A Method for Efficient Flutter Analysis

of Systems with Uncertain Modeling

Parameters

MIKAELA LOKATT

Master of Science Thesis

Stockholm, Sweden 2014
A Method for Efficient Flutter Analysis of
Systems with Uncertain Modeling
Parameters

MIKAELA LOKATT

Master’s Thesis in Systems Engineering (30 ECTS credits)

Master Programme in Aerospace Engineering (120 credits)
Royal Institute of Technology year 2014
Supervisor at KTH was Ulf Ringertz
Examiner was Per Engvist

TRITA-MAT-E 2014:19
ISRN-KTH/MAT/E--14/19--SE

Royal Institute of Technology

School of Engineering Sciences

KTH SCI
SE-100 44 Stockholm, Sweden

URL: www.kth.se/sci
Abstract
A method for efficient analysis of variations in flutter speed predictions caused by pa-
rameter variations in the mass, stiffness and aerodynamic models of a wing structure is
presented. The analysis method considers a linear uncertainty formulation and performs
a perturbation analysis based on computation of eigenvalue differentials of a nominal sys-
tem matrix. The method is applied in a test case study in which flutter speed variations
caused by variations in mass and aerodynamic properties of a delta wing model is ana-
lyzed. The report is concluded with a discussion of the validity of the results and of how
the applicability of the method is affected by the assumptions on which it builds.

1
En Metod för Effektiv Uppskattning av Fladderhastigheten för
System med Osäkra Parametrar
Sammanfattning
En metod för effektiv uppskattning av fladderhastigheten för system med osäkra para-
metrar i modellerna för massa, styvhet och aerodynamik presenteras. Metoden förutsätter
en linjär osäkerhetsmodell och uppskattar den lägsta fladderhastighet som kan orsakas av
parametervariationerna genom en analys som bygger på differentiering av egenvärden.
Metoden appliceras i en studie av en delta vinge där skillnader i fladderhastighet orsa-
kade av massvariationer och aerodynamiska osäkerheter analyseras. Rapporten avslutas
med en diskussion om resultaten från teststudien och om hur applicerbarheten av metoden
påverkas av de antaganden på vilka den bygger.

2
Contents
1 Introduction 4

2 Stability of the Nominal System 6

2.1 Modal Projection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Polar Formulation of the Modal Flutter Equation . . . . . . . . . . . . . . 6
2.3 Discretization into r-intervals . . . . . . . . . . . . . . . . . . . . . . . . . 7

3 Stability of the Perturbed System 8

3.1 Aerodynamic uncertainties . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2 Mass and stiffness uncertainties . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3 Combined uncertainties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

4 Flutter analysis of a delta wing model 11

5 Discussion 15

6 Conclusions 16

3
1 Introduction
Aircraft flying at high speed are prone to encounter so called flutter. It is a dynamic
instability which is caused by a complex interaction between aerodynamic, elastic and in-
ertial forces. The interaction between the forces causes the aircraft structure to perform
periodic, undamped oscillations, which may both cause structural damage and impose
a problem for the control system. Since the damage caused by the oscillations may be
costly, and even dangerous, it is highly desirable to be able to predict the flutter speed
already during the design process. This so that one can choose a design for which the
flutter speed is outside the speed envelop in which the aircraft is to be operated.

Throughout the years, a number of methods for flutter predictions have been developed,
see for example [1],[2],[3]. Some of them build on linearized models of the aircraft dynam-
ics and are usually able to give fairly accurate predictions of the flutter speed. However,
due to various simplifying assumptions made during the modeling process, there will most
likely be deviations between the modeled dynamics and the dynamics of the real aircraft.
As a consequence, there will also be deviations between the predicted flutter speed and
the actual flutter speed. The magnitude of the speed difference can be estimated by com-
puting the flutter speed for a number of parameter perturbations, in search of the worst
case perturbation giving the largest speed difference. However, the number of perturba-
tion combinations to be evaluated can quickly grow large and it may thus be desirable
to instead use some method which considers all possible combinations at the same time
and computes a worst case flutter speed based on the perturbation shapes and boundaries.

Throughout the last decades an area called robust flutter analysis has been developed
for this purpose. The robust analysis methods, see for example [4],[5] and [6],[7], consider
a nominal model subjected to a number of uncertain parameters and aim at answering
the question if these uncertainties could destabilize a nominally stable system. In theory,
this sort of analysis could be very efficient for analyzing if, or proving that, the flight
envelop is free of flutter instabilities. In practice, these methods are indeed very useful for
analyzing how the system responds to various perturbations, but proving that a flutter
instability could not occur in a certain speed regime may not be as simple.

One reason for this is that, for the robust stability results to be accurate, the uncer-
tainty model must cover the true parameter variations both regarding magnitude and
shape. When considering mass and stiffness uncertainties, this may not be a problem
since one usually has a relatively good idea about the mass variations of various com-
ponents and also relatively accurate material data and geometry information about the
aircraft structure. But a non-negligible part of the modeling uncertainties usually come
from the aerodynamic model, for which there may be little knowledge about both distri-
bution and size of the modeling uncertainties. Thus, it may be difficult to know if the
chosen uncertainty model covers the actual variations and, consequently, also difficult to
know if the results from the robust analysis are accurate.

The example above aims at illustrating that robust stability may be difficult to guar-
antee in a practical application. The main advantage of the robust methods may thus be
that they provide an efficient way of including several parameter variations at the same
time when analyzing a system’s response to modeling uncertainties. But it may then be

4
an alternative to use some method for a more efficient perturbation analysis instead, if
that could allow a wider variation of modeling uncertainties to be taken into account and
perhaps also be more computationally efficient.

Such a method for perturbation analysis is presented in this study. The method ex-
ploits the diagonal structure of the modal flutter equation in order to perform a pertur-
bation analysis based on computation of eigenvalues and eigenvalue differentials. The
method allows a general linear uncertainty formulation and efficiently computes pertur-
bation boundaries for the speed variations. The method is applied in a test case study
of a delta wing model in which worst case flutter speeds predicted by the perturbation
method, for parameter variations within certain bounds, are compared to worst case
speed variations found from a number of nominal analyzes for various combinations of
parameter variations within the same bounds. The study is concluded with a discussion
about various advantages and limitations of the analysis method and with suggestions for
further development.

5
2 Stability of the Nominal System
The nominal flutter equation can be written as
"  2 #
b 1
p2 M + K − ρb2 Q(p) η = 0 (1)
u 2

where M , K and Q represent the mass, stiffness and aerodynamic matrices respectively,
p = g + ik is the reduced Laplace variable, b is the aerodynamic reference length, u is the
airspeed and ρ is the density of the air.

The system (1) is formulated in the Laplace domain and can be seen as a non-linear
eigenvalue problem. The Laplace variable p represents an eigenvalue and η represents the
corresponding eigenvector. The system is said to be stable if all eigenvalues that makes
the flutter matrix singular have negative real part. This since these solutions correspond
to a stable motion with decreasing amplitude in the time domain. If some eigenvalue has
a real part g ≥ 0 the system is said to be nominally unstable. This since those solu-
tions correspond to an unstable motion with constant or increasing amplitude in the time
domain.

2.1 Modal Projection

The mass and stiffness matrices are usually obtained from a finite element model of the
aircraft structure and the dimension of the flutter problem (1) can thus be fairly large. In
order to reduce the size of the flutter problem the solution vector η can be approximated
using a so called modal basis. This does not only reduce the computational time, but it
also gives the resulting modal mass and stiffness matrices a certain structure [8].

The modal basis consists of a number of eigenvectors to the structural eigenvalue problem

[−ω 2 M + K]z = 0. (2)

where z is an eigenvector and ω 2 is an eigenvalue. If denoting the corresponding modal

basis Z = [z 1 z 2 ...z n ] and letting η = Zv, the resulting modal mass and stiffness matrices
can be written as Z T M Z = I and Z T KZ = Ω [8]. Here, I represents an identity matrix
and Ω = diag(ωn2 ). The aerodynamic modal matrix is denoted by Z T QZ = Q. Thus,
the modal flutter equation becomes
"  2 #
b 1
p2 I + Ω − ρb2 Q(p) v = 0. (3)
u 2

2.2 Polar Formulation of the Modal Flutter Equation

In this section, the property Z T M Z = I will be exposed to derive an alternative formu-
lation of the modal flutter equation (3). Using this formulation, the stability of (3) can
be analyzed by computation of eigenvalues of an ordinary linear system.

For flutter analysis, a usual approximation is to let Q(p) ≈ Q(k) close to the imagi-
nary axis [8]. If making a change of variables so that p = reiθ then close to the imaginary

6
axis k ≈ r and so Q(p) can be approximated by Q(r) in this area. Thus, (3) can be
written "  2 #
b 1
r2 ei2θ I + Ω − ρb2 Q(r) v = 0. (4)
u 2
When r 6= 0, (4) can be reformulated as
"  2 #
1 b 1
ei2θ I + 2 (Ω − ρu2 Q(r)) v = 0. (5)
r u 2

The first term in (5) is a function of θ only and, since it is the identity matrix multiplied
by a complex number, any v that is multiplied by it will become ei2θ v. Thus (5) is equal
to    
i2θ 1 b 1 2
−e v = 2 (Ω − ρu Q(r)) v. (6)
r u 2
h
2 i
Letting A(u, r) = r12 ub (Ω − 12 ρu2 Q(r)) (5) can be written as

−ei2θ v = A(u, r)v. (7)

Since | − ei2θ | = 1, this means that for (5) to have a solution, A has to have an eigenvalue
λ with magnitude equal to one. Since A is a function of u, r and does not depend on θ,
this also means that if there exists a λj with |λj | = 1 for some u, r, then it is possible to
find a solution to (5). This since −ei2θ can represent any complex number with magnitude
equal to one and θ thus can be chosen so that −ei2θ = λj .

A stability criteria can be found by noting that any angle γ in the complex plane can
be written as γ = γ0 + 2πa where 0 < γ0 ≤ 2π and a is an arbitrary, positive or nega-
tive, digit. One thus has −1 = ei(π+b·2π) and eiα = ei(α0 +c·2π) . This gives the equation
ei(2θ+pi+b·2π) = ei(α0 +c·2π) with solution θ = (α0 − π)/2 + (b + c)π. Thus, for each α0 , θ
can take any value θ = θ0 + dπ, where θ0 = (α0 − π)/2 and d = b + c. Thus, if d is an even
number, θ and θ0 will represent the same angle, while if d is an odd number the angle
represented by θ and θ0 will differ by an angle π. However, only angles corresponding
to a reduced frequency k ≥ 0 make any physical sense for the current approximation of
the aerodynamic model, so if some θa fulfills k > 0 the corresponding angle is the only
physically feasible solution. If an angle θb corresponds to k = 0, then so does the angle π
radians away, and both are physically possible.

Based on these possibilities, solutions fulfilling 0 < α0 < π are concluded to be stable,
since they give feasible solutions corresponding to π/2 < θ0 < π, i.e. solutions p = g + ik
with g < 0. Solutions with π ≤ α0 ≤ 2π are concluded to be unstable, since they give
feasible solutions corresponding to 0 ≤ θ0 ≤ π/2, i.e. solutions p = g + ik with g ≥ 0.

2.3 Discretization into r-intervals

The elements of A(u, r) are assumed to be continuous functions aij (u, r) and so the
eigenvalues of the matrix λj (u, r) will be continuous functions of the same variables [9].
However in most practical applications information about aij (u, r) will only be available
at a discrete set of points and, as a consequence, the eigenvalues will also only be known

7
at a discrete set of points. But by assuming that the eigenvalues vary linearly in a neigh-
borhood around each point, approximate eigenvalue paths can be obtained by computing
the eigenvalues at the discrete points and differentiating them with respect to u, r. The
eigenvalue derivatives are defined for simple eigenvalues but may not be available for mul-
tiple eigenvalues [10]. For this reason, the eigenvalues will be assumed to be simple in the
following sections.

In this study, the stability analysis will be performed at a constant speed and altitude.
Thus, the speed will be a constant uf and each element in A will be a function of r only,
i.e. aij (uf , r) = aij (r). In an interval close to the point rm , Rm ={r|rm ≤ r < rm + ∆rm }
a simple eigenvalue λn can be approximated by

λj (r) = λj (rm ) + (∂λj /∂r)|rm (r − rm ). (8)

Since λj (r) in this approximation is a linear function of r it can be investigated if an

eigenvalue fulfilling λj (r) = eiα is possible for some r ∈ Rm . If that is the case, the
corresponding θ can be evaluated to see if this λj (r) corresponds to a stable or unstable
solution of (5).

If no eigenvalue λj (r) = eiα is possible or if there exists a λj (r) = eiα corresponding

to a stable solution of (5), the eigenvalue is concluded to be stable at uf and for all
r ∈ Rm . If there exists a λj (r) = eiα corresponding to an unstable solution of (5), the
eigenvalue is concluded to be unstable at uf for some r ∈ Rm .

Here it may be informative to note that the assumed eigenvalue variation in (8) does
not necessarily imply that the eigenvalues are continuous in between the different Rm
intervals. However, for the stability conditions to hold it is only necessary that the eigen-
values are continuous functions in each Rm . This since, under the assumptions made, the
stability is only evaluated in one Rm interval at the time.

When evaluating the stability in a certain Rm the method says nothing about the sta-
bility in the rest of R. However, by dividing R into a number of connected intervals Rm
and evaluating the stability in each Rm it is possible to investigate the stability in all
of R. This since the different Rm together cover all of R and an instability thus cannot
be present in R without being present in some Rm . So by evaluating the stability in all
different Rm one also evaluates the stability in R.

3 Stability of the Perturbed System

The system (1) is nominally stable in a certain flight state if all solutions p have a negative
real part. If introducing uncertainties to (1), the perturbed system is stable if all possible
solutions still fulfill the stability criteria for the uncertainties under consideration. The
critical flutter speed of the perturbed system is then the lowest speed for which the
perturbed system has an unstable solution.

3.1 Aerodynamic uncertainties

In this part it will be explained how to formulate the aerodynamic uncertainties as varia-
tions in pressure coefficients at a number of aerodynamic nodes. This method is suitable

8
for aerodynamic matrices obtained from Doublet-Lattice Theory [7]. It may be useful
to note that the analysis procedure which will be employed for these uncertainties is not
restricted to this exact uncertainty formulation, but can be applied to any uncertainty
formulation that can be written as a linear combination of uncertainty matrices.

In order to be able to formulate the aerodynamic uncertainties as variations in pressure

coefficients it is assumed that the aerodynamic matrix can be divided into left and right
partitions QL , QR that relates displacements and forces through pressure coefficients in
different boxes [7], [11]. Letting C p0 =diag(cp0u ), where cp0u denotes the pressure coeffi-
cient in box u, the aerodynamic matrix can be written as Q(r) = QL C p0 QR . Introducing
uncertainties in the pressure coefficients as cpu = cp0u (1+xau eiβu ), where xau ∈ R denotes
the magnitude of uncertainty u and eiβu is a complex number fulfilling |eiβu | = 1, and
letting dC p (r, xa ) =diag(cp0u xau eiβu ) the aerodynamic uncertainty matrix can be written
as
dQ(r, xa ) = QL dC p QR . (9)
It can be seen that dC p can be represented by a matrix sum
X
dC p = cp0u xau eiβu dC u , (10)
u

where D u is a matrix with element (u, u) equal to one and all other elements equal to
zero. Thus, by the distributive law of matrix algebra, (9) can be rewritten as
X
dQ(r, xa ) = wu eiβu QL dC u QR . (11)
u

This gives the A-matrix

"   #
1 b 2 1 2 X
A(r, xa ) = 2 (Ω − ρuf (Q(r) + xau eiβu QL dC u QR ) (12)
r uf 2 u

and so the uncertain flutter equation can be written as

[ei2θ I + A(r, xa )]v = 0. (13)

Comparing (13) to (7) it can be seen that (7) is a special case of (13) obtained by letting
xa = 0. Thus, based on the same reasoning as for (7), it can be seen that the system is
stable if and only if A(r, xa ) has no eigenvalue on the lower half of the unit circle.

In order to be able to evaluate if the stability criteria is fulfilled for (13) it is assumed
that each λj varies linearly with xa . This can be seen as a reasonable assumption for
small xau and the variation can be found by computation of eigenvalue derivatives. These
derivatives are available for simple eigenvalues and can be found from
∂λj ∂A(q, r, xa )
= v ∗lj v rj , (14)
∂xau ∂xau

where v lj and v rj are the left and right eigenvectors to eigenvalue j [10]. Using (9) one
obtains
∂A(r, xa )
= eiβu v ∗lj QL dD u QR v rj . (15)
∂xau

9
It can be seen that, since eiβu represents an arbitrary complex number with magnitude
equal to one, each of these terms describes a circle in the complex plane. Thus, the sum
of the different eigenvalues differentials can be bounded within a distance daqm from the
nominal eigenvalue path.

Thus, if the distance between λj (r, xa ) and the lower half of the unit circle is larger
than daqm , the aerodynamic uncertainties cannot destabilize the system in that Rm . If
on the other hand the distance is smaller than, or equal to, daqm , the uncertainties could
destabilize the system.

3.2 Mass and stiffness uncertainties

The eigenvalue method described above exposes the property Z T M Z = I when analyz-
ing the stability of the system. To be able to use the same eigenvalue approach when mass
uncertainties are present, the uncertainties have to be expressed in such a way that this
property is not lost. One way to achieve this is to express both mass and stiffness uncer-
tainties as variations in structural eigenvalues and eigenvectors. The required eigenvalue
and eigenvector differentials are available for simple eigenvalues but may not be available
for multiple eigenvalues [12]. In the following sections, the eigenvalues will therefore be
assumed to be simple. Using this approach, any uncertainty formulation that can be
written as a linear combination of uncertainty matrices can be allowed for both mass and
stiffness uncertainties.

Denoting the nominal structural eigenvalue problem

[−ω 2 M + K]z = 0 (16)

P ∂M
one has Z T M Z = I and Z T KZ = Ω. Letting M xms = M + ∂xm xmv and K xms =
v
v
P ∂K
K+ ∂xs xsw , where xmv and xsw denote the mass and stiffness uncertainties, one
w
w
obtains the uncertain structural eigenvalue problem

[−ω 2 M xms + K xms ]z xms = 0. (17)

The maximum magnitudes of xmv and xsw are denoted by x̂mv and x̂sw respectively. If
they are small, it P
can be assumed P that the eigenvalues and eigenvectors
P ∂z varyPaccording
∂ω 2 ∂ω 2 ∂z
to ωx2ms = ω 2 + ∂xm x mv + ∂xs xsw and z xms = z + ∂xm x mv + ∂xs xsw
v w v w
v w v v
T
respectively. Denoting the corresponding modal basis Z xms one has Z xms M xms Z xms = I
and Z Txms K xms Z xms = Ωxms , which gives the uncertain flutter equation
"  2 #
i2θ 1 b 1 2 T
e I+ 2 (Ωxms − ρuf Z xms Q(r)Z xms ) v = 0. (18)
r uf 2

The eigenvalue variations caused by these uncertainties can be found from (14) and each
eigenvalue variation can be bounded within an uncertainty distance dms (r, xm , xs ). The
resulting area, within which an eigenvalue λj (r, xm , xs ) is allowed to vary, could then
be assumed to include all points within the distance dms (r, xm , xs ) from the nominal
eigenvalue λj (r, 0). However, since the eigenvalue variations in this case have certain
directions, and cannot reach all points within the uncertainty distance, this may induce

10
a certain amount of conservatism in the results. This can be avoided by bounding the
uncertainties within their convex hull instead, which can be found by use of Minkowski
sums [13]. The perturbed eigenvalue is then considered to be stable if none of the convex
hull intersects with the lower half of the unit circle. If on the other hand some point
intersects with the lower half of the unit circle, it is concluded that some uncertainty
could destabilize the eigenvalue.

3.3 Combined uncertainties

Combining the uncertainties from (13) and (18), one obtains the uncertain flutter equation
"  2 #
i2θ 1 b 1 2 T
e I+ 2 (Ωxms − ρuf Z xms Q(r, xa )Z xms ) v = 0. (19)
r uf 2

For an analysis of how the combination of aerodynamic, mass and stiffness uncertainties
affect the stability of the system, one has to analyze if the uncertainty area caused by
the combined uncertainties could have an intersection with the lower half of the unit circle.

The perhaps simplest way is to use the circular bound for both aerodynamic and mass and
stiffness uncertainties. This would give rise to a combined uncertainty distance, within
which the eigenvalues were allowed to vary. It would have the advantage of being simple,
but could give unnecessarily conservative results as discussed in Section 3.2.

Another possibility is to bound the variations within the convex hull of the combination
of the line segments from the mass and stiffness uncertainties and the uncertainty circles
from the aerodynamic uncertainties. This method may be a bit more computationally
demanding, but has the advantage that it avoids unnecessarily conservative results.

4 Flutter analysis of a delta wing model

To evaluate the practical usefulness of the perturbation analysis method a case study
was performed on a delta wing model. The case study was aimed at analyzing if the
flutter speed bounds predicted by the perturbation analysis method, for parameter vari-
ations within certain bounds, were similar to the worst case flutter speeds found when
performing a number of nominal analyzes for different combinations of parameter vari-
ations, within the same bounds. For this analysis it was considered important to have
well defined parameter variations which made physical sense. It was therefore decided to
focus on mass uncertainties caused by fuel level variations and aerodynamic uncertainties
caused by uncertain aerodynamics around the leading and trailing edges of the wing.

A relatively simple delta wing model, see Figure 1, was used. The model had a semi
span of 0.86 m and a thickness of 4 mm. It was made of a glass fiber composite material
with an elastic modulus of 28.5 GPa, a shear modulus of 6.0 GPa, a Poisson ratio of
0.1606 and a density of 2036 kg/m3 . The mass and stiffness matrices needed for the nu-
merical analysis were obtained from a finite element model and the aerodynamic matrix
was found using Doublet-Lattice aerodynamics [8].

The aerodynamic model was only considered to be accurate for reduced frequencies k

11
up to 1.5 and since the equations needed for the analysis are undefined in the origin only
r values between 0.1 and 1.5 were considered in this study. If the wing model has flutter
instabilities for reduced frequencies outside this interval, they will thus not be found in
the analysis. But since experimental flutter testing of wing models similar to the one
in this study have shown to suffer from flutter instabilities at reduced frequencies close
to the center of this interval, it is credible that also this wing model will have reduced
frequencies within the interval and that the flutter speeds found in this study actually are
the critical flutter speeds for the wing model in question.

The numerical model predicted the wing to have a flutter speed of 48.7 m/s and a flut-
ter frequency of 7.4 Hz. An experimental flutter test of a physical model for which the
geometric and material properties were assumed to be the same gave a flutter speed of
48.5 m/s and a flutter frequency of 8.4 Hz. The numerical model thus predicted a flutter
speed similar to the one of the physical model but a slightly lower flutter frequency. For a
flutter analysis of a real aircraft the numerical model would possibly need to be tuned to
obtain a better prediction of the flutter frequency of the physical model. But in this study
the main goal was not to analyze the exact flutter behavior of the physical wing model
in question, but to investigate if the perturbation analysis method was able to predict
the speed variations caused by mass variations for an arbitrary wing model. It was thus
deemed sufficient that the numerical model gave results which made physical sense and
for that purpose the match between the numerical and experimental analysis was deemed
good enough.

The mass uncertainties were chosen to be variations in the fuel levels in nine fuel tanks
at different span and chord wise positions of the wing, see Figure 2. The fuel mass was
modeled as an increased density of the material, from 2036 kg/m3 to 4000 kg/m3 , in
the corresponding parts of the wing and the flutter speed variation for density variations
within 10% of 4000 kg/m3 was analyzed. Nominal analyzes for a number of combinations
of density variations within different bounds were made. Worst case flutter speeds for the
different uncertainty bounds were found as the lowest nominal flutter speed computed
for some parameter combination within the bounds. These results were compared to the
worst case flutter speeds predicted when using the same bounds for the density variation
as input to the perturbation analysis method. The predicted worst case flutter speeds for
the different uncertainty bounds can be seen in Figure 4.

The aerodynamic uncertainties were modeled as variations in the pressure coefficients

in six patches along the leading and trailing edges of the wing, see Figure 3. Nomi-
nal analyzes for a number of combinations of pressure variations within different bounds
were made. Worst case flutter speeds for the different uncertainty bounds were found as
the lowest nominal flutter speed computed for some parameter combination within the
bounds. These results were compared to the worst case flutter speeds predicted when
using the same bounds for the pressure variation as input to the perturbation analysis
method. The predicted worst case flutter speeds for the different uncertainty bounds can
be seen in Figure 5.

An analysis of the flutter speed variations caused by the combination of mass and aero-
dynamic uncertainties was then made. Also this time, a number of nominal analyzes
were made for various parameter combinations within certain bounds and the worst case

12
 Figure 3: Numbering of
Figure 1: Wing model Figure 2: Numbering of
the patches used for the
used for the nominal flut- the fuel tanks used for the
aerodynamic uncertain-
ter analysis. mass uncertainties.
ties.

flutter speeds found from these analyzes were compared to the worst case flutter speeds
found when using the same bounds in the perturbation analysis method. The results can
be seen in Figure 6.

As can be seen in Figures 4,5, the worst case flutter speeds predicted by the pertur-
bation analysis method show close resemblance to the worst case flutter speeds found
from the nominal tests of different combinations of mass uncertainties and aerodynamic
uncertainties. For the combined uncertainty models the flutter speeds in Figure 6 show
that the predictions from the perturbation analysis method are slightly conservative as
compared to the worst case flutter speeds found from the nominal analyzes.This could
indicate that the perturbation analysis method in this case over predicted the decrease
in flutter speed caused by the combined uncertainties. However, it could also mean that
there are worse combinations of parameter variations than the ones investigated in the
nominal analyzes and that the speed bounds predicted by the perturbation method are
more accurate than the ones found from the nominal analyzes.

Another aspect to note about the results is that the worst nominal flutter speed found
for a 10 % variation in the combination of mass and aerodynamic uncertainties is 22 %
below the flutter speed for the nominal model without any uncertain parameters. One
could thus perhaps question if this large uncertainties should be analyzed by a perturba-
tion analysis method, or if one should try to improve the nominal model to reduce the
magnitude of the uncertainties instead. If only analyzing the results for flutter speeds
above 42.7 m/s, which corresponds to a 10% decrease from the nominal flutter speed,
it can be noted that the speeds predicted by the perturbation analysis and the nominal
analyzes are fairly close.

Since there are infinitely many possible combinations of parameter variations within the
bounds, it is possible that there could exist some worse combination of parameter vari-
ations which was not found when testing different combinations and that the worst case
flutter speed that could be caused by the parameter variations actually is below the one
predicted by the perturbation analysis method. However, since the parameter variations
tested in the nominal analyzes were chosen as combinations which were expected to give
a noticeable decrease in flutter speed, it is credible that a combination close to the worst

13
 Figure 4: Worst case flutter speeds for mass uncertainties.

Figure 5: Worst case flutter speeds for aerodynamic uncertainties.

Figure 6: Worst case flutter speeds for combination of mass and aerodynamic
uncertainties.

14
combination was found. This would indicate that the predicted worst case flutter speeds
are fairly close to the actual ones and that the perturbation analysis method was successful
in estimating these speed boundaries.

5 Discussion
As shown by the test case study in the previous section, the perturbation analysis method
seems to be able to predict fairly accurate bounds for the speed variations caused by var-
ious parameter variations. If the assumptions about linear variation of eigenvalues and
eigenvectors hold, this should also be the case. For most dynamical systems these as-
sumptions are reasonable if the parameter variations causing them are small enough. It is
thus reasonable to expect the speed variations predicted by the method to be accurate for
small variations in mass, stiffness and aerodynamic properties. However, lots of analysis
remains to be done to be able to tell how small the variations need to be to be considered
as small enough.

Another assumption that needs discussing is the one about continuously varying eigenval-
ues and eigenvectors. As mentioned in Section 3.1, these derivatives may be undefined for
multiple eigenvalues and for this reason the eigenvalues are, in this study, assumed to be
simple. If multiple eigenvalues appear, it is difficult to know how to interpret the results
obtained from the perturbation analysis. So far, undefined derivatives have not shown to
be a problem but it may be recommendable to keep in mind that problems could appear.
It could perhaps also be useful to investigate how to handle the problems in case of them
appearing.

The fact that the governing equations are undefined in the origin may also need dis-
cussing. For aircraft with a high wing stiffness flutter instabilities will likely appear at
fairly high values of the reduced frequency and it may thus not be a problem that the
analysis cannot be performed close to the origin. For aircraft with a low wing stiffness
however, flutter could appear at fairly low reduced frequencies. Since it may be difficult
to analyze the system’s behavior at low reduced frequencies with the above described
method, the method may not be suitable for analysis of aircraft for which one suspects
that flutter instabilities could appear at very low reduced frequencies.

Further, the perturbation method builds the analysis on the assumption that, close to
the imaginary axis, the reduced frequency is approximately equal to the polar radius. A
question that naturally arises is for what polar angles this is a reasonable assumption.
To answer the question one needs to analyze for what angles the aerodynamic model is
able to accurately represent the aerodynamic forces. No such analysis has been made in
this study but, based on arguments about small angles, it is assumed to be a reasonable
assumption for a few degrees variation about the imaginary axis. This means that the
analysis method cannot be expected to give accurate flutter predictions outside this in-
terval and that the analysis thus is restricted to predictions within this area. But the
commonly used approximation of letting the aerodynamic matrix be a function of the
reduced frequency, and thereby neglecting the effect of the damping of the motion, is also
restricted to an area close to the imaginary axis. Thus, the effect of approximating the
reduced frequency with the polar radius will hopefully not have a severe effect on the
validity of the aerodynamic model.

15
6 Conclusions
The study has shown that the suggested method for perturbation analysis seems to be
able to predict accurate bounds for flutter speed variations for small variations in mass
and aerodynamic properties. The linear assumptions on which the method builds can
likely be seen as reasonable for small parameter variations and the method can thus be
expected to give accurate flutter predictions for small variations in modeling parameters.
However, a lot of analysis remains to be done to be able to tell how small the variations
need to be for them to be considered as small enough. Since important equations and
derivatives are undefined in certain points, problems could appear if the analysis needs
to be performed close to those points. Hopefully however, such cases will not be very
common in practical applications.

16
References
[1] E.H. Dowell et al. A Modern Course in Aeroelasticity. Kluver, Dordrecht, The
Nethrlands, 2005.

[2] P. Bäck and U. Ringertz. Convergence of Methods for Nonlinear Eigenvalue Prob-
lems. AIAA Journal, 35(6):1084–1087, 1997.

[3] D. Eller. Flutter Equation as a Piecewise Quadratic Eigenvalue Problem. Journal of

Aircraft, 46(3):1068–1070, 2009.

[4] R. Lind and M. Brenner. Robust Aeroservoelastic Stability Analysis. Springer, Lon-
don, Great Britain, 1999.

[5] R. Lind. Match-Point Solutions for Robust Flutter Analysis. Journal of Aircraft,
39(1):91–99, 2002.

[6] D. Borglund. The µ − k Method for Robust Flutter Solutions. Journal of Aircraft,
41(5):1209–1216, 2004.

[7] D. Borglund. Robust Eigenvalue Analysis Using the Structured Singular Value: The
µ − p Flutter Method. AIAA Journal, 46(11):2806–2813, 2008.

[8] D. Borglund and D. Eller. Aeroelasticity of Slender-Wing Structures in Low-Speed

Airflow. KTH Aeronautical and Vehicle Engineering, Stockholm, Sweden, 2013.

[9] M. Zedek. Continuity and Location of Zeros of Linear Combinations of Polynomials.

Technical report, University of Maryland, Maryland, juni 1964.

[10] J. R. Magnus. On Differentiating Eigenvalues and Eigenvectors. Technical report,

London School of Economics, London, 1985.

[11] W. P. Rodden and E. H. Johnson. MSC/NASTRAN Aeroelastic Analysis Users

Guide - Version 68. The MacNeal-Schwendler Corp., Los Angeles, California, 1994.

[12] U. Ringertz. Structural Optimization with Aeroelasticity Constraints. Technical

report, Royal Institute of Technology, Stockholm, 1996.

[13] C. Weibel. Minkowski Sums of Polytopes: Combinatorics and Computation. Tech-

nical report, École Polytechnique Fédérale de Lausanne, Suisse, 2007.

17
 TRITA-MAT-E 2014:19
ISRN-KTH/MAT/E—14/19-SE

www.kth.se