Aeroelastic Simulations of a Sailplane

David Eller and Ulf Ringertz

Aeronautical and Vehicle Engineering
Royal Institute of Technology
SE - 100 44 Stockholm, Sweden
December 2005
Abstract
Using the ASK 21 training and aerobatics sailplane as a test case,
dierent structural and aerodynamic modeling approaches are compared.
For the structure, a simplied model consisting of a beam framework is
assembled and compared with a more complex shell nite element model.
Frequency-domain lifting-surface aerodynamics is employed for utter
analysis and compared with results obtained from three-dimensional time-
domain potential ow aerodynamics. Finally, a 4g pull-up maneuver is
performed in ight testing and numerical aeroelastic simulations of the
same maneuver are evaluated.
Introduction
The purpose of the present study is to compare and evaluate dierent levels
of modeling detail when trying to assess the aeroelastic behavior of a complete
aircraft conguration. Many dierent modeling approaches are possible depend-
ing on the application at hand. If the analysis is part of a preliminary design
process, a computational approach is the only option and standard procedures
involve a nite element model to assess the structural dynamics properties and
a linear potential ow model to obtain the unsteady aerodynamic forces.
Combining the structural model with the model of unsteady aerodynamics,
it is possible to pose the aeroelastic stability problem as a nonlinear eigenvalue
problem for which there are ecient methods available [1]. Essentially the same
computational model can be used to analyze other aeroelastic phenomena such
as control surface eciency and gust response [2].
Dierent levels of detail can be used for both the structural and aerody-
namic model. The simplest structural model is usually obtained in the form
1
E 2 D. Eller and U. Ringertz
of a beam model of the wings, fuselage and tail. Even though the topology of
the beam model is simple, signicant diculties arise when trying to accurately
estimate the properties of the beams and their interaction at connections such
as the wing-fuselage interface. A more detailed structural model, using for ex-
ample shell elements, may make it easier to dene structural properties. The
drawback is that one may have to model much more details in terms of local re-
inforcements in order to obtain accurate global stiness properties and accurate
representations of resonance frequencies and the associated modal eigenvectors.
If the analysis is concerned with the certication process of an already exist-
ing design for which a prototype has been manufactured, one may eliminate the
numerical structural analysis all together. In this case, a detailed ground vibra-
tion test is used to experimentally determine the structural dynamics in terms
of resonance frequencies and modal eigenvectors. This procedure is well de-
scribed in [3]. The measured structural dynamics properties are then combined
with a numerical model of the aerodynamic forces to perform the actual utter
analysis. This approach is mainly used for smaller general aviation aircraft and
gliders. However, even though general design guides such as those provided
by the DLR [4] are used, costly design changes to the already existing aircraft
prototype may be needed to demonstrate the aeroelastic stability required by the
certication authority as dened in the airworthiness code (CS-22 [5]).
In case the needed engineering manpower and resources are available, a nu-
merical model is used for the structural dynamics even in the early design stages.
The numerical model is then rened during the design process and nally cor-
related to ground vibration test data as it becomes available. A structural op-
timization approach [6] may be used to improve the correlation between the
numerical and the experimental data. Such an approach helps to avoid surprises
late in the development process.
The model of the unsteady aerodynamics can also be of very dierent detail
even if it is restricted to unsteady potential ow. The least complex model
would be to use a strip model were the ow is assumed to be well represented
by a locally two-dimensional analysis as described by Rodden and Johnson [7].
A more detailed model involves the use of a panel method to approximate all
lifting bodies by their mean surfaces, but still neglecting the fuselage and the
thickness of the wing and tail surfaces, using the well known Doublet-Lattice
method [7] or the similar methods described in Refs. [8, 9]. An even more
detailed model would involve a model of the fuselage and also the actual wing
prole of the lifting surfaces.
In this study, dierent modeling approaches for the structural and aerody-
namic behavior of the ASK 21 sailplane will be investigated and compared with
respect to ight mechanics and aeroelastic properties. Numerical models are
Aeroelastic Simulations of a Sailplane E 3
compared with data from ground vibration and ight testing in order to assess
the accuracy which can be obtained with the dierent simulation methods. Fi-
nally, some remarks regarding the relative eort required to create and validate
the dierent models are provided.
The ASK 21 sailplane
The reason for using the ASK 21 sailplane [10] in this study is the availability
of an instrumented aircraft of this type and also that detailed design data has
been kindly provided to the authors by the manufacturer. The ASK 21, which
is shown in Figure 1, is a two-seat sailplane for pilot training and aerobatics. It
Figure 1: The ASK 21 glider.
features a wing aspect ratio of 16.1, large wing thickness and robust glass ber
composite construction. In comparison to modern high-performance sailplanes,
the ASK 21 is relatively sti, and hence less susceptible to aeroelastic instability.
It operates at ight speeds between 80 km/h and 280 km/h and, as it is certied
for aerobatics, may be subjected to load factors from 4g to +6.5g. Due to the
high allowed load factor, elastic deformations experienced in service may still be
fairly high and likely exceed those encountered in other general aviation aircraft
considerably.
Structural model
With a relatively large wing thickness of 19.6% at the root, a moderate wing
aspect ratio, and sandwich shell construction throughout, the ASK 21 is rela-
tively light for its size. Wings, fuselage and tail shells are built from glass-ber
E 4 D. Eller and U. Ringertz
reinforced epoxy resin and polymer foam or tubus core. Most sandwich compo-
nents employ outer face sheets which are thicker than the corresponding inner
faces in order to improve the impact strength of the exposed external layers. In
lightly loaded regions such as the control surfaces, layer thickness is as low as
0.09 mm/0.18 mm (inner/outer face) with 4 mm light foam core, while even
the outboard wing shell is not thicker than 0.18 mm/0.6 mm with 9 mm foam
core. In comparison to more recent sailplane designs featuring carbon ber
sandwich construction and much higher aspect ratio, the ASK 21 is expected
to be relatively sti in terms of global deformations, but comparatively exible
with respect to localized loads.
Modal subspace formulation
When considering small elastic deformations, linear equations of motion ac-
cording to
M x +Kx = f
a
(t, x, x, . . . ) (1)
are used. Here, M is the mass matrix relating accelerations in the structural
degrees of freedom x to inertial forces, K is the stiness matrix and f
a
a vector
of aerodynamic forces. Note that the aerodynamic forces f
a
also depend on the
time history of motion and loads because these determine the shape and strength
of the wake. In order to drastically reduce the number of degrees of freedom,
the standard modal subspace technique is adopted. Let z
j
be eigenvectors and

2
j
eigenvalues solving the generalized eigenvalue problem

K
2
j
M

z
j
= 0. (2)
Then, a reduced number of eigenvectors with low eigenfrequencies
j
are used
to dene a subspace Z = [z
1
, z
2
, . . . ]. Low-frequency eigenvectors are utilized
because it is assumed that higher order modal deformations do not contribute
relevant aerodynamic loads. With mass-normalized eigenvectors, the equations
of motion in modal subspace become uncoupled according to
I q +q = f
q
(t, q, q, . . . ), (3)
with the modal displacements q such that Zq = x, modal loads are found by
projection of the aerodynamic forces into the subspace f
q
= Z
T
f
a
. For the
matrices holds
Z
T
MZ = I, and Z
T
KZ = . (4)
Here, I is the identity matrix and is a diagonal matrix of the squares of
the eigenfrequencies
j
. The modal equation of motion (3) only has a small
number of degrees of freedom, but retains the aeroelastically relevant dynamic
Aeroelastic Simulations of a Sailplane E 5
characteristics of the structure if suciently many eigenmodes are included in
the subspace Z.
Finite element models
Two nite element models of dierent type were assembled for comparison
purposes. The rst, much simpler model is based on a framework of beams,
the properties of which were chosen to match ground vibration test data, and
was initially developed by Keller [11] and co-workers [12, 13, 14]. It was later
modied in order to improve the accuracy of the control system modeling. The
nal beam nite element model contains approximately 2900 structural degrees
of freedom.
The second model is based on triangular and quadrilateral shell nite ele-
ments, as shown in Figure 2, where dierent colors are used for dierent shell
properties. In contrast to the beam model, stiness properties are not designed
to match test data, but derived from lamination plans which were kindly pro-
vided by the manufacturer. The main load-carrying structure accounts for only
roughly 50% of the empty mass of the aircraft, and the distribution of the re-
maining mass is to a large extent unknown. Therefore, mass properties were
used to match the model to data obtained from vibration testing. The model
for which results are discussed below has about 23 000 elements and 70 000
structural degrees of freedom. Both nite element models were constructed for
NASTRAN [15].
Figure 2: Shell FE model including control surfaces.
Testing and validation
The main source of validation data for the structural model was the ground
vibration test (GVT), shown in Figure 3, of the full aircraft performed at KTH
[16], along with additional vibration testing of the fuselage performed as part of
E 6 D. Eller and U. Ringertz
the present investigation. These ground vibration tests were also compared to the
Figure 3: Ground vibration test of the ASK 21 glider.
tests by Niedbal [17] performed during the original certication process. Due
to friction in the control system, most control surfaces do not move when the
aircraft is excited to vibrate, at least not with moderate excitation amplitudes.
Therefore, all aps were xed rigidly to the adjacent wing or tail surface in order
to create a well-dened structural conguration.
In Table 1, eigenfrequencies obtained from modal analysis of the beam (f
b
)
and shell (f
s
) nite element models are compared with test data (f
exp
). For
a more compact notation, the abbreviations given in Table 2 are used in the
following presentation.
In general, it should be noted that due to the placement of acceleration
sensors and excitation devices employed in the GVT, some eigenmode shapes
present in the FE models could not be detected experimentally. The rather
distinct mode at 15 Hz involves mainly torsion of the vertical n, which can
hardly be excited with a laterally acting shaker placed at the rear wheel, i.e. close
to the elastic axis of the n. Even for the higher wing bending modes above
20 Hz, accelerometer data sometimes leaves doubts about the bending mode
shape. Furthermore, very few eigenmodes can be unambiguously identied as
pure fundamental modeshapes. Instead, complex coupling patterns of fuselage,
wing and tail deformations are observed. Comparisons of eigenfrequencies alone
Aeroelastic Simulations of a Sailplane E 7
f
exp
[Hz] f
b
[Hz] f
s
[Hz] Description
2.95 2.90 2.96 sym. wing 1b.
4.09 4.00 4.11 asym. fuselage 1t.
6.50 6.64 6.23 asym. wing 1b. + fuselage 1b.
- 7.28 7.07 sym. fuselage 1b. + wing 1ipb.
7.29 7.09 7.34 asym. fuselage 1b. + n b.
8.25 - - asym. fuselage 1b. + stabilizer rotation
8.74 9.25 8.56 sym. wing 2b. + fuselage 1b. ip.
10.5 10.7 10.3 sym. wing 2b. + fuselage 1b. cp.
12.2 11.6 12.0 asym. fuselage 1b. + n t.
- 15.5 14.2 asym. n torsion + fuselage 1b.
16.9 18.8 16.5 asym. wing 2b. + stabilizer b. cp.
17.4 - - asym. wing 2b. + stabilizer b. ip.
20.4 19.8 19.7 sym. wing 2b. + fuselage 2b. + stab. b. cp.
- 27.9 20.9 asym. fuselage b. + stabilizer b
22.3 - 22.7 sym. wing 2b. + fuselage 2b. + stab. b. ip.
22.5 20.6 24.3 asym. wing 1ipb.
- 26.2 - sym. wing t. + stabilizer b. ip.
- 27.0 - sym. wing t. + stabilizer b. cp.
- - 28.1 sym. wing 2ipb. + stabilizer b.
26.9 - 28.7 sym. stabilizer 1b.
- - 28.8 asym. wing t. + wing 3b.
29.3 - 29.5 sym. wing t.
- 26.1 30.0 asym. wing t.
32.7 36.2 32.4 asym. wing 3b.
Table 1: Comparison of measured and computed eigenfrequencies.
are hence not necessarily useful, as modes with the same description in Table 1
may combine dierent magnitudes of fuselage and wing deformations, although
the modeshape appears visually similar.
For the lowest-frequency eigenmode shapes up to about 15 Hz, both nite
element models compare reasonably well with GVT data. There is one low-
frequency modeshape at 8.25 Hz which remarkably is not present in either nu-
merical model. A possible interpretation of the accelerometer data for this mode
is that the stabilizer which is attached to the vertical tail using two detachable
centerline connections, rotates about the line through the attachment points as
in a joint. The motion may even be combined with some degree of local n
bending and rear fuselage torsion. Both nite element models fail to reproduce
E 8 D. Eller and U. Ringertz
b. modeshape dominated by bending
t. modeshape dominated by torsion
2b. second fundamental bending mode
ipb. in-plane bending (xy-plane)
sym. symmetrical
asym. antisymmetrical
ip. wing & stabilizer tips moving in phase
cp. wing & stabilizer tips moving counter-phase
Table 2: Abbreviations used in Table 1
this particular type of motion accurately, although elastic connections between
n and stabilizer bodies are used. The modeled attachment stiness has a strong
impact on the frequency of the mode at 7.3 Hz, but does not appear to allow a
motion such as that observed in the GVT at 8.25 Hz.
For higher frequencies above 20 Hz, dierences become more pronounced.
Evidently, the shell nite element model produces a number of eigenmode
shapes not observed in vibration testing. As noted previously, this may partly
be due to the location of excitation devices and accelerometers. Moreover, com-
puted fuselage deformations for modes with frequencies exceeding 22 Hz and
wing deformations for all eigenmodes involving torsion show signicant local
shell bending deformations. It is considered likely that the type of local fuselage
deformations observed in the shell nite element model would not occur in
reality since inertial loads are introduced by means of local reinforcements, not
all of which are included in the nite element model. Computed shell defor-
mations for the wings, however, are considered realistic as the stiness of the
wing shell is well dened and the mass fairly evenly distributed. In contrast to
the forward fuselage, no large inertial loads, except the mass forces of the wing
itself, act normal to the surface of the wing. Local shell deformations of this
kind would be hard to detect in vibration testing of the full aircraft unless a
very large number of accelerometers were used.
Five eigenmodes of the shell nite element model have eigenfrequencies be-
tween 28 Hz and 30 Hz, while only one symmetric wing torsion mode with
possibly some bending component was experimentally observed in this interval.
Not a single eigenmode of the beam model is located in the same frequency in-
terval, which, instead, has four eigenmodes with frequencies between 26 Hz and
28 Hz, none of which appeared in vibration testing with a similarly coupled
deformation pattern. Considering the substantial dierences between experi-
mentally identied and computed eigenmodes, results obtained with the above
Aeroelastic Simulations of a Sailplane E 9
structural models should be regarded as approximations. At this point, the
beam model appears to be globally somewhat less accurate than the shell nite
element model for moderate to high eigenfrequencies.
Fuselage vibration testing
The mass distribution in the fuselage is aected by appreciable uncertainties.
A number of heavy local reinforcements, composite landing gear frames and
cockpit installations were found dicult to model. Most of these items can
probably not be represented well by means of shell elements, and accurate mass
distributions were unknown. Therefore, the isolated fuselage was subjected to
vibration testing, as shown in Figure 4 in order to collect more detailed eigen-
mode shapes. Following the testing, the shell nite element model was updated
Figure 4: Ground vibration test of the fuselage.
with three concentrated masses which in total account for the dierence between
measured weight and the structural mass of the nite element model.
Using a trial-and-error approach, the concentrated masses were placed in the
forward and central fuselage and connected to the shell structure using inter-
polation elements (Nastran element type RBE3). These elements compute the
displacement of the connected concentrated point mass as a weighted average
of a supplied set of mesh nodes. In the current case, all three masses were
E 10 D. Eller and U. Ringertz
connected to all mesh nodes in the lower half of the forward fuselage. In this
manner, the resulting inertial forces were eectively distributed among a large
number of shell elements, thus avoiding locally concentrated forces. Neverthe-
less, it eventually turned out that even the load-carrying pilot seat shells and
some minor local reinforcements in the lower cockpit shell needed to be mod-
eled in order to avoid that local shell deformation modes (breathing modes)
appear at unrealistically low frequencies.
Figures 5 and 6 show the front view of the undeformed mesh and the same
view of the deformed surface for the 14th mode. The 14th mode is dominated
by antisymmetric in-plane bending of the wing but contains a signicant com-
ponent of lateral fuselage bending. At the relatively high frequency of 24.3 Hz,
this fuselage bending combines with signicant local shell deformations.
X
Z
X
Z
Figure 5: Undeformed mesh.
X
Z
X
Z
Figure 6: Breathing mode.
A comparison of the rst two longitudinal (vertical) bending eigenmode
shapes of the fuselage is shown in Figure 7. Note that here, the classication
of modes refers to the isolated fuselage alone and that these modeshapes are
not directly related to the full aircraft eigenmodes discussed above. To make
measurements and numerical results directly comparable, the computed shell
model modeshape was interpolated to the known locations of the accelerometers.
Then, both the measured and computed vector of displacement magnitudes were
normalized to unit length. The choice of normalization is admittedly rather
arbitrary.
Aeroelastic Simulations of a Sailplane E 11
0 1 2 3 4 5 6 7 8
0.4
0.2
0
0.2
0.4
Position [m]
V
e
r
t
i
c
a
l

d
i
s
p
l
a
c
e
m
e
n
t

[

]
1
st
longitudinal bending
0 1 2 3 4 5 6 7 8
0.6
0.4
0.2
0
0.2
Position [m]
V
e
r
t
i
c
a
l

d
i
s
p
l
a
c
e
m
e
n
t

[

]
2
nd
longitudinal bending
Experiment
Shell FE model
Figure 7: Modeshape comparison for two longitudinal fuselage modes.
The normalized eigenmode shapes for the rst two longitudinal modes
match fairly well. In both cases, there is a small parallel shift between ex-
perimental and computed modeshapes, which indicates that the longitudinal
mass distribution is not accurate, while the fuselage stiness is matched well.
In the left part of Figure 7, the location of the forward and rear node of the
rst longitudinal bending mode is shown to lie slightly farther apart in the
experiment than the modal nite element solution predicts. When considering
that both the fuselage mass and the center of gravity match closely between the
experimental setup and the corresponding model, this indicates that the mass
in the real aircraft is less concentrated near the center than in the nite element
model.
Figure 8 shows the rst lateral bending and torsion modes of the fuselage.
In this case, there is a small irregularity at the 3 m position, just behind the
cockpit, where the measured acceleration data shows a localized peak. For the
torsion mode, the irregularity may be related to measurement noise, since at
this position, there are only small displacements from which the torsion angle is
computed by nite dierences. Even the inconsistency between two experiments
performed for the same excitation hints at this explanation.
The irregularity in the rst lateral bending mode occurs at the same posi-
tion. In this case, a repetition of the experiment yields practically the same data.
Since the fuselage shell in this region is locally reinforced with internal frames,
it is considered unlikely that the accelerometer senses local shell deformation.
It rather appears as if the eect of the said reinforcements is not represented
suciently accurately in the FE model. Similar to the longitudinal case, the
E 12 D. Eller and U. Ringertz
0 1 2 3 4 5 6 7
0.4
0.2
0
0.2
0.4
Position [m]
L
a
t
e
r
a
l

d
i
s
p
l
a
c
e
m
e
n
t

[

]
1
st
lateral bending
0 1 2 3 4 5 6 7
0.4
0.2
0
0.2
0.4
Position [m]
T
o
r
s
i
o
n

[

]
1
st
torsion
Experiment
Shell FE model
Experiment 1
Experiment 2
Shell FE model
Figure 8: Modeshape comparison for lateral fuselage modes.
overall match of the lateral eigenmode shapes is fairly good.
Looking at the eigenfrequencies for the isolated fuselage shown in Table 3,
it appears that the shell model of the fuselage yields slightly higher frequencies
for all three bending modes. This dierence could be related to the placement
of concentrated masses near the centerline. As a result, the inertia in bending is
too low for the same total mass. The torsion frequency, on the other hand, is
matched with better accuracy.
f
exp
[Hz] f
s
[Hz] Description
10.8 11.4 1st lateral bending
13.2 13.7 1st longitudinal bending
19.8 19.6 fuselage torsion
32.5 32.9 2nd longitudinal bending
Table 3: Measured and computed eigenfrequencies for the fuselage.
Control system model
For the beam nite element model, control surfaces are modeled as rigid el-
ements connected to the beam by means of rigid bar elements, which allow
control surface rotation as the only degree of freedom. Aileron rotation is as-
sumed to occur about the y-axis instead of the actual aileron hinge axis, which
makes an angle of only 7.2

with the y-axis. The rotational degree of freedom

Aeroelastic Simulations of a Sailplane E 13
is connected to a concentrated mass and rotational inertia corresponding to val-
ues measured by the manufacturer. Since the internal kinematics of the control
system introduces a small stiness, rotational springs are added to constrain the
control surface motion accordingly.
The shell nite element model shown in Figure 2 includes shell models
of the control surfaces. In this case, all surfaces are attached by means of
multiple hinges positioned according to the drawings. Hinge nodes on the
wing, n and stabilizer are locally reinforced with small beam elements in order
to better distribute inertial loads into the rather exible sandwich shell.The
rather detailed modeling has the advantage of including the eect of elastic
control surface motion. For this particular aircraft, the rst elastic eigenmodes
involving aileron torsion and bending appear around 25 Hz, close to the wings
fundamental torsion modes, and may hence be relevant from an aeroelastic
point of view.
Unfortunately, the detailed structural model caused diculties when at-
tempting to interpolate eigenmode shapes to the aerodynamic surface mesh.
The elastic motion of the hinged control surfaces which is part of most relevant
modeshapes could not be represented as a smoothly varying deformation of the
closed aerodynamic surface. Even comparatively small deformations involving
control surface rotation or twist lead to geometric inconsistencies such as locally
self-intersecting surface elements or much too large variation of the local surface
normal. Even when the elastic surfaces were constrained to the wing, stabilizer
and n in order to approximate a xed-stick conguration, the pronounced ex-
ibility of the very light control surface shells caused discontinuous deformations
e.g. at the outboard end of the aileron.
As an alternative, the control surface motion was modeled separately. The
aerodynamic surface mesh was deformed according to the ap deection using a
previously developed procedure which guarantees smooth surface deformations
up to moderate deection angles of at least 15

. The structural model, on the

other hand, was simplied by removing the explicit control surface components
and replacing them with a continuation of the wing, stabilizer and n shell.
In this manner, fully continuous modeshapes could be obtained for which suf-
ciently smooth interpolated deformations of the aerodynamic mesh could be
computed.
In order to account for the mass coupling eect between modal displace-
ments and control surface deections, a point-mass model was used. The equa-
tions of motion in modal form (3) are augmented with the control surface
deections
i
according to
E 14 D. Eller and U. Ringertz

I C
T
C diag(J
,i
)

q

0
0 diag(k
,i
)

q

f
q
M
h

. (5)
Here, J
,i
are the inertia terms associated with control surface rotation, in-
cluding the rotational inertia of the control surface itself and components of
the control system mechanics which are coupled to the ap deection. The cor-
responding stiness terms k
,i
constitute the lower diagonal of the new stiness
matrix. On the right hand side, the excitation force vector is augmented with
the aerodynamic hinge moments M
h
.
Since there are signicant coupling eects between the structural motion
as described by the modal coordinates q and the control surface rotation, o-
diagonal terms C are added to the mass matrix. The coupling term C
i,j
is the
hinge moment experienced by the control surface i due to an acceleration in
mode j. It is computed by considering a point mass m
i
located at the center of
mass of the control surface. Taking r
i
as the vector from the center of mass to its
projection on the hinge axis and a
j
as the translational acceleration experienced
by the center of mass due to an acceleration q
j
in mode j, the coupling term is
found to be
C
i,j
=

h
i
(m
i
a
j
r
i
), (6)
where

h
i
is the normalized direction vector of the hinge axis.
In practice, there are also kinematic coupling eects, i.e. a steady modal
deformation would lead to certain ap deections. Such eects are caused by
the kinematics of the complex internal actuation mechanism, which is not only
dicult to include in a structural model but also nonlinear, at least for some
control surface positions. Kinematic couplings are therefore not present in the
current structural models.
Remarks on structural modeling eort
Creating the structural models required a signicant amount of time. The beam
model is a strongly reduced representation of the structure, so that the deriva-
tion of beam properties involves a certain degree of heuristic approximations
and fairly many iterations until a reasonable match with experimental data can
be obtained. Without GVT data, creating such a model would be even more
dicult. Furthermore, the credibility of simulations performed with a beam
model not validated with experimental data would be poor due to the large
amount of approximations involved.
The process of creating a much more complex shell model revealed that,
with drawings and lamination plans available, a nite element model with rea-
sonably accurate global stiness properties can be created comparatively easily.
Aeroelastic Simulations of a Sailplane E 15
The by far most demanding problem in this context is the realistic modeling
of component connections such as the interface of the fuselage with the detach-
able wing. These points are usually designed with complex local reinforcements
which are not easily included in a shell model and must hence be approximated.
A nite element model with accurate stiness behavior is obviously not su-
cient for dynamic analyses, and mass properties are most likely less well dened
than laminate properties. Unless the mass distribution can be established from
accurate weight and balance calculations, it is unlikely that a credible shell model
can be created without access to vibration testing data.
Aerodynamic model
Two dierent aerodynamic models based on subsonic potential ow theory are
used in this study. The rst model is based on the ZONA6 method implemented
in the commercial code ZAERO [18], which is a frequency domain formulation
similar to the Doublet-Lattice Method (DLM). In Figure 9, the discretization of
the lifting surfaces used with this method is shown.
X
Y
Z
X
Y
Z
Figure 9: Panel mesh used with ZAERO.
The second model uses a time-domain formulation based on linear potential
ow. In contrast to lifting-surface methods, the actual aircraft geometry is
discretized, as illustrated in Figure 10. The mesh is shown in the deformed state
for a positive loadfactor of 4. The computational method is described in detail
in [19].
Comparison of aerodynamic coecients for rigid-body motion
The aerodynamic response of the aircraft to prescribed rigid-body motion is
computed in order to investigate if the numerical methods yield similar results
E 16 D. Eller and U. Ringertz
Figure 10: Surface mesh for time-domain aerodynamic method.
despite the considerably dierent discrete surface representations. Since ZONA6
is formulated in the frequency domain, the results are complex pressure coe-
cient dierences C
p
between upper and lower side of the lifting surfaces as a
function of frequency. In this context, it is common to use the reduced form
of the angular frequency according to
k =
c
2u

, (7)
which is made nondimensional with the reference chord c and the airspeed
u

. Integration of the pressure coecient over the aerodynamic surface yields

integral aerodynamic coecients, which depend on type and frequency of the
unsteady rigid-body motion. Harmonic heave and pitch motion about the center
of gravity of the aircraft were used as prescribed motion.
Frequency-domain coecients cannot be computed directly with a time-
domain method, so that an indirect approach must be taken. Here, time-
marching simulations were performed for a prescribed harmonic rigid-body
motion with reduced frequency k. The coecients are then evaluated by di-
viding the discrete Fourier transform of the output lift coecient time series
by the (known) Fourier transform of the prescribed motion. In this manner,
coecients can be determined at discrete frequencies, which is a quite costly
procedure since a new simulation is necessary for each desired frequency. Fig-
ure 11 shows the results of a number of harmonic excitation simulations next
to ZONA6 results for the same four coecients, namely lift and moment coef-
cient as a function of heave and pitch motion frequency. These time-domain
simulations were performed using timesteps of 2.5 ms, and repeated with a
timestep of 1 ms with the same mesh for comparison. Despite the considerably
higher computational cost, no signicant dierences could be found.
Aeroelastic Simulations of a Sailplane E 17
0 0.5 1 1.5
20
15
10
5
0
5
10
15
Reduced frequency
C
L
h
(
k
)
DLM, real part
DLM, imag. part
timedomain, real
timedomain, imag.
0 0.5 1 1.5
5
0
5
10
15
20
Reduced frequency
C
L

(
k
)
0 0.5 1 1.5
8
6
4
2
0
2
4
6
Reduced frequency
C
M
h
(
k
)
0 0.5 1 1.5
40
30
20
10
0
10
20
Reduced frequency
C
M

(
k
)
Figure 11: Dependence of integral aerodynamic coecients on reduced fre-
quency.
At low to moderate reduced frequencies, both methods yield similar results.
Even the characteristic wavy behavior for the moment coecient response to
harmonic heave motion, C
Mh
(k) is represented well. For k in excess of about
0.8, more dierences appear. A possible reason may be the added mass eect
of the displacement bodies, i.e. wing thickness and fuselage, which is neglected
completely in the ZONA6 model. This eect would normally be assumed to
be very small for low reduced frequencies, it increases however linearly with
frequency and may play a certain role at high reduced frequencies.
Additionally, it should be noted that the time-domain method requires the
explicit enforcement of the Kutta condition due to the location of collocation
points exactly at the trailing edge, which is not the case in the same way for
ZONA6. It is not clear to what extent the kinematic (velocity) formulation
chosen here for the Kutta condition yields accurate results for high reduced
frequencies [20, 21].
For better computational eciency, time-domain simulations can be per-
formed with a non-harmonic input signal, which contains more than a single
frequency. Good results were obtained with a frequency ramp (chirp) signal,
E 18 D. Eller and U. Ringertz
for which the frequency increases linearly with time. Using the output of a simu-
lation with a ramped frequency input, the complex coecients can be computed
at a number of frequencies from the ratio of the Fourier transform of the coef-
cient time series to the Fourier transform of the input signal. Figure 12 shows
lift and moment coecient response to a chirp excitation in heave.
0 0.5 1 1.5
20
15
10
5
0
5
10
15
Reduced frequency
C
L
h
(
k
)
0 0.5 1 1.5
8
6
4
2
0
2
4
6
Reduced frequency
C
M
h
(
k
)
harmonic motion, real
harmonic motion, imag.
identification, real
identification, imag.
Figure 12: Identication from harmonic input and chirp response.
Coecients obtained from the identication procedure are overall fairly
close to those found from single harmonic motion simulations, although small-
scale variations occur between adjacent frequencies. Such a wiggling behavior in
frequency is not realistic and must hence be regarded as computational noise.
The thin solid lines in Figure 12 are approximating 5th order polynomials
tted to the discrete frequency response data. Alternatively, cubic spline ts
with multiple segments can be used to approximate the frequency dependency.
Frequency-domain stability analysis
Substituting the damped harmonic ansatz x = xe
pt
into the aeroelastic equa-
tion of motion (1) yields

Mp
2
+K

xe
pt
= f
a
(t, p, x, . . . ) (8)
for p = + i, where is the damping constant and the angular frequency.
Using a modal subspace approach according to (3), the above becomes

Ip
2
+

qe
pt
= Z
T
f
a
, (9)
where the deformations are now expressed in terms of the structural eigenmode
shapes Z obtained from the solution of the free vibration eigenproblem (2).
Aeroelastic Simulations of a Sailplane E 19
Hence, q is a vector of complex modal (or generalized) displacements. For
small displacements, the modal aerodynamic loads (or generalized aerodynamic
forces) Z
T
f
a
are assumed to depend linearly upon the magnitude of the modal
displacements q, so that (9) can be written as a nonlinear eigenvalue problem
according to

c
2u

2
I p
2
+q

A( p)

q = 0, (10)
where p =
c
2u

+ ik. (11)
Here, q

is the dynamic pressure and A( p) relates the modal aerodynamic

forces to modal displacements. The real part of the nonlinear eigenvalue p is
the nondimensional damping value and its imaginary part corresponds to the
reduced frequency. In terms of actual deformations x, the solution can be
expressed as
x(t) = Z Re

qe
pt

, (12)
where Re() extracts the real part. In order to simplify the computation of
frequency-domain aerodynamic loads, it is often assumed that A depends on
the reduced frequency k alone, and not on the damping. This approximation is
exactly fullled for a purely imaginary solution p
crit
= ik of (10), corresponding
to neutral stability.
Results for the shell nite element model
The nonlinear eigenvalue problem (10) is solved using the robust pkalgorithm
by Back and Ringertz [22], which yields eigenvalues and eigenvectors solving
(10) for a particular dynamic pressure. Figures 13 to 14 show the damping g
and frequency of solutions for airspeeds up to 160 m/s, where the damping is
dened as
g = 2
Re(p)
Im(p)
. (13)
This type of diagram is usually called vg-plot. As the nonlinear eigenvalue
solver does not track modes but instead employs a more robust scheme based
on sorting eigenvalues by reduced frequency, the line colors in the graphs below
identify the position of a solution in the frequency order.
In Figure 13, solutions to the eigenvalue problem are shown for the shell
FE model in combination with aerodynamic loads A(k) computed using the
ZONA6 method implemented in ZAERO 6.2 [18]. Although the load matrices
are computed using this software, the pksolution is performed externally us-
ing a Matlab implementation of the algorithm by Back and Ringertz [22]. The
E 20 D. Eller and U. Ringertz
50 100 150
0.2
0.15
0.1
0.05
0
0.05
0.1
Airspeed [m/s]
D
a
m
p
i
n
g

[

]
50 100 150
0
5
10
15
20
25
30
Airspeed [m/s]
F
r
e
q
u
e
n
c
y

[
H
z
]
Figure 13: Damping and frequency for shell model and ZONA6 aerodynamics.
shell nite element model used is the simplied version which does not con-
tain structurally modelled control surface mechanisms. Aerodynamic eects of
control surface motions are not modelled either.
Figure 14 shows results computed using aerodynamic loads obtained from
the identication process described previously, applied to modal displacements
instead of rigid body motion. Time-domain simulations are performed for
50 100 150
0.2
0.15
0.1
0.05
0
0.05
0.1
Airspeed [m/s]
D
a
m
p
i
n
g

[

]
50 100 150
0
5
10
15
20
25
30
Airspeed [m/s]
F
r
e
q
u
e
n
c
y

[
H
z
]
Figure 14: Damping and frequency for identied aerodynamics matrices.
prescribed motions in one or two modal coordinates, and the time history of the
modal aerodynamic forces is recorded. The input signal is a sine which linearly
increases in frequency up to 40 Hz, corresponding to a reduced frequency of
1.47. One symmetric and one antisymmetric mode are excited simultaneously
in each simulation, and the resulting modal forces are partitioned according
Aeroelastic Simulations of a Sailplane E 21
to symmetry, relying on the assumption that symmetric deformations do not
cause antisymmetric loads and vice versa. After the simulations are completed,
complex frequency-domain loads are computed from the ratio of the Fourier
transforms of output (modal force) and input (modal displacement) signals.
Figures 13 and 14 indicate that both approaches to compute frequency-
domain aerodynamic loads give similar utter results. In Table 4, the airspeeds
for the rst two critical modes are listed with their (absolute and reduced) fre-
quencies at the utter boundary (g = 0). Note that the computed instabilities
occur far beyond the ight envelope of the ASK 21, which is certied for a
maximum permitted ight speed of 78 m/s.
Mode ZONA6 Identication Dierence
1 airspeed 126 m/s 126 m/s <1%
frequency 9.65 Hz 9.52 Hz -1.3%
k 0.254 0.251
2 airspeed 150 m/s 150 m/s <1%
frequency 16.1 Hz 15.8 Hz -1.9%
k 0.356 0.340
Table 4: Critical aeroelastic modes
While there are small dierences in the frequencies of the critical modes,
the overall behavior is quite similar. Some of the aeroelastic modes show slightly
more signicant dierences. As an example, consider the mode which becomes
unstable at 150 m/s. While the aeroelastic system using identied aerody-
namic frequency-domain loads predicts that this particular mode will achieve
its strongest damping of g = 0.18 at about 130 m/s, the model employing
ZONA6 aerodynamics predicts only g = 0.13 at the same airspeed. For air-
speeds slightly larger than 130 m/s, the same mode quickly becomes less strongly
damped.
Figure 15 to 18 show the rst critical utter modeshape for each of the two
aerodynamic methods. Since aeroelastic modes q are complex-valued, real and
imaginary part of the mode are plotted separately. The real part can be under-
stood as the deformation shape at the beginning of the neutrally stable utter
oscillation, while the imaginary part would be the shape after one quarter of a
period, according to (12). Comparing the modeshapes obtained using dierent
aerodynamic methods, considerable similarities in the deformation pattern can
be found. Symmetric wing bending is visible in both utter modes. The bending
pattern is a combination of rst and second fundamental wing bending in such
a way that mainly the outboard third of the wing appears to move. Vertical
E 22 D. Eller and U. Ringertz
Figure 15: Real part of the rst critical
utter mode (ZONA6).
Figure 16: Imaginary part of the rst
critical utter mode (ZONA6).
Figure 17: Real part of the rst critical
utter mode (identied loads).
Figure 18: Imaginary part of the rst
critical utter mode (identied loads).
fuselage bending and a minor component of symmetric stabilizer bending are
also present in both modeshapes. Even the ratio of fuselage to wing bending
appears to be similar for the two methods. However, the phase between fuselage
bending and wing tip motion is slightly dierent, as the comparison of the
imaginary part of the modes shows. Since, in this case, the same structural
model is used, the dierences must either be caused by dierent interpolation
of structural eigenmode shapes to the aerodynamic mesh, or by dierences in
the computation of modal aerodynamic forces.
The interpolation procedure employed by ZAERO is based on the tting
of a plate spline to a subset of structural mesh nodes. Visualization of the
ZONA6 mesh, deformed according to structural eigenmodes reveals that the
quality of the spline interpolation procedure depends on the chosen node subset.
Subsets with too many structural nodes lead to wavy interpolated deformation
shapes, while too small node sets do not dene the structural deformation
with sucient accuracy. For the results shown above, only a small number of
structural, i.e. only 60 of 3800 nite element nodes per half-wing were used, in
order to suppress waviness in the interpolated modeshapes. The selected nodes
cover most of the structural surface, but are not regularly spaced in any way,
which may decrease interpolation accuracy.
For the time-domain aerodynamics used in the identication process, the
deformation of the surface mesh is computed from the shell FE model by as-
Aeroelastic Simulations of a Sailplane E 23
suming that surface mesh points are connected rigidly to the nearest structural
element. Since the distance is very small (typically below 3 mm), the surface
mesh follows the shell model surface very closely except where the structural
model is not dened. Such small gaps in the FE model exist e.g. between wing
root rib and fuselage. Due to the dierent interpolation method, the deformed
shape can dier signicantly from the deformation of the ZONA6 mesh for the
same modeshape. Such dierences are more likely to occur for higher frequency
modes, where local shell deformations become important.
The rst and fourth structural eigenmode contain deformations which ap-
pear prominently in both utter modeshapes, namely rst wing bending and
rst vertical fuselage bending. In Figure 19, frequency-domain modal aerody-
namic loads computed using the two methods are compared for reduced fre-
quencies up to 1.5.
0 0.5 1 1.5
2
1.5
1
0.5
0
0.5
1
Reduced frequency
A
1
,
1
DLM, real part
DLM, imag. part
Response, real
Response, imag.
0 0.5 1 1.5
0.04
0.02
0
0.02
0.04
0.06
0.08
Reduced frequency
A
1
,
4
0 0.5 1 1.5
0.04
0.03
0.02
0.01
0
0.01
0.02
Reduced frequency
A
4
,
1
0 0.5 1 1.5
0.15
0.1
0.05
0
0.05
0.1
Reduced frequency
A
4
,
4
Figure 19: Four elements of A(k) compared.
Both methods generate similar results for the elements A
1,1
(k) and A
4,4
(k),
i.e. the aerodynamic loads in modes one and four due to harmonic excitation
of the mode itself, for reduced frequencies up to 1.2. Much less similarity can
be found in the o-diagonal terms A
1,4
(k) and A
4,1
(k). The imaginary part
E 24 D. Eller and U. Ringertz
of the former shows quite signicant deviations, which appear to be systematic
dierences in the aerodynamic model, as the values obtained from the discrete
Fourier transform are not too far from some smooth function of frequency. In
the latter case, however, the identication process using the FFT failed. Real
and imaginary part of the response are found to vary wildly with frequency
and deviate strongly from values computed with the ZONA6. Such an erratic
behavior indicates that the modal force response to chirp excitation for the
corresponding time-domain simulation is either noisy or does not resemble the
input signal suciently.
Figure 20 shows the time history of aerodynamic forces in mode 1 and 4
due to a chirp excitation of the rst structural eigenmode. While the response
0 0.2 0.4 0.6 0.8 1
3000
2000
1000
0
1000
2000
3000
Time [s]
F
q
,
1
0 0.2 0.4 0.6 0.8 1
220
240
260
280
300
320
340
Time [s]
F
q
,
4
Figure 20: Time-domain response to excitation of the rst mode.
of the rst mode itself is smooth and directly correlated to the modal veloc-
ity of the excitation signal, the force in the fourth mode is of much smaller
amplitude and more noisy. Furthermore, the time history looks more like a
response to two superimposed excitation signals. Since such a signal clearly is
not properly correlated to a single input, the simple identication procedure
fails to yield reasonable results. In this particular case, the level of noise may
be related to the extreme dierence in magnitude between modal force compo-
nents. The large values of modal force in the rst mode indicate that the surface
pressure distribution must dier substantially from the equilibrium state. Since
the modal force amplitude in the fourth mode is about 40 times smaller, the
apparent superposition in this mode can be related to a numerical integration
error. This problem is also encountered when computing small integrated force
coecients such as induced drag in the presence of large pressure dierences.
Aeroelastic Simulations of a Sailplane E 25
Results for the beam nite element model
Finally, the same utter analysis as above has been performed with the shell
nite element model replaced by the beam model and ZONA6 aerodynamics.
Damping and frequency over velocity are shown in Figure 21. Interestingly,
there is no instability in the speed range of interest, and even the development
of the eigenfrequencies of the aeroelastic modes is quite dierent. In contrast
to the utter solution based on the shell model, almost all eigenfrequencies
below 30 Hz rise with airspeed. With the shell FE model, several of the higher
frequency eigenmodes show a distinct drop in frequency above 100 m/s, with
possible divergence in the vicinity of 160 m/s.
50 100 150
0.2
0.15
0.1
0.05
0
0.05
0.1
Airspeed [m/s]
D
a
m
p
i
n
g

[

]
50 100 150
0
5
10
15
20
25
30
Airspeed [m/s]
F
r
e
q
u
e
n
c
y

[
H
z
]
Figure 21: Damping and frequency for beam FE model.
The very signicant dierences between utter solutions using the beam
and the shell models emphasize the utmost importance of an accurate nite
element model for aeroelastic analysis. That does not necessarily mean that a
high-delity shell model must be used in every case. As seen in Table 1, the
dierences in modal properties of beam and shell model cannot be regarded as
drastic, but still the utter solutions show very few similarities. This indicates
that an aeroelastic analysis may not be meaningful unless the structural model
reproduces the actual ying structure very accurately.
Maneuver simulation
Due to the diculties and risks associated with utter ight testing, a more
benign maneuver ight test serves to compare aeroelastic simulation results with
experimental data. The ight maneuver is a pull-up with a maximum load
factor of about 4, close to the saturation limit of the installed attitude and
E 26 D. Eller and U. Ringertz
heading reference system (AHRS). The test aircraft is the ASK 21 shown in
Figure 1 which is equipped with position sensors at the control surfaces, solid-
state AHRS and tted with an air data boom with vanes to measure angle of
attack and sideslip. All of these sensors are currently sampled with 10 Hz using
a digital on-board data acquisition computer. Furthermore, 16 piezoelectric
accelerometers are installed in the aircraft, which are sampled with 1 kHz.
The recorded elevator deection is used as input for the numerical sim-
ulations, where its value is shifted by a constant in order to achieve initial
equilibrium conditions. The rst simulation is performed with a exible air-
craft model using the shell nite element model with eigenmodes up to 40 Hz.
A further numerical simulation is performed with exactly the same simulation
code, but without accounting for exibility. The only dierence in the input
signal between the two simulations is in the denition of the initial state, as
both computations must be started from an equilibrium state. This requires a
dierent initial elevator deection (shifted by +2.3

) for the rigid model, and

a slightly smaller angle of attack (reduced by 0.8

). Moreover, both numeri-

cal simulations employ a simple model for the parasite drag which is added
to the induced drag obtained from surface pressure integration. Drag data was
extracted from the measured speed polar diagram in the aircraft datasheet [10]
and extrapolated for airspeeds in excess of 200 km/h. Since the maneuver is
performed in the vertical plane only, the numerical simulation of the ight me-
chanics was limited to longitudinal states, and lateral velocities and angles were
held xed.
The time history of the pull-up in terms of altitude, calibrated airspeed and
load factor is shown in Figures 22 24. Measured airspeed and altitude are de-
0 2 4 6
8
6
4
2
0
2
Time [s]
E
l
e
v
a
t
o
r

d
e
f
l
e
c
t
i
o
n

[
d
e
g
]
0 2 4 6
20
10
0
10
20
30
40
50
Time [s]
F
l
i
g
h
t

p
a
t
h

a
n
g
l
e

[
d
e
g
]
Flight test
Simulation, flexible
Simulation, rigid
Figure 22: Measured elevator deection and processed ightpath angle.
Aeroelastic Simulations of a Sailplane E 27
rived from pressure measurements using standard atmosphere conditions, while
the ight path angle in Figure 22 is computed from low-pass ltered altitude
and velocity data. The latter should therefore be understood as approximate be-
cause the angle is not measured directly. However, the initial short drop by 5

as
well as the steep change at t = 5.5 s hint at the presence of thermal drafts, since
comparison with the elevator position in the same gure indicate that these
rapid changes are not related to the commanded maneuver. Figure 23 shows
pressure altitude and airspeed, which are both reasonably well reproduced by
the numerical simulations. After six seconds simulated ight, the accumulated
0 2 4 6
140
160
180
200
220
240
Time [s]
A
i
r
s
p
e
e
d

[
k
m
/
h
]
Flight test
Simulation, flexible
Simulation, rigid
0 2 4 6
420
440
460
480
500
520
540
Time [s]
A
l
t
i
t
u
d
e

[
m
]
Figure 23: Time history for pressure altitude and calibrated airspeed.
error reaches about 10 km/h and 15 m in altitude. The measured altitude data
contains a sudden jump of about 5 m at t 0.8 s, which may be caused by pres-
sure measurement noise. Accounting for the use of an approximate drag model,
the accuracy of the numerical simulation in terms of velocity and altitude is
considered reasonable.
Figure 24 shows load factor and pitch rate as obtained from ight test, along
with values computed by numerical simulation with rigid and exible aircraft.
The global development of load factor and pitch rate is matched quite accurately
by the numerical time-domain simulation. In particular, the accumulated error
after completion of the maneuver is surprisingly small. It was expected that
dierences between computed and actual control surface eectiveness would
lead to more signicant error accumulation over the simulated 3000 timesteps.
The magnitude of vertical accelerations caused by thermal drafts can be esti-
mated by considering the measured load factor at t 5.5 s, just before the end
of the time window. Here, the load factor drops by about 0.2 for less than 0.5 s,
while the recorded elevator deection is essentially at. Consequently, neither
E 28 D. Eller and U. Ringertz
of the simulations contains this load excursion. Contrary to expectations, the
0 2 4 6
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
Time [s]
L
o
a
d

f
a
c
t
o
r

[

]
0 2 4 6
10
5
0
5
10
15
20
25
30
Time [s]
P
i
t
c
h

r
a
t
e

[
d
e
g
/
s
]
Flight test
Simulation, flexible
Simulation, rigid
Figure 24: Load factor and pitch rate.
simulation of the elastic aircraft did not show large dierences from the rigid
aircraft ight simulation. Only during the second half of the maneuver, an
error in pitch rate appears to accumulate for the rigid case. At the end, the exi-
ble aircraft simulation still matches the experimentally obtained pitch rate fairly
well, while the rigid aircraft simulation is o by -3

/s. Similarly, the ight path

angles of the two simulations diverge in the second half of the maneuver, to
reach a signicant dierence of about 10

at the end of the time window. How-

ever, since direct measurements for this angle are not available, it is dicult to
conclude if the elastic aircraft simulation improved accuracy.
The test aircraft is tted with piezoelectric acceleration sensors which are
primarily intended for vibration testing. Each sensor contains a capacitance-
resistance circuit acting as a high-pass lter with a time constant of one second.
In order to enable interpretation of the measured data over the relatively long
time range considered, the raw accelerometer data was processed by ltering
with the inverted high-pass lter. This process introduces a linear drift since
the raw electrical accelerometer data also contains a constant feed voltage signal.
Therefore, the corresponding linear term was subtracted from the acceleration
signal.
Figure 25 shows a comparison of the time history of measured wing tip
acceleration, when the cockpit acceleration signal as provided by the AHRS is
subtracted. This dierence can be regarded as an approximation of the tip mo-
tion due to elastic deection. Furthermore, acceleration data for the simulation
is shown. It is obtained by nite-dierencing the velocity of a mesh node at the
accelerometer position.
Aeroelastic Simulations of a Sailplane E 29
0 2 4 6
10
5
0
5
10
Time [s]
A
c
c
e
l
e
r
a
t
i
o
n

[
m
/
s
2
]
Flight test
Simulation
0 10 20 30 40
40
35
30
25
20
15
10
5
0
5
Frequency [Hz]
P
o
w
e
r

s
p
e
c
t
r
a
l

d
e
n
s
i
t
y

[
d
B
/
H
z
]
Figure 25: Time history and PSD estimate for wing tip acceleration.
The measured time history contains considerably larger acceleration ampli-
tudes than simulation data. Moreover, the strong peaks do not correlate to the
known ight mechanic motion in an obvious manner. The acceleration data
from the simulated maneuver shows a considerably smaller variation and even
allows to recognize two points with particularly strong signals, i.e. pronounced
wing bending and torsion motion, at t = 1.5 s and t = 3.5 s. These points
correspond to the beginning and the end of the pull-up maneuver as seen in
Figure 24, where the load factor varies rapidly. There may be a similar pattern
in the measured data, but that is somewhat masked by the much stronger and
apparently noise-related peaks.
The right part of Figure 25 shows the Yule-Walker power spectral density
(PSD) estimate of the time series shown to the left. This estimate is the PSD of
a discrete autoregressive time-domain model tted to the time series data and
yields spectrum information which contains much less variation in frequency
than the raw FFT. It is thus easier to identify the frequency content of the
corresponding signals, although the absolute value of the PSD would not be
directly useful, as it depends somewhat on the parameters used in the method.
Here, the autoregressive model was chosen to span 320 milliseconds, that is
order 320 for the measured data and order 160 for the accelerations obtained
from simulation.
First of all, the measured data comprises signicantly higher magnitudes
throughout most of the frequency range considered, which is expected from the
time history. In addition, its PSD is considerably less varying, which indicates
that random processes (such as atmospheric turbulence) could be responsible
for some of the large magnitudes. The PSD estimate of the simulated accelera-
E 30 D. Eller and U. Ringertz
tions, on the other hand, contains several articulated peaks at 10 Hz, 20.5 Hz,
23.5 Hz, 28 Hz and 30 Hz. Most of these frequencies can be identied with
structural eigenfrequencies associated with symmetric modes (see Table 1), which
is expected since the representation of the structural behavior underlying the nu-
merical simulation is based on the same modal data.
Computational eort
Structural modal analysis of the NASTRAN shell model shown in Figure 2 con-
tributed only a very minor amount of computational work; the extraction of the
rst 30 eigenmodes was usually computed within 5 minutes. Evaluation of the
aerodynamic inuence coecient matrices required for the utter analysis using
ZAERO took approximately 2 hours on the same machine. This computation
need only be performed when the geometry or discretization of the panel mesh
has changed.
Time-domain simulations for the evaluation of frequency-domain load ma-
trices using the chirp response approach were far more computationally inten-
sive. The results presented here required approximately 11 hours computation
time on a dual-processor 2.4 GHz Opteron computer. On the same machine,
the maneuver simulation of the exible aircraft with 3000 timesteps ran for 26
hours. The current mesh-deformation procedure is adapted for arbitrary struc-
tural deformations and therefore performs a number of computations which
may not be necessary if deections remain small. For that particular case, fur-
ther improvements of computational eciency are possible.
Conclusions
When comparing aeroelastic stability analysis performed with a beam and a
shell nite element model of the same aircraft, considerable dierences in aeroe-
lastic stability are found although the structural models match fairly well in
the eigenfrequencies up to 15 Hz. Even though the rst utter mode detected
in the analyses involving the shell models is close to 10 Hz, the aeroelastic
system based on the beam model is stable throughout the whole speed range
investigated. This behavior leads to the conclusion that aeroelastic stability is
highly sensitive to dierences in the structural model, even if these dierences
are limited to eigenmode shapes with moderate and high frequencies.
A time-domain boundary element method was used to identify frequency-
domain aerodynamic loads from ramped frequency response simulations. In
most cases, the aerodynamic behavior thus obtained was found to dier lit-
tle from the aerodynamic loads computed by the frequency domain method
Aeroelastic Simulations of a Sailplane E 31
ZONA6. Flutter computations performed using the pkmethod and with both
types of aerodynamic loads showed closely matching utter speeds and similar
utter modeshapes. However, the accuracy of the identication method degrades
signicantly for modal aerodynamic loads of small relative magnitude and the
computational cost of the identication process is comparatively high.
Finally, a 4g pull-up maneuver was performed with an instrumented ight
test aircraft. Numerical simulations which used the measured time history of
the elevator deection as input demonstrated that the ight mechanic behav-
ior of the sailplane was represented fairly well by the simulation. Comparing
simulations including elastic deformations with those for the rigid aircraft, dif-
ferences were found to be present, but smaller than anticipated from the large
load factor.
Since control system modeling is known to have a strong eect on aeroelas-
tic stability, in particular if the control surfaces are not mass balanced, future
eorts in this project will be concerned with the accuracy of control surface
aerodynamics and the mechanical properties of the control system. It is not
unusual that mild instabilities are tolerated in the certication analysis if the
damping required to stabilize is very small. A positive (unstable) value of the
damping parameter g less than 0.03 is sometimes accepted to account for un-
modeled damping in the structure and friction in the mechanical control system.
However, the damping dened by (13) is physically very dierent in comparison
to the stick-slip type of Coulomb friction that typical control surfaces experi-
ence. Accurate modeling of Coulomb friction requires a nonlinear time-domain
analysis for which the developed aerodynamic model is very suitable.
Acknowledgments
The authors would like to thank Michael Greiner and Gerhard Waibel of Alexan-
der Schleicher Flugzeugbau for their valuable support in providing design draw-
ings, lamination plans and even rotational inertia data for control system com-
ponents. Sebastian Heinze implemented and tested the interpolation spline data
required to use the shell nite element model with ZAERO.
References
[1] E. H. Dowell, H. C. Curtiss, R. H. Scanlan, and F. Sisto. A modern course
in aeroelasticity. Kluwer, Dordrecht, 1989.
[2] R. L. Bisplingho, H. Ashley, and R. L. Halfman. Aeroelasticity. Dover
Publications, 1955.
E 32 D. Eller and U. Ringertz
[3] W. Stender and F. Kiessling. Aeroelastic utter prevention in gliders and
small aircraft. Technical report, DFVLR e.V., Gottingen, 1990.
[4] W. Stender. Praxisnahe Abschatzungs- und Vorbeugungsmoglichkeiten
gegen die Flattergefahrdung von Segelugzeugen und kleinen Motor-
ugzeugen. Technical Report IB 151-74/6, 151-74/20 and 151-74/22,
DFVLR e.V., Braunschweig, October 1974. Part I-III.
[5] European Aviation Safety Agency. CS-22, Certication specications
for sailplanes and powered sailplanes, November 2003. Available at
www.easa.eu.int.
[6] T. Brama. The structural optimization system OPTSYS. International Series
of Numerical Mathematics, 110:187206, 1993.
[7] W. P. Rodden and E. H. Johnson. Msc.Nastran Aeroelastic Analysis Users
Guide. MacNeal-Schwendler Corp., Los Angeles, 1994.
[8] V. J. Stark. The AEREL utter prediction system. In ICAS-90-1.2.3, 1990.
[9] ZONA Technology. ZAERO Version 6.2 Theoretical Manual, 17th edition,
October 2002. Available at www.zonatech.com.
[10] Alexander Schleicher Flugzeugbau, Poppenhausen. Datasheet for the ASK 21,
July 2003. Available at www.alexander-schleicher.de.
[11] A. Keller. Aeroelastic model development for the ASK 21 glider aircraft.
Masters thesis, AVE, Kungliga Tekniska Hogskolan, Stockholm, October
2004.
[12] A. Westergren. Structural dynamics of the ASK 21 glider wing. Masters
thesis, AVE, Kungliga Tekniska Hogskolan, Stockholm, June 2003.
[13] O. Bergogne. A simulation model of the ASK 21. Technical report, AVE,
Kungliga Tekniska Hogskolan, Stockholm, August 2003.
[14] R. Lardet. Flight dynamics model of the ASK 21 sailplane. Masters thesis,
AVE, Kungliga Tekniska Hogskolan, Stockholm, May 2003.
[15] MacNeal-Schwendler Corp. Nastran Reference Manual, 2004. Msc.Nastran
2004.
[16] U. Nilsson, M. Norsell, and U. Carlsson. Ground vibration testing of
the ASK 21 glider aircraft. Technical Report TRITA/AVE 2003:32, AVE,
Kungliga Tekniska Hogskolan, Stockholm, May 2003.
Aeroelastic Simulations of a Sailplane E 33
[17] N. Niedbal and A. Bertram. Flatteruntersuchung ASK 21. Technical report,
DFVLR e.V., Gottingen, 1979.
[18] P. C. Chen, H. W. Lee, and D. D. Liu. Unsteady subsonic aerodynamics
for bodies and wings with external stores including wake eort. Journal of
Aircraft, 30(5):618628, September 1993.
[19] D. Eller. An ecient boundary element method for unsteady low-speed
aerodynamics in the time domain. Technical Report TRITA/AVE 2005:40,
AVE, Kungliga Tekniska Hogskolan, Stockholm, December 2005.
[20] J. Katz and D. Weihs. Wake rollup and the Kutta condition for airfoils
oscillating at high frequency. AIAA Journal, 19(12):16041606, December
1981.
[21] J. Katz and A. Plotkin. Low Speed Aerodynamics. Cambridge University
Press, second edition, 2001.
[22] P. Back and U. T. Ringertz. Convergence for methods for nonlinear eigen-
value problems. AIAA Journal, 35(6):10841087, June 1997.