Nationaal Lucht- en Ruimtevaartlaboratorium

National Aerospace Laborator y NLR

NLR TP 97281

Analysis of computational aeroelastic

simulations by fitting time signals

M.H.L. Hounjet, B.J.G. Eussen and M.W. Soijer

 DOCUMENT CONTROL SHEET

ORIGINATOR'S REF. SECURITY CLASS.

TP 97281 U Unclassified

ORIGINATOR
National Aerospace Laboratory NLR, Amsterdam, The Netherlands

TITLE
Analysis of computational aeroelastic simulations by fitting time signals

PRESENTED AT
CEAS International Forum on Aeroelasticity and Structural Dynamics, Rome,
Italy, June 17-20, 1997.

AUTHORS DATE pp ref

M.H.L. Hounjet, B.J.G. Eussen 970626 15 20
and M.W. Soijer

DESCRIPTORS
Aeroelasticity Parameter identification
Aircraft configurations Prediction analysis techniques
Curve fitting Time series analysis
Flight characteristic Transfer functions
MIMO (control systems) Transonic flutter
Modal response

ABSTRACT
The procedures to assess the critical aeroelastic state of an aircraft
with the AESIM simulation method are described. The main innovation is
the identification with MIMO-class methods which opens up possibilities
for efficient usage of aeroelastic simulation methods. Experience with
recent applications is reported.

217-02
ANALYSIS OF COMPUTATIONAL AEROELASTIC SIMULATIONS BY
FITTING TIME SIGNALS
Michael H.L. Hounjet and Bart J.G. Eussen
NLR, Anthony Fokkerweg 2, 1059 CM Amsterdam
email: hounjet@nlr.nl
email: eussen@nlr.nl
Marco W. Soijer
Faculty of Aerospace Engineering
Delft University of Technology, Kluyverweg 1, 2629 HS Delft
email: m.w.soijer@lr.tudelft.nl

Abstract. validation of advanced time-accurate aerodynam-

The procedures to assess the critical aeroelastic state of an ic CFD methods for the aforementioned purposes,
aircraft with the AESIM simulation method are described. resulting in a variety of methodologies for mostly
The main innovation is the identication with MIMO-class
methods which opens up possibilities for ecient usage clean wings and for rigid motions using the modal
of aeroelastic simulation methods. Experience with recent way of coupling. The recent algorithms are already
applications is reported.
Key words: Aeroelasticity, MIMO, Prony, AESIM,
DOULAT, NASTRAN, GUL, Structural Dynamics
1. Introduction
The subject is being addressed because of the
industrial need for an ecient aeroelastic simula-
tion system able to improve aircraft design with
adequate and ecient assessment of aeroelastic
uid-structure interface
behaviour (
utter and/or dynamic responses) for
at plate geometries 1 , that the coupling
transonic
ight or other nonlinear conditions.
Today's industrial aeroelastic studies are per-
formed mainly with a set of classical methods
very ecient and a further signicant reduction of
turn-around time from algorithm improvements is
not expected. Also, the industry seems to be reluc-
tant in taking on these new technologies due to
the fact that in many occasions attention was giv-
en only to the modelling of the physical unsteady
aerodynamic phenonema and much less to the com-
plete aeroelastic problems more relevant to indus-
try. This is supported by the fact that almost no
methods have been reported which deal with the
interpolation issues at the
for non-
methodology is occasionally investigated [3] and
the fact that hardly nothing (methods and strate-
gy) has been reported that deals with the system
restricted to linear assumptions due to their e- identication in computational aeroelastic simula-
ciency in assessing the critical state cases for a large tion.
state space. One of the fundamental tasks in an aeroelastic
For transonic
ow these results are questionable analysis is the determination of the frequency and
and the aforementioned methods are often matched damping of aeroelastic modes (e.g. to detect if one
with data obtained from experiment and or other of the generalized displacements becomes unsta-
CFD methods. ble and
utter will occur) which is a subset of the
The NLR AESIM method is developed with the
objective to assist in the design of future aircraft
which are subjected to increases in
exibility, aero- analysis of the time signal is a closing entry which
dynamic loading and nonlinearity and might be of
value in the early design and development phase for
assessing
ight stability and control, safety and risk
valuation and ride qualities. The method focuses
primarily on aeroelasticity at transonic and mildy
separated
ows where aerodynamic nonlinearity is
a non-negligible factor which cannot be estimated
utter analysis that give the user a
with extrapolating current methodology which is
suited for
ight at subsonic and low angle-of-attack
supersonic speeds.
Over the past 20 years research centres and uni-
versities have put eort in the development and
afore-mentioned system identication.
In most time domain aeroelastic applications the
does not get much attention and is often left out in
presentations. For the well-known Isogai case any
analysis method is good enough in contrast with
more realistic multi-DOF structures. To single out
a few modes for analysis from a multi-modes sys-
tem is not good enough compared to classical p- or
pk-methods of
complete analysis of every mode. As many dierent
time response signals may have to be analyzed, sev-
1
The volume spline method has been introduced in 1994
[1, 2] for that purpose.
eral methods for curve-tting should be available. Geometry NASTRAN
The analysis can be sensitive to light damping sys-
tems, the number of analyzed modes, coinciding Panelling FOLDIT NASAES
Elastomechanical
modes and the signal being non-physical. To over- Geo-preprocessor Preprocessor

come these problems a number of solution strate- Lin. Methods

GUL

gies is needed. CAR88 Surface grid Elastomechanical

data

Stability analysis with time-accurate CFD meth- Pade Transfer function

ods is usually performed with one of the following
BLOWUP
Grid Generator
two strategies:
Convolution

 P- or pk- method analysis. The aerodynamic Interpolation

data required for these methods might be sup-
plied by: i.Harmonic excitation which is ine-
cient when the state space 2 is large; ii. Impuls-
AERO SOLVER Motion
Full Potential stat.flex discrete
response which is prone to noise and is also ine- TLNS sine coupled
impulse gust
cient when the state space is large; iii. Diverging Monitoring
rate method [4]. This method is based on the lin- Grid Modeshapes
Cp Amplitude
earized form of the equations and is suitable in Data fit Frequency
a design cycle because the turn-around time can Damping
AESIM
be brought back to the minimum.
 Fully-coupled simulation. This method is Postprocessing
Flutter diagram Cl - α
especially useful in case of strong non-linearities Cm - α Force history

at low frequencies and a large state space. For

a single ight condition the turn-around time
is always less than the turn-around time when
NASTRAN - MIMO

method 1.i or 1.ii are used for the study of a
general stability problem. For a restrictive study,
methods 1.i and 1.ii might be more ecient than
method 2. It should be mentioned that method
1.iii can be embedded in a reasonably easy way
in a time-accurate CFD method.
The main innovation in this presentation
is the adoption of MIMO-class [5] tech-
Figure 1. AEroelastic SIMulation system
2. Aeroelastic Simulation System
At NLR much eort has been spent to cre-
ate a complete AEroelastic SIMulation system,
to be used primarily for the
utter certication
nology, with the objective to predict the of transport-type aircraft in the transonic speed
system state at multiple
ight conditions regime. Time-accurate simulation of
uid and air-
from a MIMO identication of a fully- frame structure interaction is emphasized. The
coupled simulation at a single
ight condi- AEroelastic SIMulation system is referred to as
tion, by extracting useful data (e.g. Gener- AESIM, after the name of the core program.
alized Forces) from the coupled simulation The AEroelastic SIMulation system is built around
which can be used for other purposes. the AESIM core and consists of six independent
main program modules, see gure 1:
 FOLDIT: surface grid generation.
By the aforementioned continuation it is expected  BLOWUP: grid generation.
to increase eciency, so that industrial aeroelas-  NASAES: elastomechanical data manipula-
tic studies might be performed with fully-coupled tion.
CFD methods in assessing the critical state cases  AESIM core.
for a large state space.  Output interfacing e.g. to NASTRAN or MIMO.
This presentation describes the experience with the  Linear methods library.
currently available tting methods and strategies The AESIM core program is divided into 5 individ-
in applications. ual modules and contains those subroutines which
are CPU intensive and make it possible to run the
core in stand alone mode:
2
The state space here is the union of the geometrical
state (Vibration modes etc.), the structural state (Mass,  Interpolation: Interpolation of elastomechan-
stiness, eigenfrequencies etc.) and the
uid state( Airspeed, ical and aerodynamic data (Volume Spline
altitude, etc.) Method [1, 2]).
 Aero solver: Time-accurate solving of aerody- too. They will enhance the analysis capability as
namic equations (Full Potential and TLNS3 ). depicted in gure 2.
 Motion: Either description of motions, seeded

ows, or solving of elastomechanical equations
with or without external loadings. Minimum
SIMO
Impuls responsie methode
 Monitoring: Visualization of simulated data.
Signaallengte
Frequentie-
inhouds

 Postprocessing: Recollection and assimilation

verhouding
F MIMO
SISO
of facts and gures of past simulation(s).
Recently activities have been started to extend
r Control Systems
e Impuls
q
u responsie Fully Controlled Flight

the AESIM method in the direction of ows and

e methode
n
geometries which are encountered in military air-
Frequentiedomein
t
i
craft. Then the code will be ready to prove its val- e Tijddomein
s
ue in applications with mild ow separation which SISO

is primarely responsible for inducing strong limit Exp.SineFit

cycle oscillation structural responses, which might Prony

restrict the ight envelope of the aircraft.

In this paper we give attention to the time-analysis Signaallengte (Timesteps)
and strategies to use the method eciently. The Figure 2. Deployment time-analysis methods with respect to
other parts have been presented in [8]. aeroelastic sytems

Since a wide array of time response signals is avail-

3. Time signal analysis able several ways exist to make use of the afore-
The fundamental task in an aeroelastic analysis is
the determination of the frequency and damping of
aeroelastic modes (e.g. to detect if one of the gen-
eralized displacements becomes unstable and
ut-   For every modal mode separately:
ter will occur). As many dierent time response
signals may have to be analyzed a comprehensive
mentioned time-tting tools. The most common
time response signals which can be used to deter-
mine the frequency and damping characteristics of
the discrete time signal consist of:
 the generalized coordinate,
 the velocity of the generalized coordinate,
set of methods for curve-tting should be avail-
able. In general each time response signal exists
of contributions of various modal modes, of which
the frequency and damping of each one have to be
determined.
Therefore, during an unsteady simulation the data
 the generalized force.
  Also a combination of modal modes and/or the
pressure or deformation data at selected points
can be analyzed.
Experience has learned that for a fail safe analy-
sis of an elastomechanical system the above men-
must be analyzed on-line in the time domain in tioned tting routines are applied rst to the non-
order to determine the behavior of a coupled sys- aerodynamically loaded system and next to the
tem. The main purpose of this analysis is to deter- system loaded using linear aerodynamics, [10],
mine the frequency and damping characteristics of through convolution of transfer functions, [4]. The
the discrete time signal. To full that task the fol- data from these analyses might act as a guide-
lowing methods have been embedded (see section line for the analysis of the non-linear time signal,
3.2): originating from the coupled non-linear
uid struc-
 The exponential sine t,[9] ture simulation. The analysis process has been ful-
 Prony's method, ly automated through use of scripts. This facility
 Fast Fourier Transform analysis, allows the analysis process to be repeatible and to
 Curve-tting of transfer functions.
Very recently, in a cooperation with TUDelft a fea-
sibility study has been started to apply the promis-
ing MIMO-class techniques [5] for that purpose
3
Recently the TLNS methodology as reported in [6, 7]
has been embedded as there is enough evidence that with
the Thin-Layer Navier-Stokes equations complemented by a
simple turbulence modelling the needs of the industry can
be documented.
It should be noted that also the analysis might pro-
vide a prognostic way to speed up the simulation by
allowing for larger time steps [3].
3.1. MIMO-class system identification
The main innovation in this presentation is the
be met for many congurations. adoption of MIMO [5] technology. This permits a
black box 4 evaluation of the aeroelastic system in al closely located poles have to be distinguished.
such a way that after a single fully-coupled sim- However, the closed-loop character of the identi-
ulation for one
ight condition the system state cation problem poses a major problem. Since there
for other
ight conditions (e.g. qdyn ) might be pre- is no external input or reference signal, the aero-
dicted and to extract useful data (e.g. Generalized
Forces) from the coupled simulation which can be
used for other purposes.
The main purpose is to extend the single point
application of coupled simulation methods to mul-
dynamical model that is identied will converge to
the inverse of the mechanical model. This eect is
due to the absence of an external signal that drives
the systems. In addition, the aerodynamic model is
insuciently excited to use stochastic theories on
tiple points and wayhead is given to perform
postprocessing activities, pk-, k-method etc, with
extracted data from an application of a fully-
coupled simulation.
which frequency-domain identiers are based. In
fact, the input signal is completely deterministic
and contains only the system dynamics.
Finally, frequency-domain analysis is mainly useful
for signals that are distorted by large amounts of
qdyn noise and where specic frequency ranges are of
interest. Both are not the case for the problem at
Q11 + hand.
Therefore time-domain analysis has been select-
Ax + Bu Q
12
+ ed. The use of time-domain methods to estimate
the aerodynamic model, allows for the isolation of
C x + Du Q
21 + causal and non-causal relations between the inputs
+ and the outputs. Non-causal eects are not to
Q
22 be modeled, since these are caused by the elas-
qdyn tomechanical system in the closed loop. Because
the structure of the mechanical system is known
Figure 3. Generic 2 mode aeroelastic system to be strictly non-causal, i.e. there is no direct
feedthrough from the forces to the displacements,
A 2-DOF example is presented in gure 3 where we both a lter model and a prediction model can be
want to assess the transfer functions (generalized used for the aerodynamic system.
forces) from the coupled simulation. From the cou- For identication, a multivariable regression model
pled responses x; u at a xed qdyn the MIMO anal- is chosen for the more general lter case. Details of
yses will deduct the A; B; C; D; Q11; Q12; Q21 and the aforementioned procedure will be published in
Q22. The condition of the system is equivalent to a forthcoming publication.
the amount of proportional feedback. The symbols
used in the gure are explained in the appendix. 3.2. Data analysis methods embedded in
In order to obtain an estimate for the
utter condi- AESIM
tion from a single time simulation, the aerodynam-
ics that relate the generalized aerodynamic forces The methods described in this section obtain the
to the generalized displacements are to be mod- frequencies and dampings of a dataset with con-
elled and identied. Since the system operates in a stant time step:
closed loop with the elastomechanical system, mul-
tivariable algorithms for closed-loop identication xl = x(l
t) j l = Lb : : :Le ; (1)
are required. The data that is to be used is obtained
by simulation, so the signals are not contaminated where Lb denotes the beginning and Le denotes the
by measurement noise. However, a small amount end of the data set. Furthermore the time is dened
of process noise is present, since round-o errors as:
are introduced on the propagation of the signals
through the two systems. tl = t = l
t j l = Lb : : :Le : (2)
The complicated dynamics that govern the sys-
tem for a large number of modes, suggest the use
of frequency-domain methods. This class of algo-
rithms is well applicable to systems where sever-
4
No knowledge is assumed of coecients of the structural
and aerodynamic system
3.2.1. Exponential sine tting and determination
of dampings and frequencies
The data set is approximated in time in a least
squares sense by the real valued function Xf (t)
dened by [9]:  is dened by:
 k , k 
X
Nf
k = arctan ,|  +  ; (10)
Xf (t) = ak ek t: (3) k k
k=1
where # denotes the conjugate part.
N f is the order of the t which is less or equal to From the denition of k it is clear that k is not
twice the number of modal modes N h . The coe-
cient ak is complex. For every such term there is
a companion term which is the complex conjugate,
so that the components of the sum are real. Such
a sum arises as the homogeneous solution of a lin-
ear system of ordinary dierential equations with
constant coecients. In particular dynamical sys-
uniquely dened. Here l = 0 is taken.
A.2: k are the roots of the polynomial equation:
N f + pN f N f ,1 + : : : + p1 = 0:
(11)
The zeros of the polynomial equation 11 which
occur in complex conjugate pairs are determined
tems with symmetric mass, stiness and damping using a general root nding routine, see [12]. A.3:
matrices. The coecients pk j k = 1 : : :N f are dened by
The complex k is dened as follows: the matrix equation, see [13]:
k = k + |k ; (4) 0 c1 : : : cN f 1 0 p1 1 0 ,cN f +1 1
BB c2 : : : cN f +1 C B p C B ,c f C
from which the frequencies and dampings of Xf (t) B@ .. . . . .. C CA BB@ ..2 CCA = BB@ N.. +2 CCA(12)
:
are dened: . . . .
cN f : : : c2N f ,1 pN f ,c2N f
Damped natural frequency = k ; (5)
Characteristic damping coecient = k : (6) The coecients in the matrix of equation 12 are
dened as follows:
In the least squares t the squared error dierence cl = (13)
between the output t Xf (t) and the input time- 8
history xi given by: >
> xl j (l = Le , 2N + 1 : : :Le
f
>
< 1 Pr ^Le , 2N f + 1 = Lb);
X
Le = r i=1 xl,i+1 j (l = Le , 2N f + 1 : : :Le ;
E= (Xf (tl ) , xl ))2 ; (7) >
> ^Le , 2N f + 1 > Lb);
l=Lb : r = 2 _ : : : _ Le , Lb , 2N f + 1
is minimized. where Ie denotes the end of the selected time inter-
The dampings and frequencies are obtained with val.
the following methods: In principle a set of 2N f equally spaced values of
A. Prony : the function x(t) is sucient to determine the func-
The Prony formulation derived by Prony in tion Xf (t).
1790 [11] to analyze elastic properties of gasses is However, a more robust procedure is obtained
described below. according to [13] when the set of data points con-
The k are obtained with the following procedure: tain more than 2N f points and the coecients are
Determination of k : obtained by averaging the evenly spaced values of
x(t).
A.1: Introduce the parameter: The ak 's parameters are obtained as explained
k = ek
t: (8) below.
B. Levenberg-Marquardt Method :
Which can be written as: Previous obtained k 's can be improved by the
k = rke|(k +2l): (9) Levenberg-Marquardt Method, see [14]. This iter-
From which the characteristic damping coe-
cient and the damped natural frequencies can be
obtained as:
k =
1t log(j rk j);
ative non-linear t procedure is applied here using
frozen ak parameters.
A reasonable guess to the aforementioned k and
ak parameters is necessary for reliable results. The
initial guesses for k are provided by:
1. the aforementioned Prony method.
k = k +
t2l j l = 0: 2. the natural harmonic frequencies (eigenfre-
quencies) of the structure.
 3. obtained by the curve tting method  User-supplied specication of the selected inter-
explained later on. val (subset) tb = Lb
t; te = Le
t.
The ak 's parameters are obtained as explained  When Prony's method is applied evaluation of
below. the 0s and the a0k s is performed for a range of
C. Complex curve t : averages with r = 1; 2; 4; 8 : : : and selection of
When the assumption can be made that after l = the coecients which result in the smallest error.
Lb no external forces are applied (the input-signal  Evaluation of the 0s and the a0k s is performed
to the system is zero). The 0k s can also be derived by using a sequence of modes (1; 2; 4


N f ).
by the procedure below.
After applying the discrete Fourier transform to 4. Applications
the data:
xf = F x; (14) The examples here focus on current ongoing activ-
the data is approximated by a complex curve t[15] ities with respect to the time-analysis and demon-
and the coecients of the rational polynomials are strate the status of the aeroelastic environment.
obtained:
PN a A (|!)i 4.1. T-tail
xf (|!) ' PNi=0b i i ; (15)
i=0 Bi (|! )
where:

Synct=.6261E-04

Re =.1957E+08
B0  1:

Altit=.0000E+00
(16)

Full Pot. model

Time 14:55:38
SEQNR= 68
Mach = 0.840
Alpha= 0.000

EAS* =285.8
CAS* =285.8
Dyn.p=.4004

Dt :95-11-23
Mass =1.000

NTS = 507
Freq =1.000
LS =1.000
V* =316.4

0.14
0.00

0.00
33

0.1402

0.0000
N a; N b are the order of the polynomials. N b is

NC=
SPC
C^=
C_=
dis

dis
equal to twice the number of modal modes N b =

1850

1850
2N h and in general N a  N b.
The 0k s are the roots of the denominator polyno-
z-proj: 1

y-proj 1
mial which again are obtained by a general root
2

2
nding method.

x
1550

1550
With the k 's determined as indicated above, it
only remains to determine the ak 's. The following 1250

349 1250
system of equations remains to be solved for deter-
y

z
357

125

323

90
-106

-141
mining the ak 's:
2 1 1 : : : 1 3 0 1 t 1 0 1
x-proj 1

a1e xLb
DispLacement projection mode

66 1 : : : N f 77 B CC B
2

66 .. . . .. 77 B
B a2e2t CC = BB x2 C
C
.. C
y
:
5B@ ...
125
4 L .,L . L .,L A @ . A
1 e b : : : Nef b aN f eN f t xLe
ttail
*AESIM*V0.901

(17)
-98

Equation 17 is solved by a General Linear Least

z
264

90

-82

Squares method in which the design matrix is Figure 4. Second mode shape of T-tail
inverted by one of the following procedures:
 Gauss elimination, [16] In order to demonstrate the ability of the sys-
 LU Decomposition, [17] tem to deal with existing aircraft structures a
 SVD, singular value decomposition [17]. If the transport-type T-tail fuselage conguration was
system of equations is numerically very close to considered. The elastomechanical model consisted
singular, Gaussian elimination and LU decom- of ten modes. The geometry and the second mode
position may fail to give satisfactory results. In are depicted in gure 4.
that case SVD is a more powerfull tool. The generalized coordinates of each individual
It can be concluded from the aforementioned proce- mode were calculated in time as a result of a non-
dure that a strategy has to be developed to obtain zero initial value for the acceleration of the general-
the best t. The strategy developed sofar consist ized coordinate. In gure 5 the time response infor-
of: mation is evaluated through exp. sine t signal pro-
 4.2. Embedding of lineair aerodynamics

Da-Fre(Ts): EXP.S.F
EXP.S.F
Synct=.6261E-04

Re =.1957E+08
Altit=.0000E+00
To ease applications and to build condence a cou-

-.444E-01 29.69
Full Pot. model

DATA
Time 14:52:35
SEQNR= 11
Mach = 0.840

-1.49 20.62
-1.39 27.06

-.172 30.82
-.367 43.94
-.247 45.95
-.306 50.86
-.373 53.89
-.528 67.71
-1.19 78.49
Alpha= 0.000

EAS* =285.8
CAS* =285.8
NTS = 3000
Dyn.p=.4004

Dt :95-12-14
Mass =1.000
Freq =1.000
LS =1.000
V* =316.4
pled simulation should also be run based on linear
aerodynamics. This requires the generalized forces
(transfer functions) which are in general available

73614
in the frequency domain to be tted [4, 18] and

Ti me
36807
transformed to the time-domain [19]. A feasibility
4

8
study with 2-D airloads and 3-D airloads has been

0
q

q
*10 -1

*10 -1
performed to investigate the most ecient way to
0.319

0.008

0.191
-0.303

-0.005

-0.201

73614
embed linear aerodynamics in the AESIM method.

Ti me
36807
The assumption is made that the behavior of
- AESIM - dynamic MONITORING MODE: AMPLITUDES

any unsteady parameter of interest such as an

0
q

q
*10 -1

*10 -1
0.326

0.018

0.201

0.002

aerodynamic load or a pressure coecient can be

-0.291

-0.196

73614
described by a appropriate form for the transfer

Ti me
36807
function which is a ratio of two s dependent poly-
10
2

nomials which is known as the Pade approxima-

0
q

q
*10 -1

*10 -1

*10 -1
0.399

0.051

0.224

0.124

0.008
-0.298

-0.002

-0.229

-0.109
tion:
73614
Ti me
36807
ttail

2 3
G(s) = A10++BA1ss++BA2ss2 ++BA3ss3 ++






*AESIM*V0.901

(18)
0
q

q
*10 -1

*10 -1

*10 -1
0.493

0.240

0.010

0.144

0.005
-0.056

-0.605

-0.220

-0.135

1 2 3
Figure 5. Result after exp. sine t signal processing of The complex curve tting procedure is used also
dynamic response of generalized coordinates at Mach=0.84 here to obtain the approximation.
and zero altitude in Standard Atmosphere
The rational polynomial is transformed to state
space form:5
0 x_ 1 0 0 1 0 1 0 x 1 0 0 1
0.1 0.05 0.05 0.05

1 1
0 0 0 0
@ x_2 A = B@ 0 0 1 CA @ x2 A + @ 0 A u
x_3 , B13 , BB31 , BB23 x 1
 A0 A3 1 A1 A3 B13 A2 A3 B2 
−0.1 −0.05 −0.05 −0.05
0 1000 2000 0 1000 2000 0 1000 2000 0 1000 2000

y = B3 , B3 B3 B3 , B3 B3 B3 , B3 B3
0.05 0.05 0.05 0.02

0x 1
1
0 0 0 0

@ x2 A + A3 u (19)
−0.05
0
0.02
1000 2000
−0.05

0.02
0 1000 2000
−0.05
0 1000 2000
Damping
−0.02
0 1000 2000
Frequency x3 B3
−0.7519 21.35
−10.23 24.94

This system is solved using the Newmark scheme

−0.1582 29.22
−0.01041 30.64
0 0 −0.434 43.64

embedded in AESIM, [1].

−0.2313 45.57
−0.3212 50.44
−0.3468 53.48
−0.5184 67.17
−0.02 −0.02 −1.16 77.79
0 1000 2000 0 1000 2000

Figure 6. Result after MIMO signal processing of dynamic

4.2.1. Two-dimensional application
response of generalized coordinates at Mach=0.84 and zero Calculations of unsteady airloads have been per-
altitude in Standard Atmosphere formed with DOULAT for a at plate pitching
about an axis 0:5c in front of the leading edge
(mode 2) and heaving (mode 1) at M1 = 0:5 and a
reduced frequency range up to j s j= 1:0. 6 The gen-
cessing to get damping and frequency information. eralized forces data generated by DOULAT were
Results of the MIMO-class procedure are depict- tted with the aforementioned procedure. There-
ed in gure 6. The exp. sine t results compare after the Newmark scheme was applied to oscil-
reasonably well with the simulated data. However, latory motions in the same frequency range and
the MIMO-class t results are astonishingly good. the time traces were transformed to the frequency
Both methods revealed about the same damping domain.
and frequencies. The results of these simulations 5
Equation 19 is given here for a third order system.
show that the T-tail has a stable dynamic behav- 6
The reduced frequency is dened here as k = Im(s) =
ior for the ight condition under consideration. !c
2U where c denotes the chord.
 H11 H12
at Mach=0.901. This conguration is described in
[20]. Again two modes have been selected.
1.5 2.5

A similar procedure as outlined above was applied.

2
1
1.5
Imag

Imag
1
0.5 H11 H12
3 0.2
0.5
2.5 0
0 0
−0.6 −0.4 −0.2 0 0.2 −2 −1 0 1 2
Real Real −0.2

Imag

Imag
H21 H22 1.5
4 8 −0.4
1
3 6 0.5 −0.6

2 4 0 −0.8
Imag

Imag
−3 −2.5 −2 −1.5 −1 −4 −3.8 −3.6 −3.4 −3.2
Real Real
1 2
H21 H22
0.6 3
0 0
2.5
−1 −2
0.5 1 1.5 2 2.5 −2 0 2 4 0.4 2
Real Real

Imag

Imag
1.5

Figure 7. Comparison of directly calculated and tted 0.2 1

unsteady coecients of a harmonically heaving and pitch- 0.5
ing at plate at M1 =0.5. 0 0
0 0.5 1 1.5 −1.5 −1 −0.5 0
Real Real

Figure 7 shows a comparison in the frequency Figure 9. Comparison of directly calculated and tted
domain between the original data (circle) and the unsteady coecients of the harmonically oscillating wing
445.6 at M1 =0.901.
tted data (line-cross) which show a good agree-
ment.
H11 H12
H11 H12
3 0.2
1.5 2.5

2 2.5 0
1
1.5 2
−0.2
Imag

Imag
Imag

Imag

0.5 1 1.5
−0.4
0.5 1
0 −0.6
0 0.5

−0.5 −0.5 0 −0.8

−0.6 −0.4 −0.2 0 0.2 −2 −1 0 1 −3 −2.5 −2 −1.5 −1 −4 −3.8 −3.6 −3.4
Real Real Real Real
H21 H22 H21 H22
5 8 0.6 3

4 2.5
6
3 0.4 2
4
Imag

Imag
Imag

Imag

2 1.5
2
1 0.2 1
0 0 0.5
−1 −2 0 0
0.5 1 1.5 2 2.5 −2 0 2 4 0 0.5 1 1.5 −1.5 −1 −0.5 0
Real Real Real Real

Figure 8. Comparison in the frequency domain of direct- Figure 10. Comparison in the frequency domain of directly
ly calculated and to and fro transformed unsteady coe- calculated and to and fro transformed unsteady coecients
cients of a harmonically heaving and pitching
at plate at of the harmonically oscillating wing 445.6 at M1 =0.901.
M1 =0.5.
Figure 9 shows a comparison in the frequency
Figure 8 shows a comparison in the frequency domain between the original data (circle) and the
domain between the original data (circle) and the
data (cross) obtained by analysing the time traces
which again show a good agreement. From this
the conclusion might be drawn that the aforemen-
tioned procedure is applicable in 2D.
4.2.2. Three-dimensional application
Calculations have also been performed with GUL
for the 3-D AGARD standard aeroelastic wing
tted data (line-cross) which show a good agree-
ment.
Figure 10 shows a comparison in the frequency
domain between the original data (circle) and the
data (cross) obtained by analysing the time traces
which again show a good agreement. From this
the conclusion might be drawn that the aforemen-
tioned procedure shows good promise for embed-
ding in the AESIM system.
4.3. MIMO-class application MSE(o) AIC(+) FPE(x)

1st bending
0

The applicability of the MIMO method [5] in ut-

10

ter analysis is presented for an aeroelastic inves-

tigation which was conducted for one of the 3- 0 2 4 6 8 10 12 14 16 18 20

D AGARD standard aeroelastic congurations in

0

1st torsion
10

subsonic, transonic and supersonic ow. This con-

guration is described in [20]. The conguration 0 2 4 6 8 10 12 14 16 18 20

for dynamic response I wing 445.6 model "weak-

2nd bending
0
10

ened no. 3" was selected at Mach=0.901. The data

were obtained from [3]. 0
0 2 4 6 8 10 12 14 16 18 20

2nd torsion
10
5
10

0 2 4 6 8 10 12 14 16 18 20
Model order
4
10
Figure 12. Validation criteria for AGARD aeroelastic system
at subcritical ight condition
Condition number

3
10

Figure 12 shows the validation criteria for each

2
10 of the multiple-input single-output models in the
four-mode system that was mentioned before. The
1 order of the multiple-output model is equal to the
sum of the orders of the single-output models.
10

0 −3
10 x 10
0 1 2 3 4 5 6 7 8 9 2
1st bending

Order
MSE 0.4967%
0
Figure 11. Order of excitation versus condition number for
AGARD aeroelastic system at subcritical ight condition −2
0x 10−3 50 100 150 200 250 300 350 400 450 500
5

Persistency of excitation for the four-mode sys-

1st torsion

0 MSE 0.326%

tem is illustrated in gure 11, where the order of −5

excitation is shown along the horizontal axis and 1

0x 10−3 50 100 150 200 250 300 350 400 450 500
2nd bending

the condition number for the correlation-function 0

MSE 1.079%

Toeplitz matrix is printed along the vertical axis.

From these condition numbers, the input is found
−1
0x 10−3 50 100 150 200 250 300 350 400 450 500
1

to be persistently exciting of order two at most.

2nd torsion

MSE 1.711%

The model order is determined by separating the

0

data in three subsets. The rst subset contains the

−1
0 50 100 150 200 250 300 350 400 450 500
Time

rst half of the available time series and is used for

identication. The second set contains only 12%
of the data points and is used to track the state
ight condition
vector of the new model without using the subset
Figure 13. Comparison of MIMO tted (: : :) and original
(-) generalized forces data for AGARD aeroelastic system
at subcritical
that was used in the identication step. Finally, Figure 13 shows the generalized forces as obtained
the last 38% of the original time series is used to from simulation with the identied model, togeth-
validate the new model using the estimated state er with the original data. The data is plotted for
vector as an initial condition. time points after the transition has damped out.
The validation step results in three criteria to sig- An excellent agreement is shown between both
nal overtting of the system. The normalized mean datasets (they coincide entirely).
square error indicates the best model. However, A slightly dierent model structure (MIMO ),
this may not be the optimal model since this crite- including auto regressive terms on the outputs,
rion contains no penalty for hugh models. Akaike's leads to slightly larger errors for the same simu-
information criterion (AIC) and the nal predic- lation, as shown in gure 14.
tion error criterion (FPE) both do include the num- Finally we come to the main purpose of the excer-
ber of model parameters in the resulting cost. How- cise. We increase the airspeed to a supercritical
ever, all three criteria usually indicate the same value and apply the MIMO results obtained from
optimal model order. tting the subcritical airspeed data and make the
 −3 −3
x 10 x 10
2 5

1st bending

1st bending
MSE 2.725% MSE 3.39%
0 0

−2 −5
0x 10−3 50 100 150 200 250 300 350 400 450 500 0x 10−3 50 100 150 200 250 300 350 400 450 500
5 5
1st torsion

1st torsion
0 MSE 1.52% 0 MSE 5.574%

−5 −5
0x 10−3 50 100 150 200 250 300 350 400 450 500 0x 10−3 50 100 150 200 250 300 350 400 450 500
1 1
2nd bending

2nd bending
MSE 3.774% MSE 3.924%
0 0

−1 −1
0x 10−3 50 100 150 200 250 300 350 400 450 500 0x 10−3 50 100 150 200 250 300 350 400 450 500
1 2
2nd torsion

2nd torsion
MSE 4.04%
MSE 66.75%
0 0

−1 −2
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Time Time

Figure 14. Comparison of MIMO tted (: : :) with auto Figure 16. Comparison of MIMO predicted (: : :) and ref-
regressive terms on outputs and original (-) generalized erence (-) generalized forces data for AGARD aeroelastic
forces data for AGARD aeroelastic system at subcritical system at supercritical ight condition
ight condition
5. Conclusions
In this paper the status of the NLR system
comparison with results of the aeroelastic simula- for aeroelastic simulation has been presented and
tion at the higher airspeed. demonstrated with the emphasis on time-analysis.
Experience with recent applications and ongoing
developments led to the following observations:
−3  The analysis of time signals can be carried out
satisfactorily with the available models.
x 10
5
1st bending

0 MSE 7.422%
 The MIMO-class t procedures are superior to
−5 the ones currently embedded in the AESIM
method.
0x 10−3 50 100 150 200 250 300 350 400 450 500
5
1st torsion

0 MSE 16.23%
 The MIMO-class analysis application has shown
−5 good promise for increasing the eciency of
coupled simulations by allowing results made for
0x 10−3 50 100 150 200 250 300 350 400 450 500
1
2nd bending

a single ight condition being extented to mul-

MSE 8.773%
0

tiple ight conditions.

 The procedures for utilization of lineair aerody-
−1
0x 10−3 50 100 150 200 250 300 350 400 450 500
2

namics from the frequency domain to the time

2nd torsion

MSE 47.18%
0

domain have shown good promise for embedding

in the aeroelastic simulation system.
−2
0 50 100 150 200 250 300 350 400 450 500
Time

Figure 15. Comparison of MIMO predicted (: : :) and ref-

erence (-) generalized forces data for AGARD aeroelastic
system at supercritical
ight condition Appendix
A. Aeroelastic MIMO analysis
Figure 15 and 16 depict the comparison which A very important application of the MIMO anal-
shows that the system at the supercritical airspeed ysis is the black box evaluation of the aeroelastic
is unstable and that the linear MIMO model pre- system in a way as to be able to predict the
utter
diction performs very well for the lower 3 modes. condition after just a single time-simulation.
Mode 4 is overpredicted. The system condition is equivalent to the amount
The modied model MIMO results in smaller of proportional feedback as depicted in gure 3.
errors than the model from the original model set.
Which model set results in the best estimates for To be analyzed, by the MIMO software, in gure 3
aerodynamic modeling in aeroelastical closed-loop are the aerodynamic transfer functions, Q11, Q12,
systems is yet to be investigated. Q21 and Q22, in the case of a two mode structural
system. References
The denition of the structural system is as fol-
lows:
The generalised coordinates q i for each vibration
mode may be dierent in time and are based on
the generalized modal de
1. M.H.L. Hounjet and B.J.G. Eussen, Outline and Appli-
cation of the NLR Aeroelastic Simulation Method,
ICAS-94-5.8.2, September 1994
2. M.H.L. Hounjet and J.J. Meijer ,Evaluation of Elas-
tomechanical and Aerodynamic Data Transfer Meth-
ection approach. ods for Non-planar Congurations in Computation-
The dynamic deformations are expressed in gener-
alized coordinates q i and their associated modal
mass M , damping D, stiness K and vibration
modes ~hi for N h modes which satisfy the equation:
[M ]q + [D]q_ + [K ]q = F (20)
The coupling of the structural model and the aero-
dynamic model involve the generalised aerodynam-
al Aeroelastic Analysis Contribution to CEAS Sym-
posium: 'International Forum on Aeroelasticity and
Structural Dynamics 1995'
3. B.B. Prananta and M.H.L. Hounjet, Large time step
aero-structural coupling procedures for aeroelastic sim-
ulation . In Proceedings CEAS Symposium: Interna-
tional Forum on Aerelasticity and Structural Dynam-
ics, CEAS, Rome, June 17-20 1997
4. M.H.L. Hounjet and B.J.G. Eussen , Prospects of time-
linearized unsteady calculation methods for exponen-
tially diverging motions in aeroelasticity AIAA paper
92-2122, April 1992
ic forces acting on the aircraft conguration which 5. M.W. Soijer , Frequency Domain Identication of
are obtained by solution of the aerodynamic equa- Rotorcraft State Space Models and Applications to a
BO105 Rigid Body Model Master's thesis, Delft Univer-
tions: sity of Technology, Faculty of Aerospace Engineering,
Z December 1996
6. B. Prananta, M.H.L. Hounjet and R.J. Zwaan, A Thin
F i = ,qdyn (Cp , Cpstc )~hi
Nds;
~ (21) Layer Navier-Stokes Solver and its Application for
s Aeroelastic Analysis of an Airfoil in Transonic Flow
where Cpstc is obained by a previous static simula- Proceedings CEAS Symposium: 'International Forum
on Aeroelasticity and Structural Dynamics 1995'
tion or by application of a low-pass lter. 7. B.B. Prananta and M.H.L. Hounjet, Aeroelastic Sim-
Equation (20) is transformed to rst order: ulation with Advanced CFD Methods in 2D and
3D Transonic Flow, In Proceedings Conference on
Unsteady Aerodynamics, RAeS, London, July 17-18
dQ_ + kQ = f; (22) 1996
q 0
1 0  8. B.J.G. Eussen, M.H.L. Hounjet and R.J. Zwaan
Experiences in aeroelastic simulation practices ,
where Q = q_ , f = F , d = 0 M and EUROMECH colloquium 349, Gottingen, Germany,
1996
 0 ,1  9. Bennett, Robert M. and Desmarais, Robert N. Curve
Fitting of Aeroelastic Transient Response Data with
k= K D : (23) Exponential Functions, NASA SP-415
10. M. H. L. Hounjet. Calculation of unsteady subsonic and
supersonic
ow about oscillating wings and bodies by
new panel methods, NLR TP89119 U, April 1989.
Q_ + d,1kQ = d,1f: (24) 11. C. Lancosz . Applied Analysis, Prentice Hall, 1956.
12. Stoer, J., and Bulirsch, R. Introduction to Numerical
Therefore the system in gure 3 can be dened as: Analysis, New York, Springer Verlag
13. Charles L. Keller. Determination of Complex Exponen-
tials, Least Squares and Prediction Methods, AFWAL-
A  ,d,1k (25) TR-88-3042, Air Force Wright Aeronautical Laborato-
ries
B  d,1 (26) 14. Marquardt, D. W. J. Soc. Ind. Appl. Math., Vol. 11,
pp. 431-441, 1963
C 1 (27) 15. E.C. LevyComplex-Curve Fitting IRE Transactions on
D 0 (28) Automatic Control, May 1959
16. Carnahan, Brice, Luther, H.A., and Wilkes, James O.
Applied Numerical Methods, Wiley, New York 1969
17. Forsythe, George E., Malcolm, Michael A., and Moler,
Cleve B. Computer Methods for Mathematical Com-
putations, Prentice-Hall, Englewood Clis, New York
Acknowledgement 1977
18. M. H. L. Hounjet and B. J. G. Eussen.Beyond the fre-
Research performed partly under contract with quency limit of time-linearized methods, NLR TP91216
U, June 1991.
the Netherlands Agency for Aerospace Programs 19. Edwards, J.W. Unsteady aerodynamic modeling and
NIVR contract number: 01904N active aeroelastic control NASA CR-148019
The authors like to mention B.B. Prananta of 20. Yates, C. AGARD Standard Aeroelastic Congurations
for Dynamic Response I-Wing 445.6, AGARD Report
TUDelft for making available the generalized force No. 765, 1988.
data used in the MIMO-class applications.