You are on page 1of 41

Ref.

Ares(2016)5714898 - 03/10/2016

Deliverable 2.2
“Robust analysis techniques for flexible aircraft and flutter”
Andrea Iannelli, Nandor Terkovics, Andres Marcos (UOB)
Martin Leitner, Thiemo Kier, Gertjan Looye (DLR)
Roeland de Breuker (TUD)
Due date: 03/07/2016

GA number: 636307
Project acronym: FLEXOP
Project title: FLUTTER FREE FLIGHT ENVELOPE EXPANSION FOR
ECONOMICAL PERFORMANCE IMPROVEMENT

Funding Scheme: H2020 MG-1.1-2014


Latest version of Annex I: 2.1 released on 20/11/2015
Start date of project: 01/06/2015 Duration: 42 Months

Lead Beneficiary for this deliverable: UOB


Last modified: 01/09/2016 Status: Closed
Due date: 01/07/2016

Project co-ordinator name, title and organisation: Bálint Vanek, SZTAKI


Tel: +36 1 279 6113
Fax: +36 1 466 7483
E-mail: vanek@sztaki.mta.hu
Project website address: www.flexop.eu

Dissemination Level
PU Public X
CO Confidential, only for members of the consortium (including the Commission Services)
Glossary
AIC Aerodynamics Influence Coefficient
HF High Fidelity
LF Low Fidelity
FD Frequency Domain
LFT Linear Fractional Transformation
LFR Linear Fractional Representations
LPV Linear Parameter Varying
LTI Linear Time Invariant
LTV Linear Time Varying
LUT Look-Up Tables
NL Nonlinear
DLM Doublet Lattice Method
RFA Rational Function Approximation
TRL Technology Readiness Level
MAC Modal Assurance Criteria
ASE Aeroservoelasticity
CAE Computational Aeroelasticity
RS Robust Stability
RP Robust Performance

FLEXOP D2.2 deliverable v01 y2016m07d01 2


Table of contents
1 Introduction 6
1 Executive Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2 Classical Offline techniques 8


1 State-space eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.1 Technique summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3 Theoretical description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.4 Continuous time aerodynamics aeroelastic state space models . . . . . . . . . . . 11
1.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2 p -k method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1 Technique summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Theoretical description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3 G AAM method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.1 Technique summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.3 Theoretical description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

3 Modern Offline techniques 19


1 µ analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.1 Technique summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.2 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.3 Theoretical description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2 Bifurcation Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.1 Technique summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2 Theoretical description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.3 Continuation Methods and Bifurcation Analysis linked to control law design . . . . 28
2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

4 On-board techniques 30
1 Classical damping trend tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2 Flutterometer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

FLEXOP D2.2 deliverable v01 y2016m07d01 3


2.1 Technique summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2 Theoretical description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

5 Conclusion 36

6 Bibliography 37

FLEXOP D2.2 deliverable v01 y2016m07d01 4


List of Figures
1.1 WP2 workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.1 Example wing and wake discretization using vortex ring elements. The thick solid lines
indicate the wing outline, the thin solid lines indicate the panel distribution, the dashed
lines indicate the vortex ring elements, and the thick gray line indicates the structural beam. 12
2.2 Schematic illustration of the transfer of the aerodynamic forces, indicated with the open
circles, to the structural nodes, indicated with the grey circles. The thick solid lines indi-
cate the wing outline, the dashed lines indicate the vortex ring elements, the black circles
indicate the vortex line vertices, and the thick gray line indicates the structural beam. . . . 12

3.1 Upper Linear Fractional Transformation (LFT) . . . . . . . . . . . . . . . . . . . . . . . . . 20


3.2 LFT property: perturbations at component level become structured at system level . . . . 21
3.3 LFT of an uncertain state-space model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.4 The plot of f (x ) for different values of α. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.5 Saddle-node bifurcation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.6 (a) Super- and (b) subcritical Hopf bifurcations (image taken from [1]). . . . . . . . . . . . 27

4.1 LFT representation of the flexible aircraft model . . . . . . . . . . . . . . . . . . . . . . . . 31


4.2 Dynamic pressure over time (horizontal sections correspond to time spent at each test
point) [2] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.3 The amount of uncertainty in damping ratio over the course of the flight test [2] . . . . . . 34
4.4 Predicted flutter pressure values [2] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.5 CPU time over the course of the flight test [2] . . . . . . . . . . . . . . . . . . . . . . . . . 35

FLEXOP D2.2 deliverable v01 y2016m07d01 5


Chapter 1

Introduction

1 Executive Summary
This deliverable is the second technical report of FLEXOP Work Package WP2 “Flutter Prediction and
Control Design Methods”. This work package considers flutter prediction, active suppression and the
related mathematical modelling task required to link WP2 with WP1 “Aeroelastic Tailoring Methods for
Wing Design”. As shown in Figure 1.1 (right) there are three main tasks within WP2 (i.e. modelling,
analysis and synthesis) which depend on the availability of an integrative model (left of Figure 1.1) that
reconciles the rigid-body dynamics, the fluid mechanics and the structural mechanics. This deliverable
addresses the main goal of Task 2.2 “Flutter Analysis Methodologies” which is to develop techniques for
flutter analyses in order to determine: flight envelope limits, flutter predictions, study flutter suppression.

Figure 1.1: WP2 workflow

There are two main focus for this deliverable:

• Review Offline techniques that allow to perform flutter predictions within the models developed
for analysis Task 2.1 “Control Oriented Flexible Aircraft Modelling” and made suitable for synthesis
applications by means of the ROM and LFT/LPV algorithms. These techniques are thought to be
used during the design stages and final verification phase of the aircraft, thus the classification
Offline. The identified techniques are broken down into classical versus modern. Belonging to

6
the former group can be listed state-space eigenvalues analysis (e.g. p method) and frequency
domain (FD) based algorithms (e.g. k , g and p -k methods). Modern approaches (developed in
the last two decades) include Computational AeroElasticity (CAE), boundary identification, robust
linear techniques (µ analysis) and nonlinear (NL) bifurcation theory.
• Review On-board techniques These techniques serve as basis for the hardware-in-the-loop
simulation platform that will be developed in WP3. These methods are tailored to allow the es-
timation of flutter onset during flight tests making use of flight data. Same classification as for
the Offline techniques will be employed. Classical techniques are all based on the concept to
measure and track damping values of the critical modes until a damping trend can be reliably
extrapolated to the flutter condition of zero damping. The main issue with these methods is that
they are based on curve fitting and thus inherently rely on extrapolation methods and underly-
ing hypothesis (for example linear trend of damping with the flight conditions). Modern methods
(flutterometer, wavelets, Volterra kernels) detour from the traditional damping estimates and move
towards more robust and involved predictions of flutter.

The layout of the document is as follows. Chapter 2 describes traditional algorithms for offline flutter
predictions, whereas in Chapter 3 modern techniques are assessed. Chapter 4 illustrate the on-board
flutter prediction task dealt respectively with traditional and more advanced, recent techniques.
For each technique, it is presented: a brief description summary; a state of art of its application to
aerospace systems; a theoretical description of the technique; an example of application of the tech-
nique to simple study case; a final section on advantages and disadvantages, including assessment of
its technology readiness level (TRL).

FLEXOP D2.2 deliverable v01 y2016m07d01 7


Chapter 2

Classical Offline techniques

Techniques to predict flutter instability have been deeply investigated in the second half of last century,
leading to a well-established state-of-practice for what concerns classic offline analyses. Cornerstone
works can be found in [3–6]. In this chapter an overview of the most well-known methods will be
provided. Here the common starting point is discussed.
The equation describing the aeroelastic plant can be written in compact form in frequency-domain as
follows (s is the Laplace variable)
 
Ms s 2 + Cs s + Ks X = La
     
(2.1)

Depending on the modeling process adopted to formulate the equilibrium, precise meaning can be
attributed to the operators. X is usually the vector of degrees of freedom of the system (for example,
its modal coordinates), Ms , Cs and Ks are respectively the structural mass, damping and stiffness
matrices, while La is the vector of aerodynamic loads.
Eq.2.1 is a rather general expression for the plant since it allows the adoption of unsteady aerodynamic
theories in the model providing an expression of La in the Laplace s domain framework as:
 
La (s ) = q A(s̄ ) X(s ) (2.2)

A(s̄ ) is known as the Aerodynamic Influence Coefficient (AIC) matrix which is a function of the dimen-
sionless Laplace variable s̄ = sb
V .
This is a common starting point for the classic flutter analysis techniques, that is, flutter analysis of flight
vehicles is based on the stability boundaries of the frequency-domain aeroelastic equation of motion in
modal coordinates.

1 State-space eigenvalues

1.1 Technique summary


A very well-known approach to study the stability of a system is to describe it in state-space and anal-
yse the spectrum of the state-matrix.
The flow unsteadiness leads often to formulate the aeroelastic problem in frequency-domain, as tes-
tified by Eqs.2.1-2.2. As known, it is always possible to switch between a frequency-domain and a

8
state-space formulation, if the former is expressed with a rational dependence on the Laplace variable
s . It is not the case in Eq.2.2, since the unsteady aerodynamic solver (as for example the Doublet
Lattice Method [7] or the harmonic gradient method [8]) doesn’t provide an analytical expression of the
operator A as a function of s̄ , but only discrete values. Thus, a proper (i.e. rational) approximation of A
is required.
Due to its importance in aeroservoelastic analyses, the rational approximation of aerodynamics opera-
tors has been deeply studied in the last decades, see [9] for a comprehensive survey. In addition, this
topic was covered in the technical report of Deliverable 2.1, thus it is not thoroughly tackled here. The
foundations of this technique are however discussed in the dedicated section.

1.2 Applications
This technique is often employed in aeroservoelastic (ASE) applications, i.e. when the study of the
aeroelastic system is coupled with servo control dynamics. This is motivated by the adoption of well
consolidated control algorithms which typically require the plant to be described in LTI state-space form.
In the last decade this method has been employed in aircraft applications to tackle diverse tasks: predict
aeroservoelastic flutter and flight mechanics frequency responses [10]; study the interaction between
aircraft structural and control design schemes [11]; flutter suppression [12]; gust-load alleviation [13].
The possibility to study the flutter problem in this framework is also an embedded capability of well-
known toolboxes in the ASE community: STARS (from NASA Dryden Flight Research Center) [14] and
ZAEROS [15] (from ZONA Technology).

1.3 Theoretical description


Generally the aerodynamic approximations consist of a two-part expression based on quasi-steady
(QS) and lag physical contributions to the aerodynamic loads:
 
A(s̄ ) ≈ ΓQS + Γlag (2.3)

The formulation for ΓQS is usually common among the different approaches:

ΓQS = A2 s̄ 2 + A1 s̄ + A0
     
(2.4)

where A2 , A1 and A0 express the contribution to the loads of the instantaneous values of respectively
acceleration, speed and displacement of the degrees of freedom. In addition to this, the memory effect
of the flow, commonly referred to as time lag and resulting in phase shift and magnitude change of the
loads with respect to the quasi-steady case, has to be included. The way this task is addressed usually
differentiates the approximation algorithms. The expression of Γlag for two common approaches in the
aeroelastic community is here summarized.
Roger proposed [16] that Γlag could be approximated as:
N
X s̄  
Γlag −Rog er = AL (2.5)
s̄ + γL−2
L=3

The partial fractions inside the summation are the so-called lag terms and they basically represent
high-pass filters with the aerodynamic roots γi as cross-over frequencies. The resulting state-space
equation includes augmented states X̂ai representing the aerodynamic lags, which are equal to the
number of roots multiplied by the number of degrees of freedom.
Karpel proposed [17] a slightly different algorithm which tries to minimize the number of augmented

FLEXOP D2.2 deliverable v01 y2016m07d01 9


states required to attain a certain degree accuracy. As a result, the lag part of the approximation is now
formulated as:  1
... 0

s̄ +γ1
  . .. ..   0 
Γlag −MS = D0  .. . .  E s̄ (2.6)
1
0 ... s̄ +γN −2

In literature [18, 19] it is reported that a number of aerodynamic states ranging from 4 to 8 generally
leads to good approximations, depending of course on the approximation method adopted.
With these definitions, it is possible to recast Eq.2.1 in state-space form. The expression for the partic-
ular case of Roger’s approximation is here reported:
  
Ẋ(t )
 
0 I 0 ... 0 X(t )
 
 Ẍ(t )    −1    −1    −1    −1  
 − M K − M  B q M  A 3 ... q M AN    Ẋ(t ) 
  
 ˙ V
 X̂a3 (t )  =  0 I − b γ1 I ... 0   X̂a3 (t ) 
 
(2.7)
  
 .  
 .   .
. .
. .
. ..   .
.

 .  . . .
. .   
˙ V
 
X̂aN (t ) 0 I 0 − γ
b N −2 I X̂ aN (t )

     
where the aeroelastic inertial, damping and stiffness matrices M , B and K have been defined as:
    1
M = Ms − ρ∞ b 2 A2
 
2
    1  
B = Cs − ρ∞ bV A1 (2.8)
2
    1
K = Ks − ρ∞ V 2 A0
 
2

A generally valid short-hand state matrix expression can be obtained, whatever the rational approxima-
tion employed:
˙
" #   
X̂ E χE E χE A X̂E
˙ = (2.9)
X̂ A
χAE χAA X̂A

X̂E and X̂A are respectively the vector of elastic and aerodynamic states (classical notation for states
X has been dropped since in here is adopted to define the elastic DOFs vector). The resulting state
matrix has been partitioned into: χE E quasi-steady aeroelastic matrix; χE A coupling term of lag terms
on quasi-steady equilibrium; χAE coupling term of structural states on lag terms dynamics; χAA pure
lag terms dynamics matrix.
Eq.2.9 thus represents the starting point for this technique. The eigenvalues of the state matrix can
be evaluated for various values of the flow velocity to trace the damping and frequency of the various
modes in the system until flutter is found. This methodology has similarities with what is known in
literature as the p method [20], although in here the mathematical framework is entirely the state-space
(without converting through the Laplace transform).
A remark is here highlighted: in flutter analysis it is often desired to associate the eigenvalue with its
physical mode in order to get insight into the dynamics underlying flutter. In fact, the understanding
of the physical mechanism leading to instability could help designers to develop solutions which may
postpone it. This means that as the speed is varied, one should be capable to read the found eigen-
values as associated with their eigenmodes. This is not a straightforward task if one considers this
kind of problem, where the state-matrix is updated at each speed and the presence of aerodynamic
augmented states complicate the system spectra. Moreover, often a merging of the frequencies is ob-
served close to the instability, and simply sorting out the eigenvalues by their magnitude could lead to
wrong conclusions.

FLEXOP D2.2 deliverable v01 y2016m07d01 10


The Modal Assurance Criteria (MAC) is one of the possible algorithms that addresses the mode track-
ing problem [21]. It is a scalar value which quantifies the linearity between two mode shapes. Cursory,
a reference mode ψr and a tested mode ψt are compared in order to assess how similar (in a linear
combination sense) they are. The mathematical definition of MAC follows:

|ψr∗T ψt |2
MAC (ψr , ψt ) = (2.10)
ψr∗T ψr ψt∗T ψt
 

The output of this comparison is thus a scalar between 0 and 1, i.e. two identical modal vectors yield a
unity value, two orthogonal modal vectors yield a zero value and in-between there is a certain degree
of similarity between the reference and the tested modes.
The flutter investigation should therefore start at a low speed such that the modes are distinct and
well detectable (e.g. from a comparison with the free-vibration analysis of the system) and thus the
modes to be tracked ψ j (with j =1,2,3) can be defined and numerated with a preset order. At each
speed, eigenvalues are calculated and the meaningful eigenvalues are ordered according to the criteria
in Eq.2.10; the modal reference basis ψrj is updated at each speed and so the dynamics in terms of
mode-eigenvalue pair is tracked.
In the following section, a different way to formulate the aerodynamics loads in the time-domain frame-
work is described; this can be considered as an alternative step to perform rational approximations
when the technique previously described is employed to perform flutter analysis.

1.4 Continuous time aerodynamics aeroelastic state space models


The classical method of converting unsteady aerodynamics from frequency domain to state space time
domain is based on an approximation of the frequency-dependent aerodynamic forces, as described in
previous sections. To avoid errors that the approximation might introduce, the unsteady aerodynamic
forces have to be written directly in time domain. A popular method is the discrete unsteady vortex
lattice method. The downside of using this method is the fact that the discretisation of the governing
aerodynamic equations is such that the time step depends on the frequency spectrum of the response.
Therefore, the number of unknowns can increase rapidly and can lead to computationally inefficient
models. A solution to this is writing the aerodynamic model in a continuous time fashion. This method
was developed recently at Delft University of Technology.
The unsteady aerodynamic model that is coupled to the structural model is based on potential flow
theory implemented by means of the unsteady vortex lattice method. A detailed explanation of the
unsteady aerodynamic model used in the present approach is given in [22]. A brief explanation will be
included here for completeness. The wing is modelled as a thin wing and represented by its deformed
camber surface. The wake is assumed to be flat and leaves the trailing edge in the free-stream direction.
Since, in case of unsteady aerodynamics, the wake vorticity is no longer constant, both the wing surface
and the wake surface are discretised, resulting in a discretisation as illustrated in Fig. 2.1.
Under the assumption of small perturbations with respect to the steady state and by applying the flow-
tangency condition, the Kutta condition and Helmholtz theoreom for the transport of vorticity in the
wake, the unsteady aerodynamic model can be written as a continuous-time state-space model:

      
Γ̇w K1 K2 Γw 0
= + α̇ (2.11)
α̇ 0 0 α I
 
  Γw
Fa = L1 L2 + L3 α̇ (2.12)
α

FLEXOP D2.2 deliverable v01 y2016m07d01 11


V∞ z
y
x

Figure 2.1: Example wing and wake discretization using vortex ring elements. The thick solid lines
indicate the wing outline, the thin solid lines indicate the panel distribution, the dashed lines indicate the
vortex ring elements, and the thick gray line indicates the structural beam.

where Γw are the vortex strengths of the aerodynamic panels of the free wake panels and α are the
perturbation angles of attack of the aerodynamic panels on the wing surface. Once the vortex strength
distribution in the free wake is resolved, the vortex strength distribution on the wing surface and in the
trailing edge wake panels can be determined by inserting these in the equations for the flow tangency
condition and the Kutta condition. The input of the state-space model is given by the time derivative of
the perturbation angle of attack and is defined separately for each aerodynamic wing panel.
Compressibility has been accounted for by means of the Prandtl-Glauert transformation, which limits the
applicability of the unsteady aerodynamic model to low to moderate compressible reduced frequencies
and no shocks, so care should be taken in applying this model to a combination of high Mach number
and high reduced frequency. However, within these limitations, the present model is still expected
to provide a good approximation of the effect of compressibility on the unsteady aerodynamics. The
effect of transonic aerodynamic effects could be included in future work, for example, by means of a
correction of the unsteady aerodynamic loads by CFD simulations. The main advantage of the present
model, however, is that it allows for efficient discrete gust simulations in time domain.
y

Figure 2.2: Schematic illustration of the transfer of the aerodynamic forces, indicated with the open
circles, to the structural nodes, indicated with the grey circles. The thick solid lines indicate the wing
outline, the dashed lines indicate the vortex ring elements, the black circles indicate the vortex line
vertices, and the thick gray line indicates the structural beam.

In order to couple the structural and aerodynamic models, the aerodynamic forces need to be trans-
ferred to the structural nodes as input to the structural model and the time derivative of the perturba-
tion angle of attack induced by the structural deformations on the aerodynamic panels needs to be
determined as input to the unsteady aerodynamic model. Similar to the steady aerodynamic model,
the aerodynamic forces are first transferred to the structural beam by means of a rigid link, and then
converted into statically equivalent nodal forces, as is illustrated in Fig. 2.2, resulting in the following

FLEXOP D2.2 deliverable v01 y2016m07d01 12


transformation equation:

Fs = TAS Fa (2.13)

The perturbation angle of attack on each aerodynamic panel is composed of four components:

ḣ θ̇ (x − xb )
α = αair + θ − + (2.14)
V∞ V∞

where αair is the perturbance angle of attack induced by the free stream flow, θ is the structural wing
twist, − Vḣ∞ is the perturbance angle of attack introduced by the local plunge motion of the wing, and
θ̇(x −xb )
V∞ is the perturbance angle of attack introduced by the local pitch rate of the wing, where x defines
the location of the aerodynamic panel and xb defines the location of the beam axis. The local structural
deformations are found by a linear interpolation of the nodal structural deformations to the spanwise
aerodynamic stations, resulting in the following transformation equation for each of the structural trans-
lations, rotations, and velocities:

xa = TSA xs (2.15)

where xs contains translations, rotations or velocities at the structural nodes, and xa contains the cor-
responding degrees of freedom at the aerodynamic spanwise stations.
Starting from the aerodynamic state-space system, as defined by Eqs.2.11-2.12, using Eq.2.14, the
aerodynamic state equation can be linked to the different components of the perturbation angle of
attack. Similarly, using Eq.2.13, the structural state-space system can be linked to the aerodynamic
forces. Combining both relations the aeroelastic state equation is obtained as:

ẋ = Ass x + Bss α̇air (2.16)

where the state vector x is defined as:

 T
x= Γw αair ṗ p (2.17)

In order to complete the dynamic aeroelastic state-space system, the aerodynamic forces and mo-
ments and the structural degrees of freedom are selected as outputs, resulting in the following output
equations:

 
F
= Css x + Dss α̇air (2.18)
p

The resulting aeroelastic state space system is thus obtained in a continuous time fashion without any
approximations of the unsteady aerodynamics. The number of states in the system is determined by
the required accuracy of the solution rather than the frequency bandwidth of the response of interest.

1.5 Conclusions
This technique represents a standard approach to deal with flutter analysis when it is instrumental
to pose the problem in the state-space framework. Aeroservoelastic investigations are an example.

FLEXOP D2.2 deliverable v01 y2016m07d01 13


Recent results (belonging to the last decade) cited in the state-of-art section demonstrate how it is still
at the present stage an invaluable method to fulfill these tasks.
The main disadvantages are related to the constraint to analyze the problem within the framework of
Eq.2.9. Since often the aerodynamics for flutter analyses is provided at discrete values of the reduced
frequencies k , this implies that rational approximations of the operators have to be adopted. Although
a vast amount of research has been conducted in this topic, this entails a loss of accuracy in the
predictions and an increase of the size of the problem since the state-matrix is augmented with the
aerodynamic states.

2 p -k method

2.1 Technique summary


The p -k method [6] solves the flutter problem directly in the frequency-domain framework where Eq.2.1
is stated. The AIC matrix is thus allowed to have the more general dependence on the variable s̄ and
can be straightforwardly provided by a dedicated unsteady solver, without intermediate adjustments
(e.g. rational approximations). Indeed this method is typical of those applications that employ unsteady-
potential-flow based aerodynamic solvers (e.g. DLM) providing A(ik ) under the hypothesis of oscillatory
motion of the body, which is common in the flutter community.
The objective is to find the flutter determinant roots s̄ such that nonzero solutions for X exist in Eq.2.1.
The complexity arises since A(s̄ ) is a trascendental function of the frequency of oscillations, known at
discrete values of s̄ and thus iterative solutions have to be sought.
Flutter occurrence is featured by a condition of pure harmonic motion and thus the critical speed is the
lowest one having at least one eigenvalue with γ = 0.

2.2 Applications
The p -k method is a well-established and consolidated technique in the aeroelastic research commu-
nity and aeronautical industry.
The former has employed this technique as a tool to address different issues of practical interest: eval-
uate the role of unsteady aerodynamics in defining flutter constraints for aeroelastic optimization [23];
dynamic aeroelastic tailoring of transport wings [24]. Among these contributions, it’s worth mention-
ing the work [25] which extends the applicability of this technique to closed-loop stability analysis of
aeroeservoelastic plants. This contribution shows that it is possible to tackle problems featuring flutter
and control applications within the frequency-domain framework, i.e. without making use of rational
approximation as requested by the method discussed in Sec.1. A conceptually similar goal was later
on pursued in [26].
This technique is also one of the options (PK) found in the NASTRAN aeroelastic solver [27], where this
method is also called ’British method’ since it was originally proposed (though with some differences
compared to the version here discussed) in [28] by the British Aeronautical Research Council.

FLEXOP D2.2 deliverable v01 y2016m07d01 14


2.3 Theoretical description
The solutions sought by the p -k method are the dimensionless conjugate complex roots s̄ , named p :
b
p = s̄ = γ k ± ik = s
V
(2.19)
1 an+1
2γ = ln( )
π an
where k is the mode reduced frequency and γ is the associated rate of decay of the amplitudes an (i.e.
a measure of its damping).
The starting equation for the application of the standard p -k method [6] is formulated as follows, recalling
that the aerodynamic matrix is only available for oscillatory motion (s̄ =i k):
2
 
2V V   
   E
p 2 Ms + p Cs + Ks − q A(ik ) X = 0 (2.20)
b b

A contradiction in this formulation may be noticed, since structural contributions are not treated as
harmonic (p is not purely imaginary) as instead is done for the aerodynamic term (because this operator
is only available for oscillatory motion). The rationale for this approach is that for sinusoidal motions
with slowly increasing or decreasing amplitude (i.e. small rate of decay), an aerodynamic approximation
based on constant amplitude is sufficient.
In practice, from physical knowledge of the problem (or previous flutter analysis) modes of interest are
selected and determined from vibrational analysis. To complete a frequency-damping-velocity diagram,
at the first selected speed an initial trial for each mode must be input. A possible trial is provided by:

2π fvj b
p0j = 0 + i = ik0j (2.21)
V
where fvj is the natural frequency of the vibration mode j . Of course, the first speed V should be
sufficiently small such that fvj is reasonably close to the aeroelastic frequency f j of the mode at that
speed. A second guess for the root is needed in order to initialize the matrix-determinant iteration
algorithm. A possible choice is:
p1j = −0.01k0j + ik0j (2.22)
Once A(ik0 ) and A(ik1 ) are computed, the corresponding values of the determinant |F(p0 )| and |F(p1 )|
are obtained. The Regula Falsi method gives the iterated value p2 (dependence of all parameters on j
is omitted)
(p1 F0 ) − (p0 F1 )
p2 = (2.23)
F0 − F1
The process is repeated until a specified degree of convergence is attained. From the converged root
pc =δc +ikc , frequency and damping can be computed:
V kc
f =
2π b
(2.24)
<p
2γ = 2
kc
The whole procedure can then be repeated for all the modes j of interest. In general the meaningful
modes in flutter analysis are the first 5-6.
The weak point of this technique lies in the twofold way the solution p is treated, i.e. as pure imaginary
root in aerodynamic calculation and as complex root for the structural term. This is not a critical issue
since in a condition of neutral stability p = ik , hence the flutter speed is correctly predicted. However
refinements of the methods have been later proposed to include aerodynamic contributions due to

FLEXOP D2.2 deliverable v01 y2016m07d01 15


generic (non-harmonic) motions, in order to better track frequencies and more importantly, damping for
subcritical speeds. In [29] Eq.2.20 was modified introducing an aerodynamic damping matrix (last term
on the left hand side):

V2
 
V   
p 2 2 Ms + p Cs + Ks − q A(ik ) − q γ AI XE = 0
    
b b (2.25)
R I
A(ik ) = A + i A

It can be proved however that this is valid only at small k or linearly varying A(ik ). A further refine-
ment which extends its validity to the entire reduced frequency range was proposed in [3], known as
g method. It adds to Eq.2.20 a first-order approximation damping term which is rigorously derived
from the available AIC calculated considering simple harmonic motion, and thus improves the results
accuracy near zero damping. This version is implemented in the aeroelastic code ZAERO [15].

2.4 Conclusions
This section has described contents and applications of the p -k method, which can be legitimately con-
sidered the state-of-practice in many applications concerning flutter calculation. One of the reasons is
that it is tailored to deal with aerodynamics formulations which are trascendental functions of the oscil-
lation frequency - the industry-accepted standard for flutter prediction of aircraft [30].
The major weak point, consisting in less accurate prediction of frequencies and damping for subcrit-
ical speeds, was analysed at the end of last paragraph. Refinements of the original algorithms were
proposed which tackle this issue. These three forms, namely the original one proposed by Hassig [6],
the first correction by Rodden [29] and the further refinement by Chen [3], are commonly referenced in
literature [31] as different versions of the p -k method.
Another critical aspect of the technique concerns its applicability to problems not limited to flutter speed
calculation, as for example aeroservoelastic investigations and control applications. Historically, these
tasks have been addressed recasting the problem in state-space due to modern control design tech-
niques requiring the equations of motion to be defined in a first-order, time-domain form. Recent studies
proved that it is possible to overcome this hurdle and formulate the problem entirely in the framework of
the p -k method (in particular this is demonstrated for the g method). However these last attempts are
still at the research stage and no many applications can yet be found, therefore it is believed that for
ASE applications the preferred approach is still the one in Sec.1.

3 G AAM method

3.1 Technique summary


The Generalized Aeroelastic Analysis Method (G AAM ) [5] solves the basic flutter equations of Eq.2.1
directly without any guidelines or restrictions made with respect to the Laplace domain expansion in s .
The aerodynamic loads La are usually (as in all other practical flutter methods) given by the generalized
aerodynamic influence coefficient matrices A(ik ) for a given reduced frequency k and a set of structural
frequencies, normal modes and rigid body mode shapes. The difference lies in the fact that they are
evaluated for the current Laplace variable which allows the damping to freely move in the whole Laplace
domain. This generalization of the harmonic oscillation assumption has been proven by Edwards in e.g.
[32]. Existing doublet lattice codes can be used with the G AAM , however modifications have to be made
to the kernel to be able to handle complex numbers. The method is iterative but self-converging (similar

FLEXOP D2.2 deliverable v01 y2016m07d01 16


to the most common p -k implementations) in the sense that the roots of the system matrix are directly
fed to the doublet lattice solver to build the next state space system. The G AAM is the most exact flutter
solution method in terms of accuracy in the damping and frequency evolution and only restricted by e.g.
approximations made in calculating the aerodynamic coefficients like in the doublet lattice case.

3.2 Applications
The G AAM is comparatively new and, to the authors knowledge, hasn’t been applied in industrial flutter
determination. Up to this point its application was of pure scientific nature. Edwards first introduced the
method in [5] yet proved that the unsteady linear potential theory equations are valid for arbitrary num-
bers already in e.g. [32]. The published application examples only encompass basic wing structures
such as the well known ha145a (2D/3D airfoils), ha145b (BAH wing) and AGARD 445.6 wing models.

3.3 Theoretical description


The G AAM method searches for the converged eigenvalues of the basic generalized aeroelastic system
of equations,
[Ms s 2 + Bs s + Ks − qA(s̄ , Ma)]{q (s )} = 0 (2.26)
The difference between most other flutter methods exists in terms of how the unsteady aerodynamic
coefficients are directly evaluated for the generalized complex variable ŝ = k + i (γ k ) instead of the har-
monic reduced frequency k followed by some variable-order damping expansion in the Laplace plane.
In practical applications this is usually realized in two ways. The most straightforward approach is to
evaluate the AIC with the help of a doublet lattice code that supports complex reduced frequencies as
input. Often times those codes expect the k variable to be a real number and will fail if it is complex,
hence adaptions have to be made towards complex number support. Another option is to approxi-
mate the complex conjugate coefficients by rational functions, knowing that the approximation quality
strongly influences flutter prediction accuracy (similar to using the state space system approach of the
p method). Adaptations then only have to be made regarding the equation outlined in 2.3. After the
source for unsteady aerodynamics has been established, the eigenvalues λ = s 2 = γω + i ω of equa-
tion 2.26 are determined and used to re-evaluate the aerodynamic coefficients iteratively. This process
of directly matching the aerodynamic with the aeroelastic eigenvalues/frequencies is also found in the
most common p -k method implementations, while the G -Methods (Chen [3],van Zyl [33]) employ a
range-interpolation scheme to find the converged eigenvalues. Concerning the root update, Edwards
suggests including a relaxation factor that allows to tune the algorithm’s performance depending on the
case at hand. The iteration scheme reads,

sn+1 = sn + rf ac (s i − sn ) (2.27)

where s i is the closest eigenvalue to eigenvalue sn of the last iteration. The convergence criterion is
given by the absolute difference of eigenvalues in consecutive iterations, |s i − sn | < conv . Good starting
points s0 for the process are the wind-off structural frequencies from normal mode analysis. Frequency
and damping are extracted from the converged eigenvalues through,

g = 2<(λ)/=(λ)
(2.28)
f = =(λ)

The usual continuity shaping of the damping curves in the vicinity of the real axis g = (2<(λ)cr ef )/(ln(2)V )
can still be applied, yet as in all other methods important details concerning divergence might get lost
in doing so.

FLEXOP D2.2 deliverable v01 y2016m07d01 17


3.4 Conclusions
The G AAM can be considered to be the most advanced of the classic flutter methods. Yet its employ-
ment in practical flutter analysis is inefficient compared to e.g. the p -k method. For one the calculation
times using a in-the-loop doublet lattice solver are very high. Secondly in cases where rational function
approximations are used (such as in loads, maneuver and control analysis model generation frame-
works) the application of the time-domain equivalent p method, that naturally supports the analysis of
closed loop systems, seems far more practical. Nevertheless the G AAM leads to the most accurate
results (with doublet lattice kernel in-the-loop) in terms of root evolution, because of its unrestricted
damping expansion. Furthermore, no approximations are being made other than those potentially in-
herited by the aerodynamic source. Its best application therefore probably lies in the domain of flutter
method verification by generating reference results.

FLEXOP D2.2 deliverable v01 y2016m07d01 18


Chapter 3

Modern Offline techniques

1 µ analysis

1.1 Technique summary


Despite the large amount of effort spent in flutter understanding, it is acknowledged that predictions
based only on computational analyses (Ch.2) are not totally reliable. Currently, this is compensated by
the addition of conservative safety margins to the analysis results and expensive flutter tests campaigns.
One of the main criticalities arises from the sensitiveness of aeroelastic instability to small variations in
parameter and modeling assumptions. In addressing this issue, in the last ten years researches looked
at robust modeling and analysis techniques from the robust control community, specifically Linear Frac-
tional Transformation (LFT) models [34] and µ analysis [35]. The so-called flutter robust analysis aims
to quantify the gap between the prediction of the nominal stability analysis (model without uncertainties)
and the worst-case scenario when the whole set of uncertainty is contemplated. This is believed to be
a powerful tool when used as a complement to the classic techniques in that it could highlight weak
points of the model requiring more refinement and conversely identify parameters that can be coarsely
estimated as they do not have a strong influence on the results.

1.2 Applications
Cornerstone works proving the application of µ for robust flutter analyses have been provided by re-
searchers from NASA [36], Stockholm University [37] and Technion [38].
Each of the aforementioned robust flutter approaches used the same underlying µ analysis tools but
different LFT model development path, in addition to relying on different aerodynamic approximations
(e.g. Roger or Minimum State). The goal of each of those robust flutter studies was to provide an end-
to-end process, from robust modeling to robust analysis, and demonstrate the validity of the technique.
Applications to realistic aircraft configurations were given, and indeed in [39] this approach was applied
to compute F/A-18 SRA aircraft robust flutter margins. They are evaluated with respect to an uncertainty
set generated by comparison of the nominal model with extensive flight data. The outcome indicated
that the envelope is safe from aeroelastic instabilities, although the margins lay distinctly closer to the
flight envelope than previously estimated by classic techniques.
The capability of this method to predict behaviors exhibited in flight is showcased in [40], where aeroe-

19
lastic instabilities of the F-16A/B aircraft in heavy store configuration are studied within this framework.
In particular, limit-cycle oscillations (LCOs) measured in flight-test runs were detected numerically con-
sidering parameter-dependent µ analyses with variable fuel mass and Mach.
In [11] an example is provided of the capabilities of the µ framework for robust ASE design of MIMO
uncertain systems with the aim to achieve robust stability (RS) and robust closed-loop performance
requirements, applied to a fighter aircraft, with particular emphasis on inaccuracies affecting inertial
properties.
µ analysis can advantageously be employed as a tool to determine the subsystems of an ASE plant
that more affect the stability and performance requirements, investigating the effects of the different
uncertainty levels in the robust predictions. This has been shown in [41] with application to a flexible
UAV test-bed built and tested to demonstrate flutter suppression and active wing control in flight.

1.3 Theoretical description


The LFT concept has been already discussed in the technical report of Deliverable 2.1, thus here only
a cursory description of this topic is given. The interested reader is referred to [42].
Let M be a complex matrix defined coefficient matrix and partitioned as
 
M11 M12
∈ C(p1 +p2 )×(q1 +q2 )
 
M = (3.1)
M21 M22

and let ∆u ∈ Cq1 ×p1 . The upper LFT with respect to ∆u is defined as the map

Fu (M, •) : Cq1 ×p1 −→ Cp2 ×q2


(3.2)
Fu (M, ∆u ) = M22 + M21 ∆u (I − M11 ∆u )−1 M12

The terminology is motivated by the feedback representation usually adopted to depict an LFT. It can
be seen that Fu is obtained from closing the upper loop in Fig.3.1.

Figure 3.1: Upper Linear Fractional Transformation (LFT)

Each map is associated to a set of equations, where y can be interpreted as an output of the system
when u is applied, while w and z are signals closing the feedback loop. If the relation between u and
y is sought, the expression in Eq3.2 is obtained. This suggests an interpretation for the upper LFT. If
M is taken as a proper transfer matrix, Fu is the closed-loop transfer matrix from input u to output y
when the nominal plant M22 is subject to a perturbation matrix ∆. M11 ,M12 and M21 reflect a priori
knowledge on how the perturbation affects the nominal map. Once all varying or uncertain parameters
are pulled out of the nominal plant, the problem appears as a nominal system subject to an artificial
feedback.
A crucial feature apparent from Eq.3.2 is that the LFT is well posed if and only if the inverse of
(I − M11 ∆) exists, where M11 is by definition the transfer matrix seen by the perturbation block.
Another important property of the LFTs is that general (i.e. unstructured) perturbations ∆i at compo-
nent level become structured at system level. This is a result of the process of building up the LFT:
term-by-term all the components, which are themselves uncertainties, are isolated in the plant, pulled

FLEXOP D2.2 deliverable v01 y2016m07d01 20


out and rearranged in the M-∆ structure leading to the result in Fig.3.2. This enables to exploit ad-hoc
techniques for robust analyses with structured uncertainties. In this regard, an insightful example of
LFT framework application is provided by the structured singular value µ.

Figure 3.2: LFT property: perturbations at component level become structured at system level

The structured singular value (s.s.v.) [43] is a matrix function denoted by µ∆ (M ). The mathematical
definition follows:
1
µ∆ (M ) = (3.3)
min∆ (σ̄(∆) : det(I − M∆) = 0)
where σ̄(∆) is the maximum singular value of ∆. Note that this definition can be specialized to de-
termine whether the LFT Fu (M, ∆) is well posed once the generic matrix M in the above definition is
replaced by M11 and ∆ belongs to the corresponding uncertainty set ∆. The definition points out the
dependence of the s.s.v. on two elements: the matrix M ∈ Cn×n and the uncertainty structure ∆. While
the genesis of the former has been previously illustrated, it’s worth here to better address the definition
of the uncertainties.
As previously highlighted, the LFT modeling leads, whatever are the uncertainties at a component level,
to a structured uncertainty at system level. Defining the structure means specifying three things: the
type of each block, the total number of blocks and their dimensions. Generally, two types of block are
considered: repeated scalar and full blocks. The positive integers r1 , ..., rS and m1 , ..., mF are intro-
duced such that the i th repeated scalar block is ri × ri , the j th full block is mj × mj and the total number
of repeated blocks is respectively S and F . This allows to define a set of matrices ∆ ¯ ⊂ Cn×n as

¯ = {diag[δ1 Ir1 , ..., δS IrS , ∆S +1 , ..., ∆S +F ] :



(3.4)
δi ∈ C, ∆S +j ∈ Cmj ×mj , 1 ≤ i ≤ S , 1 ≤ j ≤ F }
PS PF
where Ir is a r × r identity matrix and it has to be i =1 ri + j =1 mj = n. For ease of calculation and
interpretation, typically a norm bounded subset of ∆¯ is introduced

B∆ ¯ : σ̄(∆) ≤ 1}
¯ = {∆ ∈ ∆ (3.5)

Operatively, this normalization of the range of variation for the uncertain parameters consists simply in
a scaling of the matrix M associated to the set, once the ranges of possible values assumed by the
parameters defining the LFT has been given.
Eq.3.3 can then be specialized to the study of the robust stability (RS) of the plant represented by
Fu (M, ∆). At a fixed ω, the coefficient matrix M is a constant complex valued matrix; in particular M11
is known, which is the transfer matrix from the signal w to z. The µ calculation can then be formulated
as (β is a real positive scalar):
1
µB∆¯ (M11 ) = (3.6)
min(β : ∃∆ ∈ βB∆
¯ det(I − M11 ∆) = 0)
:

The result can then be interpreted as follows: if µB∆¯ (M11 ) ≤ 1 then there is no ∆ inside the allowable
¯ such that the determinant condition is satisfied, that is Fu (M, ∆) is well posed and thus the plant
set B∆
associated is robust stable within the range of uncertainties considered. On the contrary, if µB∆¯ (M11 ) ≥

FLEXOP D2.2 deliverable v01 y2016m07d01 21


1, then a candidate (i.e. belonging to the allowed set) perturbation matrix exists which violates the well-
posedness, i.e. the closed loop in Fig.3.1 is unstable. In particular, the reciprocal of µ (auxiliary notation
is dropped for clarity) provides directly a measure (i.e. its kk∞ norm β) of the smallest structured
uncertainty matrix that causes instability.
Finally, it is important to highlight that the s.s.v. is a robust stability test but can be used also for robust
performance (RP).
It is known that µ∆ (M ) is non-polynomial hard [44] with either pure real or mixed real-complex uncer-
tainties. This forced researchers in the last two decades to work on developing efficient algorithms for
upper and lower bound calculations. An ample review of algorithms and techniques related to µ analy-
sis can be found in the documentation of the major toolboxes devoted to µ analysis [45–47].
The upper bound µUB provides the maximum size perturbation k∆UB k∞ = 1/µUB for which RS is guar-
anteed. The default algorithm of the Robust Control Toolbox [46] is based on the search of scaling
matrices satisfying the minimization of a maximal singular value [48]. An alternative, starting from the
same premise, is based on LMI solvers and attempt to achieve the minimization of a maximal general-
ized eigenvalue [49].
The lower bound µLB guarantees the minimum size perturbation k∆LB k∞ = 1/µLB for which RS is
guaranteed to be violated. The default algorithm is based on power iterations [50], which provides high
accuracy for pure complex problems. Convergence problems may be encountered for mixed problems
and, in particular, for pure real uncertainties. If the bounds are close in magnitude then the conserva-
tiveness in the calculation of µ is small, otherwise nothing can be said on the guaranteed robustness of
the system for perturbations within [1/µUB , 1/µLB ]. The great importance to have a reliable lower bound
is also dictated by the fact that it provides the matrix ∆LB = ∆cr satisfying the determinant condition.
As a general summary on the calculation of accurate bounds, it can be affirmed that the difficulty is
mainly related to the size of the ∆ matrix and whether is complex, real or mixed. In the first case,
polynomial time algorithms are available to compute upper and lower bounds which converge exactly
to µ or remain tight for high-order problems. For system whose uncertainty description gives rise to
mixed problems, upper bounds algorithms (as those previously cited) are generally quite tight, whereas
the quality of the lower bounds considerably depend on the proportion between complex and real un-
certainties. For pure real cases, convergence of upper and lower bounds is only demonstrated for
exponential time algorithms whose computational burden prevents their use for many practical prob-
lems. Polynomial time upper bounds give generally good results, while it is usually tough to obtain
accurate predictions from lower bound calculations, even though ad-hoc algorithms have been recently
proposed [51, 52].
In conclusion, the selection of the uncertainties considered in the LFT model plays a crucial role in
providing reliable and accurate robust results. For this reason, the last part of this section presents an
overview on the possible strategies for the LFT modeling of systems for flutter purposes.
Firstly, general ways to define uncertainties in the operators are here summarized.
Parametric uncertainties are used to describe parameters whose values are not known with a satisfying
level of confidence. Considering a generic uncertain parameter d , σd is used to define the uncertainty
level and its relation with the nominal value d0 can be written as

d = d0 (1 + σd δd ) (3.7)

This expression is often referred to as multiplicative uncertainty. Analogously, an additive uncertainty is


defined as follows
d = d0 + σd δd (3.8)
At a matrix level, the operator D which is affected by parametric uncertainties can be respectively
expressed as:        
D = D0 ( I + ∆D WD )
        (3.9)
D = D0 + VD ∆D WD

FLEXOP D2.2 deliverable v01 y2016m07d01 22


VD and WD are scaling matrices which, provided the uncertainty level σd for each parameter, give a
perturbation matrix ∆D belonging to the norm-bounded set B∆ ¯ defined in Eq.3.5.
If the definition in Eq.3.2 is recalled, it is apparent that the operator D is expressed as an LFT with the
term inside the brackets null. When the symbolic expression of the uncertainties within the operators is
provided, this task can be performed using available toolboxes [46, 53]. They directly provide Vj and Wj
where j =M,C,K,A in the aeroelastic case (i.e. respectively mass, damping, stiffness and aerodynamics
uncertainties).
Additive uncertainties can be employed also to take into account so-called neglected dynamics. In this
case ∆D is an unstructured (full) complex block which enables to account for the effect of dynamics not
retained in the model, as for example residualization errors in reduced-order models [54].
The amount of uncertainty included in the model has to be defined carefully in order to avoid results
either extremely conservative or poorly robust. Depending on the source of the data adopted in the
model, possible criteria can be adopted to assign uncertainty levels (e.g. measurements errors of data
obtained experimentally; fitting errors in the approximation of computational data). A possible attempt
to rationalise this stage of the LFT modeling process could be the one based on model validation;
basically, these techniques assess whether the proposed uncertain model generates a specified input-
output data set, taken as reference. In particular the approach proposed in [55] makes advantage of
µ analysis itself, i.e. through a µ calculation involving the operators of the proposed model it can be
assessed whether it is capable of reproducing the reference data and a measure of how far it is to
exactly reproduce them is provided.
The uncertain aeroelastic plant can use either Eq.2.1 or Eq.2.9 as the nominal model. The latter de-
scribes the equilibrium in the state-space framework, while in the first this is stated in frequency domain.
These formulations are equivalent as long as all the terms involved in the plant description are rational
function of the Laplace variable s . It has been discussed that this often is not the case when unsteady
aerodynamics theories are employed and this issue is commonly addressed by the adoption of rational
approximations of the AIC operator.
In robust analysis applications, this discrepancy in the starting point of the plant has two important
consequences. Firstly, the LFT model development path, interpreted here as the process that from the
nominal plant definition leads to the framework enabling the application of µ technique, changes. Sec-
ondly, the effect on the RS predictions considering as uncertain the aerodynamic parameters changes
depending on the expression adopted for the AIC (whether it is the original frequency-dependent or its
approximation, and in this case also based on the approximation employed).
The strategy adopted to address the LFT modeling with the plant expressed in state-space stems from
the observation that the LFT can be viewed as a realization technique [53]. Hence the LFT formulae can
be used to establish the relationship between a generic transfer matrix and its state space realization,
once a proper definition for the operators M and ∆ is adopted as written in Eq.3.10:
(
ẋ = Āx + B̄u
y = C̄x + D̄u
G(s ) = D̄ + C̄(s In − Ā)−1 B̄ = Fu (M, ∆) (3.10)
 
Ā B̄ 1
M= ∆ = In
C̄ D̄ s

This shows that the problem of a plant described through its state-space realization and affected by
uncertainties can be seen as an LFT problem with the ∆ matrix split in two blocks: ∆u containing the
perturbations affecting the state-space matrices, and 1s In as previously introduced, with n the number
of states. The coefficient matrix M is partitioned correspondingly, as depicted in Fig.3.3
Calculating the frequency response of this LFT basically means to evaluate the second block at s =jω.
At this point, the LFT is back to the form originally introduced: the sub-matrix M11 of M, sought by the
µ algorithm to initialize the calculation of Eq.3.6, is the transfer matrix seen by the perturbation block

FLEXOP D2.2 deliverable v01 y2016m07d01 23


∆u . In particular, it can be demonstrated that due to the LFTs properties, the transfer matrix from w to
z is an LFT of the terms in the upper left block of the coefficient matrix in Fig.3.3, that is:

M11 (s ) = Ēw + Ēx (s In − Ā)−1 Āw = Fu (P, ∆)


(3.11)
 
Ā Āw 1
P= ∆ = In
Ēx Ēw s

Figure 3.3: LFT of an uncertain state-space model

The partitioned matrix P is composed by the nominal state-matrix Ā and other blocks, i.e. Āw , Ēx and
Ēw , arising from the uncertainty description of the plant.
The framework is now completely developed and ready to enable RS analysis of the uncertain plant
described by means of Eq.2.9. The whole procedure is automated in the Robust Control Toolbox [46]
which enables to perform robust analyses once the uncertain parameters are defined in the state-
matrix.
This problem formulation was adopted by both the NASA and Technion researchers, which in addition
to robust stability analysis, also applied it to robust optimization and/or on-line robust predictions during
flight tests (see Sec.1.2). The well-consolidated and state-space based algorithms available for these
purposes could represent a strong reason to perform LFT modeling in this framework.
The process of building up the framework for µ analyses when the plant is described in frequency-
domain (Eq.2.1) was proposed in [37] and it is now briefly considered. The general case of the dynamic
equilibrium of the plant with an input force U reads as:
 
Ms s 2 + Cs s + Ks − q A(s̄ ) X(s ) = U(s )
       
(3.12)

The idea is to recast this problem as in Fig.3.1, where y = X and u = U, in order to determine the
transfer matrix M11 :  
w= ∆ z
   
z = M11 w + M12 U (3.13)
   
X = M21 w + M22 U
The first step is to make explicit the dependence on perturbations of the four operators in Eq.2.1, taking
advantage of the relations in Eq.3.9. These expressions can thus be substituted leading, at a fixed ω,
to:  −1    
M22 X= V ∆ W X+U
 −1

2
       
 (3.14)
M22 (ω) = − ω Ms0 + i ω Cs0 + Ks0 − q A0 (k )

FLEXOP D2.2 deliverable v01 y2016m07d01 24


∆ is a block diagonal matrix holding the ∆j matrices and V, W are built respectively from Vj and
Wj . The inverse of M22 exists provided that the nominal system is stable in the considered conditions
(defined by frequency ω and wind speed V ).
Eq.3.14 can be finally recast in the template given by Eq.3.13 through these final steps:
   
w = ∆ z ;z = W X
 −1  
M22 X= V w+U
(        (3.15)
z = W M22 V w + W M22 U
    
X = M22 V w + M22 U

The last two expressions provide the sought partition of the coefficient matrix M, at a given fre-
  ω 
quency (the dependence on ω is inside M22 ). In particular, it is of interest for the analyses M11 =
W M22 V . For each value of the frequency ω in the range of interest, the terms involved in its
definition are known and then the algorithm for the calculation of µ can be initialized. It is clear that the
development path followed to get to the calculation of M11 is different within this second approach.
The main advantages of handling the analysis in this framework are here discussed. For a real aero-
nautical application (which relies on frequency-domain aerodynamic operators for flutter analyses) this
approach ensures that both nominal and robust stability refers to the same starting equation (i.e. no
approximations are used). Uncertainties can directly be expressed in the original AIC matrix, and not
in the operators defining the approximate expressions, enhancing physical considerations in the uncer-
tain definition. Matrix coefficients are complex and so are the associated uncertainties, with notable
improvement on the results accuracy. The weighting matrices defining the range of variation of the un-
certainties can have whatever dependence on the frequency, while in the former approach the rational
dependence constraint holds.

1.4 Conclusions
The use of µ analysis technique to deal with robust flutter predictions has seen a fast-growing interest in
the last decade. Research community has proven that this is an effective tool to study models affected
by uncertainties obtaining a worst-case prediction that is relatively inexpensive from a computational
point of view, as well as reliable.
An important feature of this technique is that it enriches the content of the predictions with something
more than a simple binomial-type of output, i.e. either the system is robustly stable or not within the
defined uncertainty set. It gathers various information which may help out to take design actions,
sketch out control strategies to prevent the onset of undesired phenomena, and point out weak points
of the nominal computational model to improve for more accurate results. As example of auxiliary
capabilites, it enables to perform trend (i.e. sensitivity) analyses and to show the adverse combinations
of parameters leading to the instability onset. In addition, due to its frequency-domain formulation, the
output of the analyses allow to detect multiple instability behaviors (if there are) providing an insight on
the different mechanisms taking place.
A downside of this approach is that at the present stage no fully automatic tools are available. A certain
number of software (commercial and freeware) are available to deal with single parts of the process, but
there is still no comparison with what is available for the classic nominal analyses techniques in terms
of end-to-end capabilities. As a result, the application of this method require a satisfactory knowledge
of the framework to initialise the analyses.
The numerical accuracy of the algorithms is another critical issue. It has been mentioned that the exact
calculation of µ is not feasible for practical applications and therefore various algorithms, tailored for
different uncertainty problems, are available for upper and lower bounds calculations. It is important
indeed to achieve closeness of the bounds in magnitude in order to achieve a tight estimation of the

FLEXOP D2.2 deliverable v01 y2016m07d01 25


guaranteed robustness of the system. Nonetheless, practical experience in the last 20 years by the
community has shown that with care and engineering insight most cases present no issues.

2 Bifurcation Analysis

2.1 Technique summary


Bifurcation anaysis is a technique for analysing the behaviour of nonlinear systems in terms of their
parameters. By changing the parameters of a nonlinear system, equilibria (or, as they are also referred
to: fixed points) can cease to exist, new equilibria can occur or stability of existing equilibria can change.
These phenomena mean a qualitative change to the dynamics of the system. The points where these
qualitative changes happen are called ‘bifurcations’. With bifurcation analysis multiple solutions of a
system as a function of the parameters can be traced providing valuable additional insight to the system
dynamics when compared to linear analysis.

2.2 Theoretical description


Consider, for example, a system described by the following nonlinear differential equation

ẋ = f (x ) = α + x 2 (3.16)

and its dependence on the parameter α. The fixed points of Equation (3.16) √ are defined by f (x ∗ ) = 0,
∗ 2 ∗
that is, when −α = (x ) . If α < 0 there exist two fixed points at x1,2 = ± α; the stabilty of which
can be determined by the gradient of f (x ). However, when α = 0 these two fixed points coalesce and
become one, x ∗ = 0. Moreover, if α > 0 then there are no fixed points at all; see Figure 3.4, which
shows Equation (3.16) in state space. Therefore, as the parameter α changes from negative to positive,
the system that had two equilibria, becomes a system with no equilbrium solution. The α = 0 condition
is called a ‘bifurcation’ (a saddle-node type in this case), since at this parameter value the system’s
behaviour changes qualitatively.

Figure 3.4: The plot of f (x ) for different values of α.

A convenient way to represent the behaviour of system (3.16) is the bifurcation diagram (see Figure 3.5).
In Figure 3.5 the equilibria of (3.16) are plotted against the values of parameter α. As discussed, if α < 0
there exist one stable and one unstable fixed points and as soon as α becomes positive there exist no

FLEXOP D2.2 deliverable v01 y2016m07d01 26


equilibrium solution. In an engineering system (such as that represented by Equation (3.16)), at the
bifurcation point not only the stable equilibrium loses stability but it ceases to exist all together.

Figure 3.5: Saddle-node bifurcation.

Another important example in engineering systems is the Hopf bifurcation. It can only happen in
second- or higher order nonlinear systems and it occurs when a pair of complex eigenvalues of the
Jacobian of the system crosses the imaginary axis. It corresponds to the phenomenon whereby an
equilibrium of a system loses stabilty and switches to a stable periodic oscillation (also referred to as
limit cycle oscillation or periodic orbit). There are two types of Hopf bifurcations: super- and subcritical.
To understand the difference between them, let us consider their typical bifurcation diagrams

Figure 3.6: (a) Super- and (b) subcritical Hopf bifurcations (image taken from [1]).

In Figure 3.6 both the equilibrium and the periodic solution of the system are presented in one plot. On
the horizontal axis, as in Figure 3.5, there is the parameter (in this case c ). As opposed to Figure 3.5,
however, here the vertical axis is not associated with the value of the state at the equilibrium but with
the amplitude of the steady state limit cycle oscillation, r . The equilibrium is, therefore, only represented
on the horizontal axis, since its ‘amplitude’ is zero. Let us consider the solutions that exist for different
values of parameter c . According to Figure 3.6, for values lower than co there exists a stable equilibrium
in both the super- and the subcritical cases. At co this equilibrium loses stability and remains unstable
for higher values of c . However, due to the nonlinearities in the system, other – periodic – solutions
co-exists with the stable equilibrium for different parameter values. These periodic solutions are repre-
sented by the dashed (unstable) and solid (stable) curves. The arrows show if the solution is attracting
or repelling, that is, to which solution the system (with an initial condition r ) converges to.

FLEXOP D2.2 deliverable v01 y2016m07d01 27


Supercritical case
In the supercritical case, for parameter values c < co there is an unstable periodic solution; it is denoted
by dashed lines. The amplitude of this unstable solution is ±r . Note that this solution is undetectable
physically due to it being unstable. On the other hand, for parameter values co < c < cF there exist both
a stable and an unstable periodic solution beside the unstable equilibrium. This means that if c becomes
greater than co , the equilibrium loses stability and a stable limit cycle oscillation emerges with amplitude
±r . The amplitude of this oscillation increases until the value of the parameter reaches cF , where the
limit cycle solution no longer exists and the system jumps to another solution. This might mean another
higher amplitude oscillation (not depicted in Figure 3.6) which could lead to structural failure. What
makes the supercritical Hopf bifurcation less critical than the subcritical case is the fact that if the limit
cycle oscillations are born, by decreasing the parameter values back to less than co the system returns
to the equilibrium state. In the aerospace context, with the parameter being airspeed/dynamic pressure,
this might correspond to, for example, a soft flutter condition.

Subcritical case
As opposed to the supercritical case for values cF < c < co there already exist a stable- and an
unstable periodic solution; see Figure 3.6b. This means that for one parameter value there are two
stable solutions for the describing nonlinear equations of motion (in the supercritical case there is only
one stable solution for one value of c , the type of which is either equlibrium or periodic orbit). This, in
turn, implies that even if the system is in equilibrium a large perturbation to the system may suddenly
push the system onto the periodic orbit without the gentle transition of the supercritical case. Moreover,
once the system exhibits the limit cycle oscillation, decreasing the value of the parameter only reduces
the amplitude of the oscillation and the parameter needs to be reduced to below cF where, due to the
fact that the periodic orbit no longer exists, the system jumps back to the equilibrium. An even more
dangerous property of this system is that if the equilibrium loses stabilty at co the system inevitably
jumps to the periodic orbit and, again, reducing the parameter value only reduces the amplitude of
the oscillation and the system can only recover if c is decreased to below cF , that may be far beyond
the original bifurcation point co . In an aerospace context this might correspond to a hard flutter case
whereby once the oscillation occurs the airspeed need to be reduced below the original value of onset
to recover.
Since these features of dynamical systems cannot be studied with linear techniques, bifurcation analy-
sis is a useful technique to detect multiple solutions of a nonlinear system; for further details of bifurca-
tion theory and computation of bifurcation diagrams see for example [56–58].

2.3 Continuation Methods and Bifurcation Analysis linked to control


law design
As indicated in the previous section, bifurcation diagrams illustrate the mechanisms underlying the
dynamics of a nonlinear system by showing the parameter dependence of the steady states and their
stability. These diagrams are typically generated numerically, using so-called continuation methods to
track stable and unstable steady state solutions of various types; the process is known as numerical
continuation.
An opportunity afforded by the adoption of bifurcation methods to support control law design is the
possibility to observe not just the local impact of closing the loop but indeed a more global perspective
[59, 60]. In particular, where there are multiple branches of steady state solutions for a specific range of
parameters, bifurcation analysis of the closed-loop system can indicate the effect of the controller on all
the branches. For example, in the case of the sub-critical Hopf bifurcation, does the feedback remove
the unwanted solution branch? or, does it merely move it in state-parameter space such that it may still

FLEXOP D2.2 deliverable v01 y2016m07d01 28


affect behaviour in the presence of sufficient perturbation or transient dynamics?
Numerical continuation also offers the prospect of being directly linked to a control law design technique.
This has been applied in the past for techniques such as eigenstructure assignment and LQR in aircraft
control [61–63]

2.4 Conclusions
In this section an introduction to bifurcation analysis was presented. Two typical examples of bifur-
cations in nonlinear engineering systems were discussed in detail which revealed the benefits of the
technique. It was shown how multiple solutions of a nonlinear system could be potentially result in
instability even in a region where linear stability analysis shows stable behaviour. Moreover, a short
description of how bifurcation analysis may be used in flight control was given.

FLEXOP D2.2 deliverable v01 y2016m07d01 29


Chapter 4

On-board techniques

Off line flutter modelling and analysis are useful tools to understand the dynamics of the system that is
to be flight tested and give an initial estimation of the to-be-expected flutter speed. However, modelling
is always an approximation and, in order to validate the model, flight tests are necessary. In order
for ensuring the safety of the test on-line prediction of flutter is beneficial. The flutter condition can
essentially be described by the damping ratio of one of the vibration modes becoming zero. Classical
on-line prediction methods are based on extrapolating damping measurements in order to estimate the
onset of flutter. A more advanced method, the flutterometer, uses µ-analysis to compute flutter margins
when uncertainties in the model are considered.

1 Classical damping trend tracking


The following methods are based on estimating damping trends during flight test. They are all common
in identifying the vibration modes that are critical to flutter and measuring and tracking the damping and
frequencies of these modes until a damping trend can be reliably extrapolated to the flutter condition of
zero damping. These methods use airstream turbulence in the transonic region as excitation and collect
subcritical structural responses via strain gauges placed on the wing. Data from time histories of the
responses are collected and converted to frequency spectra. The classic data acquisition techniques
are[64]:

• Peak-Hold method
This method uses the peaks of the frequency spectrum to identify the vibration modes, and then
simply tracks the response amplitude to indicate the damping trend. The name of the method
comes from the ‘peak-hold’ setting of the early spectrum analysers.

• Cross-Spectrum method
As opposed to the Peak-Hold method where only one transducer output is measured, this method
is based on two transducer outputs – the one corresponding to bending motion and another cor-
responding to the torsional motion. Here, the spectral conversion results in a cross-spectrum
where, similarly to the ‘peak-hold’ method, the peaks correspond to vibrations modes. However,
the amplitude of the response is now in a logarithmic relationship with the damping.

• Randomdec method
This method is based on the FFT of the time history signal. The time history is passed through

30
a filter with a centre frequency corresponding to that for the chosen response peak. A so-called
Randomdec signature is formed from the filtered data and the damping values of the signature
time history are calculated by means of a linear least square fit of the logarithmic values of the
signature peaks.

Following the measurements, the damping trend is then extrapolated from the available values. The
main issue with these methods is that they are based on curve fitting and then extrapolation, which
very much depends on the fitting method. Also they do not consider the possibility of nonlinearities
in damping and, therefore, may predict stability with certain flight conditions where the nonlinearities
result in unstable behaviour.

2 Flutterometer

2.1 Technique summary


The flutterometer is an on-line flutter prediction tool developed by Lind & Brenner [2]. It calculates
robust flutter margins, for given flight conditions during flight tests by means of µ-analysis. In order to
achieve this, the flutterometer uses a pre-defined nominal model as well as measured flight data. The
measurements may be used to update the nominal model and/or the associated uncertainty description
for model validation.

2.2 Theoretical description


Basic Principles
The flutterometer works with a nominal model of the flexible aircraft and a chosen set of uncertain
parameter. These uncertainties may be any structural or aerodynamic parameter of the model as long
as it is measurable. The model is represented in an LFT-form (see Figure 4.1). In Figure 4.1 P0 is
the nominal flexible model, ∆ is the structured uncertainty set. In this representation the true airspeed
(or, alternatively dynamic pressure) δq is also included in the uncertainty description even though it is
not considered as an uncertain parameter. This is simply due to the requirements from the µ-analysis.
Then, with respect to a given set of ∆, the robust flutter margin is given as the smallest velocity (or
dynamic pressure) that leads to instability for all uncertainty combinations.

Figure 4.1: LFT representation of the flexible aircraft model

The evaluation of the predicted flutter speed is as follows. At each test point, that is when the aircraft is

FLEXOP D2.2 deliverable v01 y2016m07d01 31


trimmed at a stable condition, a signal capable of exciting structural modes over a range of frequencies
must be introduced to the wing (an example system for wingtip excitation is discussed in [65]).The
structural response to the excitation is measured and the transfer function is computed between the
excitation and response. This transfer function may then be used to estimate the model properties at
the current test point which in turn may be used for updating the nominal model and/or the uncertainty
description. The flutter margin is computed at each test point with regards to the updated uncertainty
description by means of on-line µ-analysis.

Model Update
The reliability of the flutter margins predicted by the flutterometer depends on both the quality of the
measured data and the efficiency of updating the model and the associated uncertainties. A simple
way of estimating the parameters of the model at the test point is not to consider specific structural or
aerodynamic parameters but those of the combined aeroelastic model, hence, this technique updates
the model by taking into account the parameters of the coupled dynamics. The correct parameters
may surely be of interest too, however as shown in [66], the parameters of the aeroelastic model are
sufficient for flutter prediction.
The theoretical model may be considered in the form of
 
A + Â B + B̂
Y =Pu= u, (4.1)
C + Ĉ D + D̂
where P is the state-space model used for the theoretical prediction of the response Y to the excitation
u . It consists of the theoretical model {A, B , C , D } and the update matrix, {Â, B̂ , Ĉ , D̂ }. On the other
hand, another state space model may be defined in the form of
 
Ā B̄
Ȳ = P̄ u = u, (4.2)
C̄ D̄

where P̄ relates the measured response Ȳ for the same excitation. In this way, the model update is
based on errors in the parameters instead of in the actual value of the parameters. The objective of the
parameter estimation is then to choose optimal values for {Â, B̂ , Ĉ , D̂ } such that Y becomes as close
as possible to Ȳ .
The set {Â, B̂ , Ĉ , D̂ } may be identified by transforming P and P̄ into modal form:
   
R + R̂ I + Î B1 R̄ Ī B̄1
P =  I + Î R + R̂ B2  , P̄ =  Ī R̄ B̄2  . (4.3)
C1 + Ĉ1 C2 + Ĉ2 D + D̂ C̄1 C̄2 D̄

When the matrices in Equation (4.3) are expressed as transfer functions, the values of the updates
R̂ , Î , B̂1 , B̂2 , Ĉ1 , Ĉ2 , D̂ may be derived from the respective values of the theoretical and measured matri-
ces by setting Y = Ȳ ; for details see [66].
There are different ways to choose the optimal update for the nominal model. For example, one may
choose {Â, B̂ , Ĉ , D̂ } such that, at the current test point, the system identified by the measurements
becomes the updated nominal model. Although, this gives the smallest deviation between the identified
and theoretical models, it does not take into account the inaccuracies of the measurement and does not
give a reasonable uncertainty description either. Another way is to update the model with the running
averages of errors between the identified and theoretical models at each test point. This technique has
the advantage of considering consistent errors in the model, since it is assumed that random errors
have an average around zero [66].
The same set of measurement data may be used to update the uncertainty set. Updating the uncer-
tainty set means that the size of the uncertainty associated with the nominal model is increased until

FLEXOP D2.2 deliverable v01 y2016m07d01 32


the model can be validated by the flight data. This may be done locally at each test point considering
the current flight data or globally when all previous flight data are considered. For further information
see, for example, [2, 66]

Flight Data
The efficiency of updating the model and/or the uncertainty description and, hence, the accuracy of the
flutter margin prediction depends on the efficient use of the measured flight data. Due to nonlinearities
and noise in the system, the flight data does not satisfy linearity conditions in general. Since the
nominal model is represented in a linear LFT, in order to update this model, data representing the linear
dynamics of the measured system is required; therefore, the data needs filtering. An efficient way to
obtain the linear component of the data set is to perform a Volterra series expansion on the data.
The Volterra series expansion is an approximation of a nonlinear system based on the idea that the
system’s output at time t not only depends on the input at time t , but also on values of the input at
earlier time instants t − τ (that is, the method considers the system’s ‘memory’). The approximation is
an infinite series of integral operators of increasing order given by:

y (t ) = y1 (t ) + y2 (t ) + · · · + yn (t ), (4.4)

where Z t
y1 ( t ) = h1 (ξ) u (t − ξ)d ξ (4.5)
0

and Z tZ η
y2 ( t ) = h2 (ξ, η) u (t − ξ) u (t − η)d ξ d η. (4.6)
0 0

The functions h1 (ξ) and h2 (ξ, η) in Equation (4.5) and Equation (4.6) are the first and second order
Volterra kernels respectively, which are representative of the linear and higher order dynamics of the
system. By determining the optimal values for Volterra kernels, and by a convolution of the excitation
signal with the first order kernel, the linear component of the measurements may be filtered from the
measurement data; for details on Volterra kernels see, for example; [67, 68]. Updating the model may
then be based on the linear component of the flight data only.
This approach, however, still does not take into account the unmodelled dynamics. In order to do so,
identification of an operator that correlates to the unknown dynamics is necessary. This may be done,
for example, by considering the higher-order Volterra kernels [69]. Moreover, a more accurate prediction
may be obtained if the dependency of the unmodelled dynamics on the airspeed is considered. This
may be achieved by identifying parameter dependent Volterra kernels. This means that the Volterra
kernels are identified at a set of flight conditions and then a parameter dependent state-space model is
computed to represent those kernels. The resulting model is able to predict the measured responsed
at current and future airspeeds during a flight test [70].

2.3 Applications
The flutterometer has been flight tested on the NASA F/A-18 Systems Research Aircraft [71]. Details of
the results are discussed in detail in [2] however a summary of the results are given in this section. The
flight path is determined by the dynamic pressure and its distribution over time is shown in Figure 4.2.
The horizontal sections in Figure 4.2 correspond to the time spent at each test point. The variation
in the length of these sections are due to the different time needed for the update of the uncertainty
description and the subsequent µ-analysis. The dynamics of the aircraft is time-varying since the total
mass changes due to the amount of fuel decreasing. In this particular application of the flutterometer,

FLEXOP D2.2 deliverable v01 y2016m07d01 33


only the uncertainty description is updated at the test points. Since the nominal model describes the
aircraft at the heavyweight flight condition the uncertainty of the mass parameter becomes larger. Even
so, the greatest change in uncertainty belongs to the damping ratio. The amount of uncertainty in
damping as a percentage of the nominal damping ratio is shown in Figure 4.3.

Figure 4.2: Dynamic pressure over time (horizontal sections correspond to time spent at each test point)
[2]

Figure 4.3: The amount of uncertainty in damping ratio over the course of the flight test [2]

Figure 4.4 shows the predicted flutter pressure. The dotted line corresponds to the prediction of the
nominal model which is unchanged throughout the flight since the nominal model is unchanged. The
dashed line corresponds to the true flutter pressure and the solid line to the prediction of the µ-analysis.
Initially, the latter is very close to the nominal prediction, however, at the first updating instant of the
uncertainty description it drops significantly. Note, that after this the robust prediction is always below
the true flutter pressure which means that the flutterometer is able to safely predict the flutter margin of
the system.
In terms of computation times, the amount of time spent at a certain test point depends on the number
of iterations needed to update the model and/or the uncertainty description as well as the size of the
LFT model (which has an influence on the cost of the µ-analysis). Figure 4.5 shows the CPU time over
the course of the flight test (these computations were made using a 200MHz Pentium processor).
The initial horizontal section shows that the first robust flutter margin calculation does not happen
straight away. Further horizontal sections represent constant uncertainty sets with no updates. When

FLEXOP D2.2 deliverable v01 y2016m07d01 34


compared with Figure 4.3 it can be seen that the CPU time significantly increases when the uncertainty
in damping increases. Towards the end of the flight test, in the proximity of the flutter boundary, no
updates are made due to poor flight data.

Figure 4.4: Predicted flutter pressure values [2]

Figure 4.5: CPU time over the course of the flight test [2]

2.4 Conclusions
An on-line implementation that computes flutter margins throughout a flight test was presented. The
flutterometer uses flight data to update the modal and/or the uncertainty description and by means of
µ-analisys it provides an advanced and accurate prediction for the flutter velocity. The flight test results
in an F/A-18 aircraft were discussed and the basic capabilities of the flutterometer are demonstrated
via those results.

FLEXOP D2.2 deliverable v01 y2016m07d01 35


Chapter 5

Conclusion

This report gives an overview of some of the techniques used to analyse flutter in aircraft. The two
major parts of the report are the presentation of off-line and on-line methods for flutter prediction.
Chapter 2 describes traditional algorithms for offline flutter predictions, such as the State-Space eigen-
value method (that analyses the spectrum of the state-space matrix in the time domain), the p − k -
method (that solves the flutter problem directly in the frequency domain), and the GAAM method (which
generalises the solution of the flutter equation by releasing the constraints with respect to the frequency
domain expansion in s ).
In Chapter 3 modern off-line techniques are presented. µ-analysis for flutter prediction is discussed as
an advanced technique to analyse robust flutter margins in terms of uncertainties in the parameters of
the system. Moreover, a brief overview of Bifurcation analysis is given, which is a technique used to
analyse the effects of nonlinearities on the system.
Chapter 4 illustrates the on-board flutter prediction methods, and introduces the flutterometer which
uses µ-analysis for on-board flutter margin calculations (and was successfully flight-tested by NASA).

36
Chapter 6

Bibliography

[1] Walter Lacarbonara. Nonlinear structural mechanics: theory, dynamical phenomena and model-
ing. Springer Science & Business Media, 2013.
[2] Rick Lind and Marty Brenner. Flutterometer: an on-line tool to predict robust flutter margins.
Journal of Aircraft, 37(6):1105–1112, 2000.
[3] P.C. Chen. Damping perturbation method for flutter solution: The g-method. AIAA Journal,
38(9):1519–1524, 2000.
[4] E.H. Dowell, H.C. Curtiss, R.H. Scanlan, and F. Sisto. A Modern Course in Aeroelasticity. Me-
chanics: Dynamical Systems. Springer Netherlands, 2013.
[5] J. W. Edwards and C. D. Wieseman. Flutter and divergence analysis using the generalized aeroe-
lastic analysis method. Journal of Aircraft, 45(3):906–915, 2008.
[6] H. J. Hassig. An approximate true damping solution of the flutter equation by determinant iteration.
Journal of Aircraft, Vol. 33(7), 1971.
[7] E. Albano and W.P. Rodden. A doublet lattice method for calculating lift distributions on oscillating
surfaces in subsonic flows. American Institute of Aeronautics and Astronautics, 7:279–285, 1968.
[8] P. C. Chen and D. D. Liu. Harmonic gradient method for unsteady supersonic flow calculations.
Journal of Aircraft, 22(5):371–379, 1985.
[9] H. Tiffany and M. Adams. Nonlinear programming extensions to rational function approximation
methods for unsteady aerodynamic forces. Technical paper 2776, Nasa, 1988.
[10] R. Botez, A. Hiliuta, N. Stathopoulos, S. Therien, A. Rathe, and M. Dickinson. Aeroservoelastic
flutter and frequency response interactions on the cl-704 aircraft. INCAS Bulletin, Vol. 4(No. 4):pp.
41–56, 2012.
[11] B. Moulin, M. Idan, and M. Karpel. Aeroservoelastic structural and control optimization using
robust design schemes. J. of Guidance, Control and Dynamics, Vol. 25(No. 1):pp. 152–159, 2002.
[12] L.O. Bernhammer, R. De Breuker, M. Karpel, and G.J. van der Veen. Aeroelastic control using
distributed floating flaps activated by piezoelectric tabs. Journal of Aircraft, 50(3):732–740, 2013.
[13] C. Nam, P. C. Chen, D.D. Liu, and A. Chattopadhyay. Astros with smart structures and ase
modules: Application to flutter suppression and gust-load alleviation. 41st AIAA/ASME/AHS/ASC
Structures, Structural Dynamics, and Materials Conference,Atlanta, GA, USA, April 2000.

37
[14] K. Gupta. Ae/ase simulation of aerospace vehicles using stars software. 45th
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics & Materials Conference, Struc-
tures, Structural Dynamics, and Materials and Co-located Conferences, 19-22 April 2004.
[15] AZ ZONA Thechnology Inc., Scottsdale. ZAERO Version 5.2. 2004. Theoretical Manual, Version
7.1.
[16] K.L. Roger. Airplane math modeling methods for active control design. AGARD-CP-228, August
1977.
[17] M. Karpel. Design for active and passive flutter suppression and gust alleviation. Nasa Report,
(3482), 1981.
[18] R. L. Bisplinghoff, H. Ashley, and R. L. Halfman. Aeroelasticity. Addison-Wesley, New York, 1955.
[19] I. Abel. An analytical technique for predicting the characteristics of a flexible wing equipped with
an active flutter-suppression system and comparison with wind-tunnel data. Nasa Technical Paper
1367, 1979.

[20] D. H Hodges and G A. Pierce. Introduction to structural dynamics and aeroelasticity; 2nd ed.
Cambridge aerospace series. Cambridge University Press, New York, NY, 2011.
[21] R. J. Allemang and D. L. Brown. A correlation coefficient for modal vector analysis. 1st International
Modal Analysis Conference, 1982.

[22] N.P.M. Werter, R. De Breuker, and M. M. Abdalla. Continuous-time state-space unsteady aero-
dynamic modelling for efficient aeroelastic load analysis. In International Forum on Structural
Dynamics and Aeroelasticity, St. Petersburg, Russia, 2015.
[23] B.K. Stanford. Role of unsteady aerodynamics during aeroelastic optimization. AIAA Journal,
53(12):3826–3831, 2015.

[24] B. Stanford and C. Wieseman, C.D. Jutte. Aeroelastic tailoring of transport wings including tran-
sonic flutter constraints. 56th AIAA/ASCE/AHS/ASC Structures, Structural Dynamics, and Materi-
als Conference, 2015.
[25] M. Karpel, B. Moulin, and P.C. Chen. Extension of the g-method flutter solution to aeroservoelastic
stability analysis. Journal of Aircraft, 42(3):789–792, 2005.
[26] Z. Wang and P.C. Chen. Adapted k-method for frequency-domain ase control stability margin
analysis. 55th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Con-
ference, 2014.
[27] W. P. Rodden and E. H. Johnson. User Guide V 68 MSC/NASTRAN Aeroelastic Analysis.
MacNeal-Schwendler Corporation, 1994.
[28] A.J. Lawrence and P. Jackson. Comparison of different methods of assessing the free oscillatory
characteristics of aeroelastic systems. Current paper 2776, Aeronautical Research Council, 1970.
[29] W.P. Rodden, R.L. Harder, and E.D. Bellinger. Aeroelastic addition to nastran. Technical Report
3094, 1979.
[30] R. Yurkovich. Status of unsteady aerodynamic prediction for flutter of high-performance aircraf.
Journal of Aircraft, 40(5):832–842, 2003.
[31] L.H. van Zyl and M.S. Maserumule. Divergence and the p-k flutter equation. Journal of Aircraft,
38(3):584–586, 2001.

FLEXOP D2.2 deliverable v01 y2016m07d01 38


[32] J. W. Edwards. Applications of Laplace transform methods to airfoil motion and stability calcula-
tions. In 20th Structures, Structural Dynamics and Materials Conference, number AIAA 1979-772,
1979.
[33] L. H. van Zyl. More on the generalised aeroelastic analysis method (gaam). In International Forum
on Aeroelasticity and Structural Dynamics, number IF-095, 2005.
[34] J. Doyle, A. Packard, and K. Zhou. Review of lfts, lmis, and mu. Number Proceedings. Vol. 2. IEEE
Conference on Decision and Control, 30th, Brighton, United Kingdom, Dec. 11-13, 1991.
[35] Andrew Packard and J. Doyle. The complex structured singular value. Automatica, 29(1):pp.
71–109, 1993.
[36] R. Lind and M. Brenner. Robust Aeroservoelastic Stability Analysis. Advances in Industrial Control.
Springer London, 2012.
[37] D Borglund. The µ-k method for robust flutter solutions. Journal of Aircraft, 41(5):1209–1216,
2004. QC 20100525 QC 20110923.

[38] M. Idan, M. Karpel, and B. Moulin. Aeroservoelastic interaction between aircraft structural and
control design schemes. J. of Guidance, Control and Dynamics, Vol. 22(No. 4):pp. 513–519, 1999.
[39] R. Lind and M. Brenner. Robust flutter margins of an f/a-18 aircraft from aeroelastic flight data. J.
of Guidance, Control and Dynamics, Vol. 20(No. 3):pp. 597–604, 1997.

[40] S. Bennani, B. Beuker, J.W. van Staveren, and J.J. Meijer. Flutter analysis for the f-16a/b in heavy
store configuration. Journal of Aircraft, 42(6):1566–1575, 2005.
[41] A. Kotikalpudi, H. Pfifer, and P. Seiler. Sensitivity of robust flutter boundary to model uncertainties
in aeroservoelastic systems. AIAA Atmospheric Flight Mechanics Conference,, January 2016.

[42] Kemin Zhou, John C. Doyle, and Keith Glover. Robust and Optimal Control. Prentice-Hall, Inc.,
Upper Saddle River, NJ, USA, 1996.
[43] J. Doyle. Structured uncertainty in control system design. IEEE Conference on Decision and
Control, 25th, Fort Lauderdale, FL, USA, Dec. 11-13, 1985.
[44] R.D. Braatz, P.M. Young, J.C. Doyle, and M. Morari. Computational-complexity of mu-calculation.
IEEE Transactions on Automatic Control, 39(5):1000–1002, 1994.
[45] G.J. Balas, J.C. Doyle, K. Glover, A. Packard, and R. Smith. µ Analysis and Synthesis Toolbox.
1998.
[46] G.J. Balas, R. Chiang, A. Packard, and M. Safonov. Robust Control toolbox. 2005.

[47] G. Ferreres, J.M. Biannic, and J.F. Magni. A skew µ toolbox (smt) for robustness analysis. IEEE
International Symposium on Computer Aided Control Systems Design, 2-4 September 2004.
[48] P.M. Young, M.O. Newlin, and J. Doyle. Practical computation of the mixed µ problem. American
Control Conference, Chicago, IL, USA, December 1992.

[49] M.K.H. Fan, A.L. Tits, and J. Doyle. Robustness in the presence of mixed parametric uncertainty
and unmodeled dynamics. IEEE Transactions on Automatic Control, Vol. 36(No. 1):pp. 25–38,
1991.
[50] A. Packard, M.K.H Fan, and J. Doyle. A power method for the structured singular value. Proc. of
the 27th Conference on Decision and Control, Austin, TX, USA, December 1988.

FLEXOP D2.2 deliverable v01 y2016m07d01 39


[51] P. Seiler, A. Packard, and G. J. Balas. A gain-based lower bound algorithm for real and mixed µ
problems. Automatica, Vol. 46(No. 3):pp. 493–500, 2010.
[52] M.J. Hayes, D.G. Bates, and I. Postlethwaite. New tools for computing tight bounds on the real
structured singular value. Journal of Guidance, Control, and Dynamics, Vol. 24(6):pp. 1204–1213,
2001.
[53] J.F. Magni. Linear fractional representation toolbox modelling, order reduction, gain scheduling.
Technical report tr 6/08162, DCSD, ONERA, Systems Control and Flight Dynamics Department,
2004.

[54] M. Karpel, B. Moulin, and M. Idan. Robust aeroservoelastic design with structural variations and
modeling uncertainties. Journal of Aircraft, 40(5):946–952, 2003.
[55] A. Kumar and G.J. Balas. An approach to model validation in the mu framework. Proc. of the
American Control Conference, Baltimore, Maryland, June 1994.
[56] Steven H Strogatz. Nonlinear dynamics and chaos: with applications to physics, biology, chemistry,
and engineering. Westview press, 2014.
[57] Yuri A Kuznetsov. Elements of applied bifurcation theory, volume 112. Springer Science & Busi-
ness Media, 2013.
[58] Eusebius J Doedel, Thomas F Fairgrieve, Björn Sandstede, Alan R Champneys, Yuri A Kuznetsov,
and Xianjun Wang. Auto-07p: Continuation and bifurcation software for ordinary differential equa-
tions. 2007.
[59] Stephen J Gill, Mark H Lowenberg, Simon A Neild, Luis G Crespo, Bernd Krauskopf, and Guil-
hem Puyou. Nonlinear dynamics of aircraft controller characteristics outside the standard flight
envelope. Journal of Guidance, Control, and Dynamics, 38(12):2301–2308, 2015.

[60] Xiao Fan Wang, Mario di Bernardo, David P Stoten, MH Lowenberg, and G Charles. Bifurcation
tailoring via newton flow-aided adaptive control. International Journal of Bifurcation and Chaos,
13(03):677–684, 2003.
[61] Thomas Richardson, Mark Lowenberg, Mario DiBernardo, and Guy Charles. Design of a gain-
scheduled flight control system using bifurcation analysis. Journal of guidance, control, and dy-
namics, 29(2):444–453, 2006.
[62] TS Richardson and MH Lowenberg. A continuation design framework for nonlinear flight control
problems. Aeronautical Journal, 110(1104):85–96, 2006.
[63] CD C. Jones, MH Lowenberg, and TS Richardson. Tailored dynamic gain-scheduled control.
Journal of guidance, control, and dynamics, 29(6):1271–1281, 2006.
[64] CL Ruhlin, JJ Watson, RH Ricketts, and RV Doggett Jr. Evaluation of four subcritical response
methods for on-line prediction flutter onset in wind-tunnel tests.[conducted in the langley transonic
dynamics tunnel. 1982.
[65] Lura Vernon. In-flight investigation of a rotating cylinder-based structural excitation system for
flutter testing. 1993.
[66] Rick Lind and Joao Pedro Mortagua. Reducing conservatism in flutterometer predictions using
volterra modeling with modal parameter estimation. Journal of aircraft, 42(4):998–1004, 2005.
[67] Richard J Prazenica, Rick Lind, and Andrew J Kurdila. Uncertainty estimation from volterra kernels
for robust flutter analysis. Journal of guidance, control, and dynamics, 26(2):331–339, 2003.

FLEXOP D2.2 deliverable v01 y2016m07d01 40


[68] Martin Schetzen. The volterra and wiener theories of nonlinear systems. 1980.
[69] Rick Lind, Richard J Prazenica, and Martin J Brenner. Estimating nonlinearity using volterra ker-
nels in feedback with linear models. nonlinear dynamics, 39(1-2):3–23, 2005.

[70] Rick Lind, Richard J Prazenica, Martin J Brenner, and Dario H Baldelli. Identifying parameter-
dependent volterra kernels to predict aeroelastic instabilities. AIAA journal, 43(12):2496–2502,
2005.
[71] Martin J Brenner, Richard C Lind, and David F Voracek. Overview of recent flight flutter testing
research at NASA Dryden, volume 4792. National Aeronautics and Space Administration, Office
of Management, Scientific and Technical Information Program, 1997.

FLEXOP D2.2 deliverable v01 y2016m07d01 41

You might also like