 6 7
6 7 6 7 6 7 6 7
6
Km 0 0 07 6
R
ðD þ 2OÞ 0 07 6 I 7 607
6 7 6 7 6 7 6 7
Ac ¼ 6
6
7;
7 Acu ¼ 6
6
7;
7 Bc ¼ 6
6
7;
7 Cc ¼ 6
6 7
7
6 0 0
L I7 6 0 0 0 07 6 0 7 6L7
6 7 6 7 6 7 6 7
4 5 4 5 4 5 4 5
0 0
Km 0 R ðD þ 2OÞ 0 0
I 0

Since the ‘control input’ in Equation (16)

gH½’ep ðkÞ þ q* v ðkÞ
L*qp ðkÞ þ Lns ðkÞ þ dqðkÞ

*
þ H½WmT ðkÞyðkÞ

is held constant over the sampling period and the parameters are updated in a discrete time, it is
convenient for the further analysis to discretize the trajectory and trajectory estimate error
systems (16) as follows:

xc ðk þ 1Þ ¼ ðI þ Ac DtÞxc ðkÞ þ Acu Dtxc ðkÞ þ O1 ðDt2 Þxc ðkÞ

Bc Dtg½’ep ðkÞ þ q* v ðkÞ
L*qp ðkÞ þ Lns ðkÞ þ dqðkÞ

*
þ Bc DtWmT ðkÞyðkÞ þ O2 ðDt2 Þ þ CðkÞy þ Bc bd ðkÞ þ Cc nd ðkÞ ð17Þ

where O1;2 ðDt2 Þ are the terms of order Dt2 and denote how the remainder (the error) depends on
Dt; and
Z tkþ1
CðkÞ ¼ eðAc þAcu Þðtkþ1
tÞ Bc DWmT ðtÞ dt
tk

where bd ðkÞ  ð0; Sb DtÞ; nd  ð0; Sp DtÞ: Notice that, in the numerical integration of Equation
(16), the intensity of the sampled noises bd ðkÞ and nd ðkÞ can be approximately calculated by
multiplying the intensity of noises bðtÞ and nðtÞ by Dt:

Copyright # 2005 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2005; 19:485–503
 DISCRETE TIME ADAPTIVE CONTROL FOR A MEMS GYROSCOPE 493

Utilizing the parameter adaptation law

* þ 1Þ ¼ yðkÞ
yðk *
DtGWm ðkÞgð’ep ðkÞ þ q’* p ðkÞ þ dqðkÞÞ

*
¼ yðkÞ
DtGWm ðkÞgð’ep ðkÞ þ q* v ðkÞ

L*qp ðkÞ þ Lns ðkÞ þ dqðkÞÞ ð18Þ

the extended error dynamics can be described by the following discrete time system:
xd ðk þ 1Þ ¼ Ad ðkÞxd ðkÞ þ Adu xd ðkÞ þ OðDt2 Þxd ðkÞ

þ Cd ðkÞy þ Gd ðkÞwd ðkÞ ð19Þ

* T ; wd ¼ ½bd nd dqd T and
where xd ¼ ½ep e’ p q* p q* v y
2 3
I Dt 0 0 0
6 7
6 7
6
Km Dt I
gDt gLDt
gDt WmT Dt 7
6 7
6 7
Ad ðkÞ ¼ 6 6 0 0 I
LDt Dt 0 7
7
6 7
6 T 7
6 0 gDt
ðKm þ gLÞDt I þ gDt
Wm Dt 7
4 5
0
GWm gDt GWm gLDt
GWm gDt I

2 3
0 0 0 0 0
6 7
6 7
6
RDt
ðD þ 2OÞDt 0 0 07
6 7
6 7
Adu ¼6
6 0 0 0 0 077
6 7
6 7
6 RDt ðD þ 2OÞDt 0 0 07
4 5
0 0 0 0 0

2 3
0 0 0
6 7
6 7
6 I
gL
g 7
6 7
6 7
Gd ðkÞ ¼ 6
6 0 L 0 7
7
6 7
6 7
6
I gL g 7
4 5
0
GWm ðkÞgL
GWm ðkÞg

2 3
CðkÞ
Cd ðkÞ ¼ 4 5; wd ðkÞ  ð0; Sd Þ
0

Copyright # 2005 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2005; 19:485–503
494 S. PARK AND R. HOROWITZ

where Sd ¼ diagðSc Dt; Squan Dt2 Þ; Sc ¼ diagfSb ; Sp g and Squan is the power spectral density of the
quantization noise for dq:
The higher order term, OðDt2 Þ; is composed of O1 ðDt2 Þ and O2 ðDt2 Þ; and is given by
2 3
o1 o2
OðDt2 Þ ¼ 4 5Dt2 þ    Dt3 þ   
0 0

o1 ¼ 12 ðAc þ Acu ÞððAc þ Acu Þ þ Bc ½0
ggL
gÞ ð20Þ

o2 ¼ 12 ðAc þ Acu ÞBc WmT ðkÞ

In order to prove the stability of the closed-loop system, we consider the stochastic expectation
propagation of Equation (19). Since the propagation equation has the same form as its
deterministic counterpart, we can consider the deterministic case, i.e. wd ðkÞ ¼ 0: Since Ad ðkÞ in
Equation (19) is a periodic discrete time-varying matrix with period kT ¼ 4p2 =ðo1 o2 DtÞ; where
o1 and o2 are the model reference frequencies, we can utilize a similar analytical procedure to be
the one that was employed in Reference [5] to analyse the stability of the continuous time
varying error dynamics. To do that, we consider a discrete time version of Floquet theory [5, 7].

Lemma (Discrete time Floquet theory)

Consider the following discrete time periodic time-varying linear system with period kT
xðk þ 1Þ ¼ AðkÞxðkÞ; where AðkÞ ¼ Aðk þ kT Þ
where AðkÞ is a non-singular matrix. Then, there exists a periodic transformation which converts
the periodic time-varying linear system to a time invariant linear system.
Now, we are ready to prove the stability of the error dynamics of Equation (19).

Theorem 3 (Stability)
Consider the adaptive control law and adaptation laws, given by Equations (10) and (12), and
the adaptive observer (8), which result in the error dynamics (19). If the sampling time Dt is
sufficiently small, it is always possible to choose an observer gain L which will result in all error
signals, ep ; e’ p ; q* p ; q* v and y* being bounded. These bounds are a function of the discretization
error introduced by sampling and the zero-order hold.

Proof
Let FðkÞ be a state transition matrix associated with the periodic matrix Ad ðkÞ in Equation (19), i.e.
Fd ðk þ 1Þ ¼ Ad ðkÞFd ðkÞ ð21Þ
then the state transition matrix can be written as a product of two matrices as
% kd
Fd ðkÞ ¼ Fd ðkÞA
where Fd ðkÞ is a discrete periodic nonsingular matrix with period kT which satisfies condition,
% d is a constant matrix and the stability of the linear time varying known
Fd ð0Þ ¼ Fd ðkT Þ ¼ I: A
dynamics
xdn ðk þ 1Þ ¼ Ad ðkÞxdn ðkÞ

Copyright # 2005 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2005; 19:485–503
 DISCRETE TIME ADAPTIVE CONTROL FOR A MEMS GYROSCOPE 495

is determined by eigenvalues of A % d : Similarly to the continuous time case, an appropriate

observer gain L always exists such that the matrix A % d will be stable. In order to determine Fd ðkÞ;
the state transition matrix Fd ðkÞ must be computed. Unfortunately, it is hard to analytically
compute Fd ðkÞ: Instead, the transition matrix at the end of one period is numerically computed
from Equation (21) utilizing the initial condition Fd ð0Þ ¼ I and A % d is obtained from A%d ¼
F
kT ðkT Þ:
The remainder of the stability proof is based on the Lyapunov function approach. Consider
the following Lyapunov function candidate:
VðkÞ ¼ xTd ðkÞFd
T ðkÞMFd
1 ðkÞxd ðkÞ
where M is the solution of Lyapunov function, A % Td M A
% d
M ¼
I: Since A
% d is asymptotically

T
1
stable, M > 0 and Fd ðkÞMFd ðkÞ > 0 for all k50: Using the following relationships,
A% d ¼ Fd
1 ðk þ 1ÞAd ðkÞFd ðkÞ; and calculating the difference of V along the trajectory (19) gives
DVðkÞ ¼ Vðk þ 1Þ
VðkÞ

4
a2min jjxd jj2 þ a2max b2 lmax ðMÞjjxd jj2

þ 2a2max eblmax ðMÞjjxd jj2

þ a2max Z2 lmax ðMÞjjyjj2

þ 2a2max Zelmax ðMÞjjxd jj jjyjj
þ 2a2max bZlmax ðMÞjjxd jj jjyjj
where amin ¼ min04k4kT jjF
1 ðkÞjj; amax ¼ max04k4kT jjF
1 ðkÞjj; Z ¼ max04k4kT jjCd ðkÞjj;
b ¼ jjAdu þ OðDt2 Þjj; e ¼ jjA % d jj; and lmax ðMÞ is a maximum eigenvalue of M: If the signal
discretization error DWm ðtÞ ¼ H½Wm ðkÞ
Wm ðtÞ is zero, or equally Z ¼ 0; the origin of the state
* is guaranteed to be locally uniformly exponentially stable within the
space ðep ; e’ p ; q* p ; q* v ; yÞ
2 2
domain of amin > amax lmax ðMÞbðb þ 2eÞ: Since some amount of discretization error is always
present, the trajectory, trajectory estimation and parameter estimation errors will not converge
to zero, but rather to a compact residual set. This residual set is defined by
8 0 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi19
< s2 þ s22 þ s1 s3 =
@ ¼ xd : jjxd jj4Zjjyjj @ A
: s1 ;

where
s1 ¼ a2min
a2max lmax ðMÞbðb þ 2eÞ
s2 ¼ a2max lmax ðMÞðb þ eÞ; s3 ¼ a2max lmax ðMÞ
and the theorem is proved. &

The bound of Dt for the analysis of Theorem 3 to be valid may be computed from the
condition of s1 > 0; i.e.
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
jjAcu jj2 þ 4jjOðÞjjb * n

jjAcu jj bn
Dt5  ð22Þ
*
2jjOðÞjj jjAcu jj

Copyright # 2005 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2005; 19:485–503
496 S. PARK AND R. HOROWITZ

where
a2min *
bn  2
; OðDt2 Þ ¼ Dt2 OðÞ
2eamax lmax ðMÞ

5. PERFORMANCE ANALYSIS

5.1. Magnitude of parameter estimation biases

As discussed in the stability analysis, the introduction of the signal discretization error
function DWm ðtÞ; which was defined in Equation (13), prevents the error dynamics
from asymptotically converging to zero and introduces bias in the estimates of the angular
rate and fabrication imperfections. In order to reduce bias of the estimates, it is necessary to
make this discretization error small by achieving a fast computation rate in the feedforward
control path. To choose an appropriate computation rate in the feedforward control path, it is
necessary to determine the relationship between the computation rate and the magnitude of the
bias estimate.
Consider the signal discretization error term CðkÞy in Equation (17) and its average power. If
its power is equal to the power spectral density of a fictitious noise term Gc wðkÞ; where
2 3
0 0
6 7
6 7 2 3
6 I
gL 7 bd
6 7
Gc ¼ 66
7; w ¼ 4 5  ð0; Sc DtÞ
7
6 0 L 7 nd
6 7
4 5

I gL
where Sc ¼ diagfSb ; Sp g; the effect of the signal discretization error on the error dynamics of
ep ; e’ p ; q* p ; q* v and y* will be same to that of the noise in the stochastic average sense. Thus,
AVGðCðkÞyyT C T ðkÞÞ ¼ Gc Sc GTc Dt ð23Þ
Equation (23) can be used to determine the following bound:
maxjjCðkÞyjj2 ¼ jjGc jj2 jjSc jjDt ð24Þ
Using Euler’s numerical integration approximation, jjCðkÞyjj can be further approximated as
follows:
jjCðkÞyjj4jjBc jj jjDWmT ðtÞyjjmax Dt ð25Þ
The right-hand side of Equation (25) can be further expanded as follows:
jjDWmT ðtÞyjj ¼ jjðD þ 2OÞD’qm ðtÞ þ RDqm ðtÞjj

4 jjD’qm ðtÞjj jjD þ 2Ojj þ jjDqm ðtÞjjRjj

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 
4 Dt 2 4 2 4 2 2
X0 o1 þ Y0 o2 jjD þ 2Ojj þ X0 o1 þ Y0 o2 jjRjj 2 2 ð26Þ

Copyright # 2005 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2005; 19:485–503
 DISCRETE TIME ADAPTIVE CONTROL FOR A MEMS GYROSCOPE 497

Under the decoupling condition of the angular estimate and fabrication imperfection estimates
[5], i.e. X0 o1 ¼ Y0 o2 ; Equation (24) becomes
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2Dt2 X0 o1 ðo1 jjD þ 2Ojj þ jjRjjÞ ¼ jjGc jj DtjjSc jj
Therefore, the required sampling time to achieve parameter estimation bias magnitude below or
equal to the Brownian and position measurement noise floor is given by
pffiffiffiffiffiffiffiffiffi !2=3
jjGc jj jjSc jj
Dt ¼ ð27Þ
2X0 o1 ðo1 jjD þ 2Ojj þ jjRjjÞ
where D is the damping matrix, R is the resonant frequency modelling error matrix, and O is the
unknown ‘input’ angular rate. Equation (27) is a useful criterion for choosing an approximate
computation rate for the feedforward control path. Equation (27) suggests that the magnitude
of biases of the angular rate and fabrication imperfection estimates is proportional to the
computation rate elevated to the 32 as shown in Figure 2. However, it is important to note that
‘magnitude’ in this context means the norm of a vector composed of the angular rate and
fabrication imperfection estimate errors. Thus, individual biases such as the angular rate
estimation bias, may be equal to or less than this estimate.
When the sampling rate of the discrete time control algorithm in Equations (10) and (12) is
increased significantly, the computational burden of the control algorithm may become too
high, and may cause unnecessarily over-sampling in the parameter adaptation algorithm. This
analysis suggests the introduction of a two-rate discrete time control as a compromise between

Computation Rate vs Biases in Angular Rate Estimation

101
Biases in Angular Rate Estimation (deg/sec)

x : simulation

t 3/2 line

100

10-1

position sensing noise floor

10-2

105 106
Computation Rate in Feed forward Control (Hz)
Figure 2. Computation rate in the feed forward control path vs bias in the parameter estimations.

Copyright # 2005 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2005; 19:485–503
498 S. PARK AND R. HOROWITZ

minimizing bias estimate due to discretization errors and attaining an implemental controller
computational burden. In this two-rate discrete time controller, the parameter adaptation
algorithm is updated at a slower rate than the reference signal which is used in the feedforward
control.

5.2. Convergent rate analysis

In order to apply averaging analysis, consider the error dynamics given by Equations (16) and
(18). We will ignore the signal discretization error in the feedforward control path which causes
the bias of the states. As in the continuous time counter part, we first assume that all noise
terms are zero. If we do a state co-ordinate transformation from xc ¼ ½ep e’ p q* p q* v T to
xa ¼ ½ep e’ p q* p q’* p T ; then
x’ a ¼ Aa xa þ Acu xa

*

Bc gH½’ep ðkÞ þ q’* p ðkÞ þ Bc H½WmT ðkÞyðkÞ ð28Þ

* þ 1Þ ¼ yðkÞ
yðk *
DtGWm ðkÞgð’ep ðkÞ þ q’* p ðkÞÞ
where
2 3 2 3
0 I 0 0 0
6 7 6 7
6
K 0 0 0 7 6 I 7
6 m 7 6 7
Aa ¼ 6
6
7;
7 Bc ¼ 6
6
7
7
6 0 0 0 I 7 6 0 7
4 5 4 5
0 0
Km
L
I
We also assume that sampling rate is small so that the exponential matrix can be approximated
by the first-order Taylor expansion eAa Dt I þ Aa Dt: After discretizing Equation (28), we obtain
*
xa ðk þ 1Þ ¼ Ab xa ðkÞ þ Abu xa ðkÞ þ Bb H½W T ðkÞyðkÞ m

* þ 1Þ ¼ yðkÞ
yðk *
DtGWm ðkÞgð’ep ðkÞ þ q’* p ðkÞÞ ð29Þ
where
Ab expðAa* DtÞ; Abu ¼ Acu Dt

Z Dt
Bb  expðAa* ðDt
sÞÞBc ds
0

2 3
0 I 0 0
6 7
6
K
g 0
g 7
6 m 7
Aa* ¼ 6
6
7
7
6 0 0 0 I 7
4 5
0 g
Km g
L

Copyright # 2005 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2005; 19:485–503
 DISCRETE TIME ADAPTIVE CONTROL FOR A MEMS GYROSCOPE 499

Using the discrete time version of averaging technique [7], the average dynamics of the
parameter estimation error equation in Equation (29) is given by
y* avg ðk þ 1ÞðI
DtGAVGðWm ðkÞM# d ðW T ðkÞÞÞy* avg ðkÞ
m ð30Þ
# d is the transfer function matrix,
where M
M# d ðzÞ ¼ Cb ðzI
Ab Þ
1 Bb ; Cb ¼ ½0 g 0 g ð31Þ
and z is the Z-transform variable. Equation (30) is the sampled and zero-order hold input
# d in Equation (31)
version of the continuous time result obtained in Reference [5]. Therefore, M
is given by
 
sDt

#
1
1 1
e #
M d ðzÞ ¼ ð1
z ÞZ L M o ðsÞ
s
# o ðsÞ is given
where L
1 and Z; respectively, denote the inverse Laplace and Z-transforms, and M
# # #
in Reference [5] as M o ðsÞ ¼ diagfM o1 ðsÞ; M o2 ðsÞg; where

# 01 ðsÞ ¼ g1 L1 s2
M
S 4 þ L1 s3 þ ð2o21 þ g1 L1 Þs2 þ o21 L1 s þ o41

# 02 ðsÞ ¼ g2 L2 s2
M
S4 þ L2 s3 þ ð2o22 þ g2 L2 Þs2 þ o22 L2 s þ o42
Since the sampling frequency ð2p=DtÞ is larger than the reference resonant frequencies o1 and
o2 ; the filtered steady state response through M # d ðW T Þ is given by
m
2 3T
X0 sinðo1 kÞ 0
6 7
6 7
6 A1 Y0 sinðo2 k þ f1 Þ A X sinðo k þ f Þ 7
6 2 0 1 2 7
6 7
6 7
6 0 Y sinðo kÞ 7
6 0 2 7
6 7
6 7
M# d ðW T Þ ¼ 6 X o cosðo kÞ 0 7 ð32Þ
m 6 0 1 1 7
6 7
6 7
6 A1 Y0 o2 cosðo2 k þ f Þ A2 X0 o1 cosðo1 k þ f2 Þ 7
6 1 7
6 7
6 7
6 0 Y0 o2 cosðo2 kÞ 7
6 7
4 5

2A1 Y0 o2 cosðo2 k þ f1 Þ 2A2 X0 o1 cosðo1 k þ f2 Þ
where
g1 L1 o22
A1 ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
ðg1 L1 o22
ðo21
o22 Þ2 Þ2 þ L21 o22 ðo21
o22 Þ2

L1 o2 ðo21
o22 Þ
f1 ¼ tan
1
g1 L1 o22
ðo21
o22 Þ2

Copyright # 2005 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2005; 19:485–503
500 S. PARK AND R. HOROWITZ

g2 L2 o21
A2 ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
ðg2 L2 o21
ðo21
o22 Þ2 Þ2 þ L22 o21 ðo21
o22 Þ2

L2 o1 ðo21
o22 Þ

f2 ¼ tan
1
g2 L2 o21
ðo21
o22 Þ2

Note that Equation (32) has the same form as its continuous time counterpart equation in
Reference [5]. Therefore, every convergence property mentioned in the continuous time case is
preserved here.
Applying the decoupling condition X0 o1 ¼ Y0 o2 ; the bandwidth of the adaptive controlled
gyroscope is approximately given by
O* zavg ðk þ 1Þð1
2DtgO X 2 o2 ÞO
* zavg ðkÞ ð33Þ
0 1

or
’* 2 2*
O zavg 
2gO X0 o1 Ozavg

since the parameter estimation dynamics is much slower than the sampling rate. This is exactly
same result that was obtained for the continuous time observer-based adaptive control case,
and the bandwidth of the adaptive controlled gyroscope is also approximately given by
BW2gO X02 o21 : Thus, the bandwidth of a MEMS gyroscope under the observer-based discrete
time adaptive control is also proportional to the adaptation gain gO and the energy of oscillation
of the reference model. Other statements made in Reference [5] regarding the comparison
between the analytical convergence rate of the angular rate estimate and simulation results for
various resonant frequency ratios, and control and observer gains are also valid.

5.3. Resolution analysis

The resolution analysis for this hybrid control system that will be presented in this section is
very similar to those in Reference [5]. The error expectation propagation of Equation (19) is
given by
E½xd ðk þ 1Þ ¼ Ad ðkÞE½xd ðkÞ þ Adu E½xd ðkÞ þ Cd ðkÞy ð34Þ
where E½ denotes stochastic expectation. Notice that the expectation equation has the same
form as its deterministic counterpart. Therefore, the mean trajectory under a stochastic
environment is also biased because of the Cd ðkÞy term. Defining covariance as Pd ¼
E½ðxd
E½xd Þðxd
E½xd ÞT ; the covariance propagation equation is given by
Pd ðk þ 1ÞAd ðkÞPd ðkÞATd ðkÞ þ Gd ðkÞSd GTd ðkÞ ð35Þ
The covariance Pd can easily be pre-computed independently of Equation (35). The standard
deviation of the angular rate estimate error is obtained from the covariance matrix Pd as
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
smeas ¼ HPd H T ð36Þ
where H ¼ ½0114 1: In this case, the resolution is a summation of the standard deviation of
angular rate estimate error computed from (36) and the bias in the angular rate estimate. The

Copyright # 2005 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2005; 19:485–503
 DISCRETE TIME ADAPTIVE CONTROL FOR A MEMS GYROSCOPE 501

ultimate achievable resolution can also be calculated by setting Sp ¼ 0 and computing smeas with
Equation (36). As in the convergence rate analysis presented in Reference [5], the same results
regarding the effects of various design parameters such as control gains and parameter
adaptation gains on the variance of the angular rate estimate error, can be stated in the discrete
time case. In summary, the resolution of the gyroscope can be adjusted independently by the
angular rate adaptation gain, without affecting the other fabrication imperfection estimation
dynamics.
The effect of the quantization noise dqd ðkÞ on the velocity estimation may be explored in a
similar way as the analysis used in Section 5.1. Other noises which might be caused by the
interface connection between digital processor and a MEMS gyroscope can be included in
position measurement noise. Overall, gyroscope resolution can be estimated by following
formula:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
resolution ¼ s2meas þ s2quan ð37Þ

where s2meas is a variance of the angular rate estimate error due to measurement noise, and s2quan
is a variance of the angular rate estimate error due to quantization noise.

6. SIMULATIONS

A computer simulation study was conducted using the preliminary design data of the MIT-SOI
MEMS gyroscope, to test the analytical results presented in this paper and verify its predicted
performance. The data of some of the gyroscope parameters is given by Table I and is same as
that used in Reference [5]. The numerical values for the controller used in the simulation are:
g ¼ 1; gR ¼ 1=5; gD ¼ 1=10; gO ¼ 1=50; and L ¼ diagf1; 1g: Notice that these numerical values
are shown in non-dimensional units, which are non-dimensionalized based on the proof-mass,
length of one micron and the x-axis nominal natural frequency.
The estimate of the angular rate response to the step input angular rate is shown in Figure 3.
In this figure, the upper and lower bounds, which corresponds to the analytically estimated
standard deviation calculated by Equation (37) are also plotted. Figure 4 shows the estimate of
angular rate response to the sinusoidal input angular rate. These simulation results support the
theoretical results obtained in the previous section, regarding predicted gyroscope bandwidth
and resolution.

Table I. Key parameters of the gyroscope.

Parameter Value
Mass 5:095  10
7 kg
x-axis frequency 4:17 kHz
y-axis frequency 5:11 kHz
Quality factor 104
Brownian noise PSD 1:47  10
26 N2 s
Position noise PSD 1:49  10
27 m2 s

Copyright # 2005 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2005; 19:485–503
502 S. PARK AND R. HOROWITZ

Angular Rate Response

8
5 deg/sec step input
7 estimate

4
deg/sec

Analytic Bound
3

-1

-2
0.018 0.02 0.022 0.024 0.026 0.028 0.03 0.032 0.034 0.036
sec
Figure 3. Time response of angular rate estimate to the 58=s step input.

Angular Rate Response

8
5 deg/sec sinusoid in 50 Hz
6 estimate

2
deg/sec

-2

-4

-6

-8
0.015 0.02 0.025 0.03 0.035 0.04 0.045
sec
Figure 4. Time response of angular rate estimate to the 58=s sinusoid input at 50 Hz:

Copyright # 2005 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2005; 19:485–503
 DISCRETE TIME ADAPTIVE CONTROL FOR A MEMS GYROSCOPE 503

7. CONCLUSIONS

A new adaptive control algorithm developed in Reference [5] offers several advantages such as a
larger operation bandwidth, absence of zero-rate output, self-calibration and a large robustness
to parameter variations. However, it is a continuous time controller. In this paper, a discrete
time version of this control system was developed, which can be implemented using digital
processors, and its stability was proven. The convergence and stochastic analysis presented in
this paper showed that all convergence and stochastic properties of a continuous time adaptive
control were preserved.
However, a stochastic analysis showed that the estimates of the angular rate and the
fabrication imperfections are biased due to the signal discretization errors in the feedforward
control path introduced by the sampler and holder. Thus, a two-rate discrete time control was
proposed as a compromise between the measurement biases due to discretization errors and the
computational burden imposed on the controller due to a fast sampling rate. In this control
scheme, the parameter adaptation algorithm is updated at a slower rate than the reference signal
which is used in the feedforward control. An analysis method was also developed for
determining the trade-off between the controller sampling frequency and the magnitude of the
angular rate estimate biased errors.
A simulation study using the preliminary design data of the MIT-SOI MEMS gyroscope was
conducted, to test the analytical results derived in this paper and to verify the predicted
performance of the proposed controlled schemes. Simulation results were in strong agreement
with the analytically derived predicted results and performance estimates.

ACKNOWLEDGEMENTS

This research was supported by DARPA under Contract N66001-97-C-8643.

REFERENCES
1. Yazdi N, Ayazi F, Najafi K. Micromachined inertial sensors. Proceedings of IEEE 1998; 86(8):1640–1659.
2. Shkel A, Howe RT, Horowitz R. Modeling and simulation of micromachined gyroscopes in the presence of
imperfection. International Conference on Modelling and Simulation of Microsystems, Puerto Rico, 1999; 605–608.
3. Jiang X, Seeger J, Kraft M, Boser BE. A monolithic surface micromachined z-axis gyroscope with digital output.
IEEE 2000 Symposium on VLSI Circuits, Honolulu, HI, 2000; 16–19.
4. Chang S, Chia M, Castillo-Borelley P, Higdon W, Jiang Q, Johnson J, Obedier L, Putty M, Shi Q, Sparks D,
Zarabadi S. An electroformed CMOS integrated angular rate sensor. Sensors and Actuators 1998; A66:138–143.
5. Park S, Horowitz R. New adaptive mode of operation for MEMS gyroscopes. ASME Journal of Dynamic Systems,
Measurement and Controls 2004; 126(4):708–718.
6. Juneau TN. Micromachined dual input axis rate gyroscope. Ph.D. Thesis, U.C. Berkeley, 1997.
7. Bai EW, Fu LC, Sastry SS. Averaging analysis for discrete time and sampled data adaptive systems. IEEE
Transactions on Circuits and Systems 1998; 35(2):137–148.

Copyright # 2005 John Wiley & Sons, Ltd. Int. J. Adapt. Control Signal Process. 2005; 19:485–503