IEEE SENSORS JOURNAL, VOL. 5, NO.

5, OCTOBER 2005

1019

Analyzing and Simulation of MEMS in VHDL-AMS

Based on Reduced-Order FE Models
Michael Schlegel, Fouad Bennini, Jan E. Mehner, Gran Herrmann, Dietmar Mller, and Wolfram Dtzel

AbstractThis paper deals with the computer-aided generation

of reduced-order macromodels (ROMs) for system level simulation using VHDL-AMS. The focus is thereby set on the application of ROMs in system simulation. It also gives an survey of our
approach of reduced-order modeling.VHDL-AMS is used to describe microelectromechanical systems based on lumped and rigid
bodies for several years. But such models often have a limited accuracy if the system contains flexible bodies or nonhomogeneous electrical fields. Methods of reduced-order modeling of linear systems
to increase the system model accuracy are also available for some
years. The work presented in this paper combines the advantages
of reduced-order modeling and VHDL-AMS system level simulation, whereas our approch of reduced-order modeling is able to
handle nonlinear systems too.The application of ROMs on system
level and their export to VHDL-AMS have been demonstrated by
two surface micromachined vibration sensor arrays which are intended for wear state recognition. Additionally, a new approach
called multi architecture modeling has been applied which simplifies the integration of macromodels into the system. As a result
of this work, we are able to simulate certain classes of nonlinear
ROMs with an error less than 1% compared with the finite-element method simulation. Also, simulation time of ROMs is shorter
than simulation time of models based on lumped and rigid bodies.
Index TermsReduced-order modeling, sensor array, system
model, VHDL-AMS.

I. INTRODUCTION

EFORE constructing a prototype, a simulation of the

whole system is necessary to check the functionality of
the individual components and their interaction. For microelectromechanical systems (MEMS), a design environment is
needed which allows simulation of different physical domains
including coupling effects. VHDL-AMS is a flexible system
description language suited to handle such requirements. Additionally, it allows to describe and simulate the system at
different levels of abstraction.
System level modeling and simulation have become state of
the art in MEMS design due to increasing system complexity.
But high-level models have the disadvantage that they often
cause a considerable error in simulation. This error occurs because physical phenomena must be simplified to be expressed

Manuscript received December 17, 2003; revised April 26, 2004. This work
was supported by Projects A1 Component Design and A2 System Design
of the SFB 379 (collaborative research center), German Science Foundation
(DFG). The associate editor coordinating the review of this paper and approving
it for publication was Dr. Jan Soderkvist.
The authors are with the Department of Electrical Engineering and
Information Technology, Chemnitz University of Technology, Chemnitz
09126, Germany (e-mail: [email protected];
[email protected]; [email protected]; [email protected]; [email protected];
[email protected]).
Digital Object Identifier 10.1109/JSEN.2004.841445

in an equation. VHDL-AMS or similar simulation systems can

only solve differential algebraic equations and no partial differential equations. If the partial differential equations cannot
be solved analytically then an approximated solution must be
used which increases the error additionally. But using accurate models from the physical level [e.g., finite-element method
(FEM) models] is very cumbersome for daily tasks on system
level.
Parameter reduction of FEM models (reduced-order modeling) is an approach filling this gap in the design flow. Using
shape function methods to generate reduced-order macromodels (ROMs) of linear systems becomes state of the art.
Common engineering approaches are Krylov-Subspace for
first-order systems [1] and the modal superposition method for
second-order systems [2]. Using the Krylov-Subspace enforces
the transformation of the governing equation of motion in a
first-order system. This leads to doubling the dimension of the
original system and to ill conditioned matrices especially by
microstructures where the mass and stiffness matrices are much
different in their magnitudes. Recently presented in [3], the
first moments of the governing equation can be used directly
to compute the shape function for the ROM, for example, via
the block Arnoldi algorithm. However, most of MEMS show a
nonlinear behavior. In contrast to the linear case there exists no
formal rule for nonlinear-order reduction. The main problem
is that nonlinear system matrices are not directly available.
The application of the shape function method to generate
macromodels of nonlinear coupled systems assumes that the
nonlinear deformation state of the structure must be mapped
correctly by the involved shape functions in the operating
range of the system. The most promising approach to capture
nonlinear effects is to collect several snapshots of the system
response by superposing the shape functions represented by
Eigenvectors and running a series of full-order finite element
simulations [4], [5]. Further developments are needed to speed
up the generation pass of the ROM, which are discussed in the
following.
For a most effective usage of these methods in system design,
it is required to export the ROM to VHDL-AMS with an interface matching the system structure.
II. SYSTEM EXPERIMENTAL PROTOTYPE
VIBRATION SENSOR ARRAY
Wear state recognition on machine tools during the normal
work process is an important fundamental of product quality improvement. Today this is done by measuring the forces on the
tool. An alternative method is to control machine and tool vibrations. Approaches to tool vibration measurement have existed
for many years. But these approaches are using piezo ceramic
wide band sensors which are quite expensive [6], [7].

1530-437X/$20.00 2005 IEEE

1020

IEEE SENSORS JOURNAL, VOL. 5, NO. 5, OCTOBER 2005

Fig. 3. Extracted detail of a mass-spring resonator.

Fig. 1. System structure of the measurement system.

Fig. 4. Schematic of the sensor array.

The high-voltage amplifier generates voltages up to 30 V for

the natural frequency tuning of the sensor. A micro controller
starts or stops the measurement, activates or deactivates resonators and starts self-calibration. It tunes the natural frequency
of the resonators and transmits measured data to the classification unit. The classifier decides by a fuzzy pattern classification
algorithm whether the data are produced by an sharp or worn
out tool. This algorithm is realized in hardware using an FPGA.
Fig. 2.

Photograph of the vibration sensor.

The new sensor array which is being developed at Chemnitz

University of Technology is fabricated by SCREAM technology
[8]. Because of the narrow-band resonance working mode, the
signal processing can be simplified due to a better signal to noise
ratio and because no fast Fourier transform is needed. So a lowcost sensor system can be built.
The sensor system consists of a sensor array, an analog signal
processing unit and a high-voltage amplifier. The system is controlled by a micro controller. It also includes a fuzzy pattern
classification system [9].
An environmental model called virtual machine tool provides the stimuli for simulation. It reproduces datameasured
at a real machine toolin time or frequency domain or generates fictitious data. The center of the measuring system is the
vibration sensor array. This surface micromachined vibration
sensor consists of eight laterally moving mass-spring resonators
and is intended to work as a frequency selective device at its resonance (see Fig. 1).
Using the stress stiffening effect in the flexures (type A), the
natural frequency is tuned by a voltage
controlled tensile
force, which is electrostatically generated in pairwise arranged
lever systems. Vibration detection is effected electrostatically by
the symmetrical comb capacitor attached to the proof mass were
a dc voltage can be induced. Type B has a similar structure. But
in opposition to type A the tuning is performed by an additional
curved comb structure at the mass. This structure behaves like
a spring with negative coefficient, so the natural frequency can
be tuned directly. Fig. 2 shows both types of sensors.

III. ABSTRACT MODELS OF THE SENSOR

The first VHDL-AMS models of the sensor are based on rigid
and lumped bodies, as described several times in MEMS publications, e.g., [10][12]. So only a brief summary of these models
is given here.
In the version A, the tuning of the natural frequencies is done
by using the stress-stiffening effect. Fig. 3 shows the functional
principle of the mass-spring system and the natural frequency
tuning by the stress-stiffening effect. Fig. 4 shows a schematic
of the model. It consists of eight springs, seismic masses and
dampers.
The stress-stiffening unit is driven by an electrostatic force
. This force is caused by the voltage
is calculated as
(1)
The capacitance of the comb structure can be calculated by
the simplification of a homogeneous field of a plate capacitor
(2)
(3)
where is the number of combs. The force
lever mechanism
a normal force
[8]

generates by a

(4)

SCHLEGEL et al.: ANALYZING AND SIMULATION OF MEMS IN VHDL-AMS

1021

The modeling of the sensor of type B and the displacementcurrent transducer is similar and described in [13]. The simplifications of the physical behavior cause a simulation error of
up to 20%. This is accurate enough to check the global system
functionality. But for more significant simulation results, e.g.,
to test the fuzzy pattern classifier, a more accurate sensor model
is required. To get these more accurate models we used an approach of reduced-order modeling which has been developed
within our collaborative research center (SFB 379). A survey of
this approach is given in Section IV.
IV. REDUCED-ORDER MODELING
Reduced-order modeling using modal basis functions was
originally developed by [4] and has been continuously improved
by several authors. In [5], we have shown that the approach
is able to cover electrostatic structural interactions of arbitrary
two-dimensional (2-D) and three-dimensional (3-D) structures,
allowing multiple electrodes, geometrical nonlinearities and initial pre-stress conditions while supporting static, transient and
harmonic analyzes.
The basic idea of the ROM is to approximate the deformation
state of the finite element model by a series of weighted modal
shape functions (eigenmodes)
(5)
where
are the time dependent nodal displacements of the
FE-model and
the eigenmodes which are scaled by time
dependent modal amplitude . In general, (5) describes a coordinate transformation of finite element displacement coordinates (local coordinate) to modal coordinates of the macromodel
(global coordinate). The deformation state of the structure given
is now repreby nodal displacements
sented by a linear combination of
modes weighted by their
amplitudes
where
.
The governing equation of motion describing the ROM of
electrostatic actuated MEMS structures in modal coordinates is

(6)
where
is the modal mass,
is the eigenfrequency,
is
the linear modal damping ratio,
is the modal strain energy function,
is the modal capacity-stroke function, is
the number of capacities involved for microsystems with multiple electrodes, is the electrode voltage applied, and
is a
local force acting at the th node. The current at each electrode
is defined by

(7)
Essential prerequisites to establish (6) and (7) are proper
modal strain energy and capacity-stroke functions. Both are

derived from a series of FE runs at various deflection states in

the operating range. The acquired data are fitted to polynomial
functions in order to compute the local derivatives, which
describe force and stiffness terms. As a matter of fact, shape
function methods can be applied to nonlinear systems, too [14].
Geometric nonlinearities, as for instance, stress-stiffening,
can be regarded if the modal stiffness is computed from the
first derivative of the strain energy function with respect to
the modal amplitudes. Capacitance-stroke functions provide
nonlinear coupling between each eigenmode and the electrical
quantities (i.e., electrostatic modal forces, electrical current) if
stroke is understood as modal amplitude. Damping parameters
are assigned to each eigenmode. Modal representations of
MEMS are very efficient since just one equation per mode and
one equation per involved conductor is necessary to describe
the coupled system entirely.
The process filling the table from the series of FE runs is computationally expensive but has to be done just once. The result
is a black-box model that can be applied to any load situation.
However, it has been found that the distinction of dominant and
relevant mode shapes speeds up this process considerably. Dominant modes are characterized by large amplitude. Their interactions to all other mode shapes, dominant and relevant, are regarded throughout. Relevant modes contribute to the final solution but do not affect among each other or it is neglected. Consequently, the multivariable function fit can be reduced to a series
of functions with a lower number of variables. It turned out that
two dominant modes are sufficient for most applications. The
polynomials can then be described by the following series representation of functions with three variables

(8)
Fig. 5 shows the significance of this approximation. The stiffness matrix is computed from the second derivative of the function , respectively, to the modal amplitudes, where all couple
terms between relevant modes are set by this approximation to
zero. In the transfer function the stimulation between relevant
modes are neglected.
This approximation yields to insignificant errors. Remarkable
is that not only the number of polynomial coefficients can be
reduced but also the number of sample points. For example, if
the nonlinear strain energy is computed for five mode shapes
and six modal amplitudes of each mode, the number of data
points would be 6
7776 compared to 3
648 when
two modes are classified as dominant and the three others as
relevant. Furthermore, the fit is limited to functions with three
variables, which allows to use simple and fast algorithms like
the well known least square fit. Depending on the FE model size
and the number of mode shapes that are included in the ROM
procedure, the data acquisition is typically an over-night job.
The approach will be demonstrated at the example of one cell
of the micromechanical vibration sensor array from type A.
The first step of the ROM generation is to determine which
modes are really significant, and to estimate a proper amplitude range for each mode. Several criteria can be applied, for
instance, the lowest eigenmodes of a modal analysis, modes in
operating direction, or modes, which contribute to the deflection
state at a typical test load.

1022

IEEE SENSORS JOURNAL, VOL. 5, NO. 5, OCTOBER 2005

Fig. 6. Harmonic response of a micromechanical vibration cell from type A

for several tuning voltages.

gravitation, pressure) and significant outputs (e.g., a subset of

displacements at characteristic points of the model).
A. Problems During VHDL-AMS Export

Fig. 5.
(Top) Considering and (bottom) neglecting interaction between
relevant modes through successive polynome construction.

Applying a unit tuning voltage

and an acceleration in
operating direction as test loads to one cell of the array reveals
that its motion is dominated by the detection mode of the proof
mass (mode 1), the tuning mode of the pairwise arranged lever
systems (mode 5) and a higher mode (mode 8) which corrects
the displacement of mode 5 caused by the imposed acceleration.
Next, the dependencies of the strain energy
and of two capacities (for detection and tuning) on the modal amplitudes are
described by a fitted polynomial function. The necessary data
points are obtained by imposing each eigenmode with varying
amplitude on the mechanical model for the nonlinear strain energy and on an electrostatic space model for capacitance. In the
concept of the modal superposition method, each eigenmode
represents a single independent resonator with modal mass
and modal damping .
Fig. 6 shows the harmonic response of the proof mass excited
by acceleration in detection direction and for several tuning voltages. The results were compared with the coupled field analysis
with ANSYS. It was clearly shown that the ROM is able to capture correctly the dynamic behavior of the full FE-Model and
the stress stiffening effect (error less than 2%). Analogously,
the ROM of micromechanical vibration cells from type B were
generated only with two eigenmodes, with errors less than 1%.
V. APPLICATION OF ROMS IN SYSTEM
DESIGN WITH VHDL-AMS
The ultimate goal of reduced-order modeling is to obtain an
accurate black-box model of the microsystems behavior. Interface objects are limited to the voltage-current relationship
at each electrode, essential inputs such as external loads (e.g.,

The export of the ROM to VHDL-AMS is performed in two

steps. At first, an initialization file containing all necessary information of the macromodel, such as the fitted polynomial coefficients and orders, is generated. Then, the source code in
VHDL-AMS is generated automatically. Thereby, a set of differential algebraic equations (DAE) with nonconstant, nonlinear
coefficients emerges. The DAE can be mapped to the simultaneous statements of VHDL-AMS where nonconservative nodes
represent forces, displacements, velocities, and accelerations.
The main problems of exporting the ROM in VHDL-AMS
occur due to incomplete implementations of the VHDL-AMS
simulators. So this section describes some hints and traps when
exporting to VHDL-AMS.
The main question is how to express the fitted functions of
the nonlinear strain energy and of the capacities, which are part
of the coefficients of the DAE (derivatives of the energy
and of the capacities
,
). The algorithm consists
of a number of sequential calculations. So this algorithm must
be calculated within the VHDL-AMS sequential structures
or
.A
cannot be used because it is a digital statement. The analog
is the
statement. But
counterpart of a
this statement was not implemented in the VHDL-AMS simulators, which were available at the beginning of this work.
So this algorithm must be implemented inside a VHDL-AMS
or
. The algorithm returns more than
one result, so first an attempt was made to use a
which can return any number of values. But a
can
only be called as sequential statement within the VHDL-AMS
sequential structures described above. The so called concurrent
procedure call is not really a concurrent call. In this case, a
is placed around the procedure call. This
contains a
statement which interrupts the calculation of
the
until a digital event occurs. The parameters of
. Therefore, a digital
the procedure call are analog
event never occurs and the coefficients are only calculated at
simulation start and not during the simulation.

SCHLEGEL et al.: ANALYZING AND SIMULATION OF MEMS IN VHDL-AMS

1023

For these reasons the calculation of the coefficients (

,
,
) must take place in a
. The return
values are aggregated in a vector which is the return value of
the
. The vector is arranged in the following form:

Fig. 7. Possible problem during the conventional top-down design process due
to different interface objects.

The resulting VHDL-AMS code has the following structure.

--

The DAE, representing the movement, e.g., in , , and

direction, are described separately because the available simulators do not support a description in matrix notation properly,
as is known from MATLAB, for instance. If this feature is implemented in the VHDL-AMS simulators the exported VHDL
models will become more compact and clearer.

--

B. Using the ROM

--

Due to manufacturing problems the sensor array using tuning

by stress-stiffening effect could not be used. So the sensor with
tuning by electrostatic softening was employed. The export of
its ROM to VHDL-AMS is described above. The next step is
to insert this ROM into the system environment. Because of a
lower abstraction level of the ROM compared with the abstract
sensor model, it might happen that the abstract sensor model
uses an interface other than that of the ROM although the inputs
and outputs of the sensor are the same. This will cause a lot
of error prone work to adapt the system interface to the ROM
interface. To avoid this, it must be assured that the ROM can
use the same interface as the abstract model, where the new
methodical approach Multi Architecture Modeling has been
created.

--

-----

C. Design Method Multi Architecture Modeling

If MEMS are designed by a Top Down based method or
MEMS component models are generated by automated methods
then digital, analog electrical and nonelectrical models at different abstraction levels may appear during the design process.
A possible resulting problem of this is shown in Fig. 7. On
a high-abstraction level the interface of analog components
may use nonconservative nodes
as is common in
signal flow models. During the design process one or more of
these abstract models may be refined to a conservative system
(network, energy flow model) either manually or automatically. This new refined model has to use conservative nodes
in its interface. Due to the fact that a
cannot be connected directly to a
, it is necessary
until now to modify the interfaces of the system and component
models at every design step of a component model. In digital
components the problem is similar. On high-abstraction level
the models use clock-independend signal with real or integer
values on low-abstraction level they use clock-true bit- or
std_logic-vectors.
So an approach was created which allows to insert models of
heterogeneous components at different abstraction levels in the
same system simulation. The basic idea of the MAM is to use
the interface, which will be necessary at low-abstraction level,

1024

IEEE SENSORS JOURNAL, VOL. 5, NO. 5, OCTOBER 2005

Fig. 8. Unified interface for analog models using MAM.

already at a high-abstraction level. For analog components, in

relation to Fig. 7, this means: Because the refined model at
as interface the model
low-abstraction level needs
at high-abstraction level also has to use
instead of
(Fig. 8).
The behavior inside the abstract model is the same as it would
be without using the MAM method. In the case of an analog
interface the connection between interface and behavior is esbetablished simply by the definition of branch
tween the
and ground. In case of a digital port the
connection is made by special functions and procedures translating abstract values to clock-true bit- or std_logic-vectors and
vice versa. These functions and procedures are implemented in
a MAM supporting library. Additionally, this library provides
special A/D and D/A converters which make it possible even to
connect analog and digital representations of the same model
(e.g., an inverter described as logic function or modeled with
two MOSFETs) with its environment.
The MAM not only contains some models in a library, it also
includes a modification of the Top Down system design process.
This modified design flow using the MAM method is shown
in Fig. 9. It can be seen that the MAM method simplifies the
Top Down design process, especially if design cycles between
abstract and refined components are necessary.
Using MAM helps to avoid a lot of work because abstract
models of a component can now be replaced by detailed models
or vice versa without any modification to the interface or the
surrounding models. The MAM also allows to reconfigure a
system model within a few seconds between speed an accuracy
regarding the aim of the simulation. So even large systems can
be simulated fast on the whole and accurate in the details.
Applying this approach within the development of the sensor
models means:
1) to examine which inputs and outputs, including power
supply, the sensor will need;
2) which types of interfaces (conservative/nonconservative
analog nodes, digital nodes) the ROM will need;
3) which data types/natures these interfaces will need;
4) development of the abstract sensor model with this collected information;
5) checking the functionality of the sensor within the system
context;

Fig. 9. Simplified top-down design flow without and with using MAM.

6) development of the ROM from the sensors FEM model

observing the information collected in steps 1) to 3);
7) to insert the ROM in the system model;
8) simulation of the complete system.
Without MAM, steps 1) to 3) would drop out and step 4)
would be easier, but then step 7) would cause more work than
needed for steps 1) to 4) because the system model must be
adapted. If an error occurs after this adaptation it is not sure
whether it was caused by the ROM or the adaptation. If MAM
is used and the system model with the ROM is not working
although it does with the abstract model then the error must
be inside the ROM and the system simulation can be continued
with the abstract model until a working ROM is available. More
details on Multi Architecture Modeling are found in [15].
D. Simulation of the Sensor ROM
By using MAM, application of the ROM in the system context
is very easy. First, the eight models for the individual resonators
were tested separately to verify the method of reduced-order
modeling and the VHDL-AMS export. The results of these tests
allowed to improve the method of reduced-order modeling and
VHDL-AMS export. At the end of these tests the eight individual models were joined to the sensor ROM. Then the abstract
model was replaced by the ROM simply by changing the name
of the architecture which is to be instantiated in the VHDL-AMS
system model. No further work was necessary. A complete measurement cycle has been simulated with the ROM.
VI. RESULTS
The following simulation has been done on a SUN ULTRA60
workstation with UltraSPARC-II 296 MHz CPU. An AC-simulation with a tuning voltage of 35 V and a stimulation magnitude
of 0.6 mm led to the results shown in Tables I and II.
It was a surprise that the more accurate ROM can be simulated faster than the abstract model. This can be explained

SCHLEGEL et al.: ANALYZING AND SIMULATION OF MEMS IN VHDL-AMS

1025

TABLE I
RESULTS OF AC SIMULATION

TABLE II
RESULTS OF TRANSIENT SIMULATION

Fig. 11.

Fig. 10. Plot of system simulation using sensor ROM.

simply. The abstract model contains a set of simple equations.

The ROM contains less but more complex equations than the abstract model. Obviously, smaller numbers of complex equations
can be simulated faster than larger numbers of simple equations.
The results seem to show that the abstract model is more accurate than the ROM. But this effect occurs due to manufacturing tolerances of the real sensor. If the ROM is compared to
the FEM Model the error is less than 1% (see Table I).
The simulation results of a transient simulation for 50 ms of
the sensor and analog signal processing are shown in Table II.
Again it can be seen that the simulation with the ROM is faster
than the simulation with the abstract model. The ROM is also
more accurate.
By simulation of these system parts some side effects were
determined, e.g., a snapping of the resonators at high-tuning
voltages was taken into account when assembling the prototype.
By using Multi Architecture Modeling the system can be
reconfigured for testing purposes between abstract model and
ROM in less than 5 s.

Plot of system simulation using sensor ROM.

At last a simulation of the whole system was performed to

select stimuli patterns for the fuzzy pattern classifier. This simulation was done on a SUN Blade2 with 900 MHz CPU and 4
GByte RAM. It took about 1 h and 30 min. A simulation plot
can be seen in Figs. 10 and 11. The first curve of Fig. 10 displays a sector of the mechanical stimulation of the sensor array
containing a constant spectrum of frequencies representing machine vibrations and interferences. The second curve shows a
sector of the response of the sensor.
The first two curves in Fig. 11 display the applied tuning
voltage and the output signal of the analog signal processing
during a complete measurement cycle. The last curve shows the
data sampled by the microcontroller which are transferred to the
fuzzy pattern classifier. The different grey backgrounds of the
curves show which sensor cell of the array is activated.
A comparison of these results with an FEM simulation is
not possible because such complex systems cannot be simulated with FEM. In comparison to the manufactured system the
simulation shows an error of approximately 3% in frequency
measurement and 40% in amplitude measurement. These errors occur due to manufacturing tolerances. Regarding the highamplitude measurement error, it must be noted that the sensor
works in a resonant mode, so even small manufacturing tolerances may cause big measurement errors. These errors cannot
be predicted, neither by VHDL-AMS nor by FEM simulation.
With the results of the system simulation it has been possible to
select significant patterns for digital signal processing using the
fuzzy pattern algorithm.
VII. CONCLUSION
Within this work it was shown that the application of reduced-order modeling in system design is a powerful approach

1026

IEEE SENSORS JOURNAL, VOL. 5, NO. 5, OCTOBER 2005

to higher accuracy and simulation speed. By using the reducedorder model instead of the abstract model the simulation error
has been reduced from 20% to below 1% and the simulation
speed has been doubled.
The application of the Multi Architecture Modeling approach within this context assures an easy, fast and safe exchange of abstract and reduced-order models.
The approach of reduced-order modeling including
VHDL-AMS export is supported by a self developed EDA-Tool
in combination with a commercial FEM simulator. In the future, this approach will be able to avoid a lot of man work for
developing abstract, less accurate models.
REFERENCES
[1] P. Feldmann and R. W. Freund, Efficient linear circuit analysis by Padapproximation via the Lanczos process, IEEE Trans. Computer-Aided
Design, vol. 14, no. 7, pp. 639649, Jul. 1995.
[2] K. J. Bathe, Finite Element Procedures in Engineering Analysis. Englewood Cliffs, NJ: Prentice-Hall, 1985.
[3] S. Reitz, J. Bastian, J. Haase, P. Schneider, and P. Schwarz. System level
modeling of microsystems using order reduction methods. presented at
DTIP Conf.
[4] L. D. Gabbay, Computer-aided macromodeling for MEMS, Ph.D. dissertation, Massachusetts Inst. Technol., Cambridge, MA, 1998.
[5] F. Bennini, J. Mehner, and W. Dtzel, Computational methods for reduced order modeling of coupled domain simulations, in Proc. 11th Int.
Conf. Solid State Sensors Actuators, 2001.
[6] J. Strackeljan and D. Behr, Condition monitoring of rotating machinery
using a fuzzy pattern recognition algorithm, in Proc. 6th Int. Conf. Vibrations Rotating Machinery, Sep. 1996.
[7] G. Lipovszky, K. Solyomvari, and G. Varga, Vibration Testing of Machines and Their Maintenance. Amsterdam, The Netherlands: Elsevier, 1990.
[8] J. Mehner, D. Scheibner, and J. Wibbeler. Silicon vibration sensor arrays
with electrically tunable band selectivity. presented at MICRO SYSTEM
Technologies
[9] K. Eichhorn, Entwurf und anwendung von asics fr musterbasierte
fuzzy-klassifikationsverfahren, Ph.D. dissertation, Tech. Univ. Chemnitz, Chemnitz, Germany, 2000.
[10] B. Romanowicz et al., VHDL-1076.1 modeling examples for
microsystem simulation, in Analog and Mixed-Signal Hardware
Description Languages. Boston, MA: Kluwer, 1997.
[11] A. Dewey et al.. Energy-Based characterization of microelectromechanical systems (MEMS) and component modeling using VHDL-AMS.
presented at 2nd Int. Conf. Modeling Simulation Microsystems
[12] A. Endemao et al.. VHDL-AMS modeling, simulation and testing of
electrostatic micromotors. presented at DTIP Conf.
[13] M. Schlegel, G. Herrmann, and D. Mller. A system level model in
VHDL-AMS for a micromechanic vibration sensor array. presented at
IEEE Int. Conf. Sensors
[14] J. Mehner, L. D. Gabbay, and S. D. Senturia, Computer-Aided generation of nonlinear reduced-order dynamic macromodels, J. Microelectromech. Syst., pp. 270277, Jun. 2000.
[15] M. Schlegel, G. Herrmann, and D. Mller. Multi-Architecture-Modeling: Entwurfsmethode fr mixed-signal- und multi-domain-systemsimulation. presented at GI/ITG/GMM-Workshop: Methoden
Beschreibungssprachen Modellierung Verifikation von Schaltungen
Systemen
Michael Schlegel was born in 1974. He received the
B.S. degree in micro-system technology and the M.S.
degree from the University of Applied Science, Mittweida, Germany, in 1992 and 1998, respectively. He
is currently pursuing the Ph.D. degree at the Chemnitz University of Technology, Chemnitz, Germany.
From 1998 and 2003, he was with SIMEC
GmbH (now Ansoft Corporation), working on the
development of the VHDL-AMS simulation tools
hAMSter and Simplorer. In 2000, he joined the
Chemnitz University of Technology as a Research
Engineer in the Faculty of Electrical Engineering and Information Technology.
He works within the German Collaborative Research Center (SFB) 379, which
is funded by the German Science Foundation (DFG). His research interests are
system design with VHDL-AMS and design methodology for MEMS.

Fouad Bennini received the M.S. degree in electrical engineering and information technology from
the Chemnitz University of Technology, Chemnitz,
Germany, in 1999.
He is currently a Research Scientist at the Department of Microsystems and Precision Engineering,
Chemnitz University of Technology. His research
interests include the design and implementation of
automated macromodel generation algorithms for
electromechanical systems and the integration of
these tools into system simulation environment.

Jan E. Mehner received the Diploma and the Dr.Ing. degree in electrical engineering and information
technology from the Chemnitz University of Technology, Chemnitz, Germany, in 1989 and 1994, respectively.
He is a Scientific Assistant at the Department of
Microsystems and Precision Engineering, Chemnitz
University of Technology. From 1998 to 1999, he was
a Visiting Scientist in the field of macromodelling
with the Massachusetts Institute of Technology,
Cambridge. His research interests include analytical
and numerical methods to design microsystems, CAD-tools, and computational
algorithms for problems with coupled fields.

Gran Herrmann received two Ph.D. degrees from

the Chemnitz University of Technology, Chemnitz,
Germany.
He is a Professor of microtechnologies at the
Brandenburg University of Technology, Cottbus,
Germany, and he is a Lecturer, Researcher, and
holds the Chair of Circuit and Systems Design
at the Chemnitz University of Technology. His
research interests include VLSI design, ASICs, and
microsystems.

Dietmar Mller received the Ph.D. degrees from the

University of Dresden, Dresden, Germany, and the
Chemnitz University of Technology, Chemnitz, Germany.
He is a Professor of electrical engineering and
holds the Chair of Circuit and Systems Design at the
Chemnitz University of Technology. His research interests include VLSI design and field-programmable
logic.
Dr. Mller is a member of the Academy of Science
of Saxony, Leipzig/Germany, and the Association for
Electrical, Electronic, and Information Technologies (VDE).

Wolfram Dtzel received the Ph.D. degree from the

Chemnitz University of Technology, Chemnitz, Germany.
He is a Professor of electrical engineering and
holds the Chair of MEMS Design at the Chemnitz
University of Technology. His research interests
includes the design and simulation of MEM and
precision engineering.
Dr. Dtzel is a member of the Academy of Science
of Saxony, Leipzig/Germany, and the Association for
Electrical, Electronic and Information Technologies
(VDE).