See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/258569079

Efficient Method for Limit Cycle Flutter Analysis Based on Nonlinear

Aerodynamic Reduced-Order Models

Article in AIAA Journal · May 2012

DOI: 10.2514/1.J050581

CITATIONS READS

142 795

4 authors, including:

Weiwei Zhang Kun Ye

Northwestern Polytechnical University Northwestern Polytechnical University
191 PUBLICATIONS 3,194 CITATIONS 104 PUBLICATIONS 1,431 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Nonlinear Aerodynamic Reduced-Order Modeling Based on System Identification View project

aerothermoelastic View project

All content following this page was uploaded by Weiwei Zhang on 09 February 2015.

The user has requested enhancement of the downloaded file.

 AIAA JOURNAL
Vol. 50, No. 5, May 2012

Efficient Method for Limit Cycle Flutter Analysis by Nonlinear

Aerodynamic Reduced-Order Models

Weiwei Zhang∗
Northwestern Polytechnical University, Xi’an 710072, China
Bobin Wang†
Tsinghua University, Beijing 100084, China
and
Zhengyin Ye‡ and Jingge Quan§
Northwestern Polytechnical University, Xi’an 710072, China
DOI: 10.2514/1.J050581
Downloaded by Northwestern Polytechnic University on May 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/1.J050581

Nonlinearities can be present in an aeroelastic system because of some aerodynamic factors that occur in transonic
flight regimes or at a large angle of attack. The sources are shock wave motions and separated flows. Complex
aeroelastic problems due to the aerodynamic nonlinearity can be studied using high-fidelity computational fluid
dynamics (CFD) codes. However, the computational cost may be very high. Hence, this kind of problem is
conveniently solved using the reduced-order model (ROM) for unsteady aerodynamic analysis. Many studies have
been done using reduced-order modeling for aeroelastic analysis. However, most of the reduced-order aerodynamic
models are dynamic linear models and have loads proportional to the structural motions. In the current paper, a
nonlinear unsteady reduced-order aerodynamic model is constructed using the radial basis function neural network
model. This kind of ROM is used to analyze the limit cycle oscillation (LCO) for two linear structural models with
large shock motions in transonic flow. Unlike the input signals in the traditional design method, the signals of the self-
excited vibration of the aeroelastic system are designed as the input signals in the current paper. Coupling the
structural equations of motion and nonlinear aerodynamic ROM, the system responses are determined by time
marching the governing equations using a hybrid linear multistep algorithm. Then, the LCO change with velocities
(dynamic pressure) was analyzed. The two transonic aeroelastic examples show that both the structural responses
and the LCO characteristics simulated using the nonlinear ROM agree well with those obtained using the direct CFD
method. Moreover, the computational efficiency of the nonlinear ROM-based method is improved by one to two
orders of magnitude compared with that of the direct CFD method.

Nomenclature n = output delay order of aerodynamic model

a = location of airfoil elastic axis Q = generalized force vector
b = half-chord length q = dynamic pressure of the free flow
Cl = coefficients of lift r
= radius of gyration of airfoil about elastic axis
Cm = coefficients of pitch moment s = dimension of input vector x
c = number of neurons in the hidden layer V = velocity of the free flow
fa = generalized aerodynamic force coefficient vector V = reduced velocity
G = generalized damping matrix y = output vector radial basis function neural network
h = airfoil plunge
j = center vector for neuron j
K = generalized stiffness matrix Wi;j = weight of the linear output neuron
k = reduced frequency, !

b=V x = input vector of radial basis function neural network,
l = dimension of output vector y structural state vector
M = generalized mass matrix x
= airfoil static unbalance
m = input delay order of aerodynamic model
= airfoil pitch angle
 = mass ratio
 = generalized structural coordinate vector
Presented as Paper 2010-2723 at the 51st AIAA/ASME/ASCE/AHS/ASC j = spread of neuron
Structures, Structural Dynamics Conference, Orlando, FL, 12–15 April 2010; !h , !
= uncoupled natural frequency of plunge and pitch
received 14 April 2010; revision received 14 November 2011; accepted for
publication 6 December 2011. Copyright © 2011 by the American Institute of
Aeronautics and Astronautics, Inc. All rights reserved. Copies of this paper
may be made for personal or internal use, on condition that the copier pay the
$10.00 per-copy fee to the Copyright Clearance Center, Inc., 222 Rosewood
I. Introduction
EROELASTICITY is a multidisciplinary field of study dealing
Drive, Danvers, MA 01923; include the code 0001-1452/12 and $10.00 in
correspondence with the CCC.
∗
Associate Professor, National Key Laboratory of Science and Technology
A with structural behavior under the interaction of inertial,
structural, and aerodynamic forces. An aeroelastic system can have
on Aerodynamic Design and Research, College of Aeronautics; nonlinearities due to some phenomena that occur in the structure,
[email protected] (Corresponding Author). aerodynamics, and servo control systems. These nonlinearities affect
†
Ph.D. Student, School of Aerospace; [email protected].
‡ the aeroelastic behavior in a manner that cannot be predicted using
Professor, National Key Laboratory of Science and Technology on
Aerodynamic Design and Research, College of Aeronautics; yezy@nwpu.
linear analysis methods. The effects of different structural and aero-
edu.cn. dynamic nonlinearities on aeroelastic systems are discussed in [1–3].
§
Ph.D. Student, National Key Laboratory of Science and Technology on The typical nonlinear aeroelastic response phenomenon is the
Aerodynamic Design and Research, College of Aeronautics; pigeon729@ amplitude limit cycle oscillation (LCO), which can be considered as
163.com. bounded flutter.
1019
 1020 ZHANG ET AL.

Structural nonlinearities arise from worn hinges of control harmonic balance method has also been applied to an aeroelastic
surfaces, loose control linkages, material behavior, and other various NACA 64a010/NLR 7301 airfoil model to determine the LCO in
sources. LCOs due to classical nonlinearities, such as cubic spring, transonic flow. Computational times are decreased from hours to
free play, and hysteresis, have been investigated. For some flexible or seconds using the nonlinear ROM.
large structures, such as high-aspect-ratio wings [4–6] or flexible In 1997, Faller and Schreck [30] used recursive neural networks
panels [7,8], the effects of geometric nonlinearities on flutter and (RNNs) to model the nonlinear unsteady fluid mechanics system. For
LCOs have also been studied. the harmonic pitch motions of a wing in a wind tunnel, the input to the
Aerodynamic nonlinearities occur primarily in the transonic flight RNN is the unsteady motion history, composed of instantaneous
,
regime or at a large angle of attack. The sources are shock wave _ and angular acceleration
.
angular velocity
,
The targeted output is
motions and separated flows. With the recently well-developed either the surface pressure or the shear-stress values at time (t  t).
software and hardware technologies, the numerical simulation of The time dependence was modeled by feeding the RNN predictions
complex aeroelastic phenomena has become possible. LCOs due to back to the RNN as input throughout the motion time history. RNN
the aerodynamic nonlinearity are studied in [9–12] using high- was shown to predict unsteady boundary-layer development, separa-
fidelity computational fluid dynamics (CFD) codes. However, such tion, dynamic stall, and dynamic reattachment. Moreover, Marques
numerical models may involve hundreds, thousands, or even and Anderson [31] predicted unsteady transonic aerodynamic loads
millions of degrees of freedom, which limit their applications in using a kind of temporal neural network based on multilayer
optimal design or system control. functions. The aerodynamic database of pitching airfoil is created
Many studies on reduced-order modeling have recently been using CFD codes based on Euler equations. Three typical motions,
performed. It is a novel concept and captures the nonlinear flow char-
Downloaded by Northwestern Polytechnic University on May 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/1.J050581

comprising sinusoidal, ramp-up, and pulse-down input time his-

acteristics more efficiently than full CFD simulations. Dowell and tories, are considered in two case studies. Network models identified
Hall [13], Lucia et al. [14], and Zhang and Ye [15] have presented for fixed Mach numbers and for a range of Mach numbers are both
some overviews of reduced-order models (ROMs) and their appli- presented in the current study. The prediction precision of a moment
cations on nonlinear aeroelastic research. Two kinds of methods are coefficient is worse than that of a lift coefficient. Recently, a kind of
used to reduce the order of an aerodynamic model: the proper surrogate-based recurrence framework (SBRF) approach to reduced-
orthogonal decomposition (POD) technique and the aerodynamic order nonlinear unsteady aerodynamic modeling associated with
modeling based on system identification technology. For the system pitching/plunging airfoils for fixed or time-varying freestream Mach
identification method, both the integral model, such as the Volterra numbers was developed by Glaz et al. [32]. This kind of SBRF is
series, and the difference model, such as the autoregressive with commonly referred to as a nonlinear autoregressive moving average
exogenous input (ARX) model, can be used. In our research group, with exogenous inputs (NARMAX) model, which is the nonlinear
ROMs based on ARX have been applied on transonic flutter analysis expression of the ARMAX model. Once constructed from a limited
[16], flutter analysis at a high angle of attack [17], gust analysis [18], number of full-order CFD analyses, the SBRF predictions require a
aeroservoelastic parameterical analysis [19], and control law design fraction of a second to compute while maintaining a level of fidelity
[20]. In [21–23], ROMs have been used to study LCO with structural corresponding to CFD results that would require several hours to
nonlinearity in transonic flow. They can capture the steady nonlinear obtain.
characteristics of the flows, but the nonlinear sources of LCOs are In the current paper, using the radial basis function (RBF) neural
from the structural system (the control surface free play, geometric network model, the reduced-order aerodynamic model for nonlinear
stiffness nonlinearities, and piecewise linear model in [21–23], flow (M of 0.8, with pitch amplitude up to 10 deg and large-scale
respectively). In the preceding studies, the reduced-order aero- shock motion) is constructed in time domain. The nonlinear ROM is
dynamic models (such as POD [21,22], first-order Volterra series then applied to analyze LCOs for two aeroelastic airfoils with linear
[23], and ARX [16–20]) are dynamic linear models in which the structural parameters in transonic flow.
aerodynamic load is proportional to the structural motion. They are
suitable for transonic flutter computations with the assumption of
small-amplitude vibrations.
A kind of nonlinear system identification model based on the II. Nonlinear Reduced-Order Model Based on Radial
second-type Volterra series in [24] has been used to capture the Basis Function Neural Network
transonic flutter boundary. The unsteady flow aroused by small- The artificial neural network is a new approach to the approxi-
amplitude structural vibrations in transonic flow is essentially a linear mation of a nonlinear system. The use of neural networks in
dynamic system [1,13]. Raveh [25] found that the accuracy of this nonlinear system modeling was well documented in [33].
kind of ROM is dependent on the choice of input amplitude and The RBF emerged as a variant of artificial neural network in the
computational time step and that the kernels are more sensitive to late 1980s. RBF neural networks are suitable for modeling nonlinear
inaccuracies in the impulse responses used for their identification data. They can be trained in one stage rather than using an iterative
compared with that in the step responses. The introduction of second- process as in the multilayer perceptron network and can also learn the
order kernels does not significantly improve the predictions, given application quickly. With enough hidden neurons, the RBF
indicating that true nonlinear identification is difficult to perform. network can approximate any continuous function with arbitrary
The use of a first-order, step-type ROM can accurately predict mildly precision. An RBF neural network is an artificial neural network that
nonlinear responses. In a recent review paper, Silva [26] suggests that uses RBFs as activation functions. It typically has three layers: an
the application of the Volterra theory to multi-degree-of-freedom input layer, a hidden layer with a nonlinear RBF activation function,
systems is incomplete. He states that “An important issue that needed and a linear output layer. Depending on the number of hidden
to be addressed is the simultaneous excitation of multiple degrees of neurons, two kinds of RBF networks exist. They are the regular-
freedom to properly identify any nonlinear crosscoupling of the ization and generalized RBF networks. The former is used in the
degrees of freedom.” In a recent work [27], the multi-input Volterra current paper. In the regularization RBF network, the number of
theory has been applied to the mildly nonlinear multi-degree-of- hidden units is equal to the size of the training sample, and the
freedom aerodynamic system (NACA 0012 airfoil for forced pitch activation functions of the hidden layer are defined by a set of Green’s
and plunge motion at an M of 0.6; dynamic pitch amplitude is about functions centered at the training data points. The linear weights of
1 deg). The identified cross kernels have been found to increase the the output layer are the only unknown parameters of the network.
modeling accuracy of the Volterra ROM significantly. However, the Figure 1 shows the scheme of a regularization RBF network. An
sensitivity analysis shows that the accuracy of the ROM has effects input vector x is used as input to all RBFs, each with different
on some parameters, such as impulse magnitude, time step, and so parameters. The neurons in the hidden layer contain Gaussian
on. A method based on the Taylor series expansion of a frequency- transfer functions of which outputs are inversely proportional to the
domain harmonic balance computational fluid dynamic solver distance from the center of the neuron. The output of the network is a
residual has been developed by Thomas et al. [28,29]. This kind of linear combination of the output from RBFs.
 ZHANG ET AL. 1021

uk
x1 ρ (|| x-v1||) Σ y1 uk-1

uk-m
x2 ρ (|| x-v2||) Σ y2 RBF
yk
Network
yk-1
yk-2

xs ρ (|| x-vc||) Σ yl
yk-n
Fig. 1 Scheme of RBF neural network.
Delay
Fig. 2 Scheme of recursive RBF neural network.
For a regularization RBF neural network, x ! y, x 2 Rs , y 2 Rl
(s is the dimension of the input vector x, and l is the dimension of the
output vector y), the output of the network is thus where D is the desired response vector and matrix G is the
pseudoinverse of matrix G and is defined as
X
c
Downloaded by Northwestern Polytechnic University on May 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/1.J050581

yi  wi;j
kx 
j k; 1il (1)

j1
G   GT G1 GT

where c is the number of neurons in the hidden layer,
j is the center where
vector for neuron j, and wi;j are the weights of the linear output G  fgij g
neuron.
Basically, each input is connected to a hidden neuron. The norm is

typically taken to be the Euclidean distance, and the basis function is kxi  j k2
gij  exp
taken to be Gaussian: 2 2j


kx 
j k2
kx 
j k  exp ; j > 0 (2) Although the preceding RBF neural network is a feedforward
2j2 neural network with three layers, the output is only a nonlinear
mapping of the inputs. Hence, this kind of model cannot be used for
The parameter j is the spread (radius) of neuron j. With larger dynamic systems. An unsteady flow with a large amplitude of shock
spreads, neurons at a distance from a point have greater influence. wave motion or flow separation is a typical nonlinear dynamic
The Gaussian basis functions are local in the sense that system. Dynamic delays are needed to model this kind of system. In
the current paper, the recursive RBF neural network is used to
lim kx 
j k  0 (3) construct the nonlinear unsteady aerodynamic model. The scheme of
kxk!1
a recursive RBF neural network is shown in Fig. 2, and its
The changing parameters of one neuron have little effect on input corresponding function can be written as
values that are far from the center of that neuron.
In the RBF network, the variances , the positions of the centers  ff ak gT  fk ; k1 ; k2 ; . . . ; km ; f ak1 ; f ak2 ; . . . ; f akn  (4)
of the RBFs, and the linear weights w associated with the output layer
where f is a nonlinear function,  is the vector of system inputs
are all unknown parameters that need to be learned. Depending on
(generalized structural coordinate vector), and f a is the vector of
how the centers of the RBF network are specified, four different
system outputs (generalized aerodynamic force coefficient vector).
learning strategies can be followed in designing an RBF network.
m and n are the delay orders that are determined by the user. Vectors
They are the fixed centers selected at random method, the self-
Tk ; Tk1 ; Tk2 ; . . . ; Tkm ; f Tak1 ; f Tak2 ; . . . ; f Takn T and f ak corre-
organized selection of centers method, the supervised selection of
spond to vectors x and y in Fig. 1, respectively.
centers method, and the strict interpolation with regularization
The excitation signals (the generalized displacement) were
method.
inputted into the unsteady aerodynamic solver, and the corre-
The present paper uses the fixed centers selected at random
sponding outputs (generalized aerodynamic loads) were generated.
method. The activation functions of the hidden units are assumed to
With the obtained training signal, the nonlinear ROM can be
be the fixed RBFs. The location of the centers can be chosen
constructed using the recursive RBF neural network. Coupling the
randomly from the training data set. For the regularization RBF
structural equations of motion and nonlinear aerodynamic ROM, we
network designed in the current paper, the number of hidden units is
determined the system responses by time marching the governing
equal to the number of training samples so that one hidden unit is
used for each input vector xi, i  1; 2; . . . ; N, where N is the size of
the training sample. Hence, the centers of the RBFs are the training
samples. The standard deviation of all the Gaussian RBFs, which is Unsteady flow solver
fixed according to the spread of the centers, is defined as
dmax Identification
  p
 of RBF network
2m1
Structural
where dmax is the maximum distance between the chosen centers and equations Nonlinear ROM of
m1 is the number of centers. This formula makes the RBFs neither too aerodynamics
peaked nor too flat. In this approach, the only parameters that need to
be learned are the linear weights in the output layer of the network. A
simple procedure is to use the pseudoinverse method [33–35].
Generalized aerodynamic loads
Specifically,
q
W G D 
Fig. 3 Flowchart of aeroelastic analysis based on ROM.
 1022 ZHANG ET AL.

equations, and the limit cycle behaviors changing with dynamic responses of nonlinear dynamic systems constructed with an
pressures were analyzed. The flowchart of this kind of time simu- artificial neural network. A coupled fluid–structure simulation
lation method based on nonlinear ROM is shown in Fig. 3. algorithm (referred to as the hybrid multistep algorithm) [36] can be
used to integrate Eq. (6) in the time domain along with any unsteady
flow solver or ROM to determine the characteristics of the coupled
III. Time Marching Method aeroelastic system. Several such time domain simulations are needed
of the Aeroelastic Equations to determine the LCO characteristics for each Mach number.
The following equation is the structural motion equation: A Euler-equation-based unsteady flow solver has the ability to
simulate the flow with a strong shock motion, which is used in the
M 
 G_  K  Q  q
f a (5) following studies. A finite volume based unstructured Euler solver is
used for unsteady flow computations. It is suitable for moving bodies
M, G, K, Q and q are the generalized mass, damping, and stiffness using an arbitrary Lagrangian–Eulerian formulation and adding a
matrices, generalized force vector, and dynamic pressure of the dynamically deforming mesh algorithm. Spatial discretization is
free flow, respectively. By defining the structural state vector x accomplished by a cell-centered finite volume formulation using a
1 ; 2 ; . . . ; n ; _1 ; _2 ; . . . ; _n ; T , we obtain the following structural center differencing scheme or AUSM  up wind scheme. A second-
equation in state-space form: order accurate full implicit scheme is used to integrate the equations
in time domain, and a fourth-order Runge–Kutta time marching
x_  fx; t  A
x  B
Qx; t (6) method is used in the pseudotime step. Local time stepping and
Downloaded by Northwestern Polytechnic University on May 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/1.J050581

implicit residual smoothing accelerate its convergence. The vali-

where dation with experiments for static and unsteady aerodynamic coef-
  ficients has been documented previously. A more detailed process
O I can be found in [37].
A
M1 K M1 G A type of second-order hybrid linear multistep scheme for coupled
fluid–structure simulation [36] is used to integrate Eq. (6) in the time
and domain along with any unsteady load solver (such as the direct CFD
method or the RBF-based nonlinear ROM) to determine the char-
  acteristics of the coupled aeroelastic system.
O Equation (6) can be integrated by the standard explicit second-
B
M1 order Adams linear multistep scheme:

Equation (6) is a first-order differential equation. t t

xn1  xn  3f n  f n1   xn  3A
xn  A
xn1 
A two-dimensional aeroelastic model with 2 degrees of freedom in 2 2
plunge and pitch is illustrated in Fig. 4. h and
are the airfoil plunge t
and pitch angle, respectively.Cl and Cm are the coefficients of the lift  3B
Qn  B:Qn1  (7)
2
and pitch moments, respectively. r
is the gyration radius of the
airfoil around the elastic axis. x
is the airfoil static unbalance It can also be integrated by implicit second-order Adams linear
(dimensionless distance between the center of gravity and the hinge multistep scheme using the predictor-corrector procedure:
axis). !h and !
are the uncoupled natural frequencies of the plunge 8
and pitch. b, , V, and V  are the half-chord length, mass ratio, > x n1  xn  t 3f n  f n1 xn  t 3A
xn  A
xn1 
>
> 2 2
velocity of the free flow, and reduced velocity, respectively. <  t 3B
Q  B
Q 
2 n n1
For the two-dimensional two-degree-of-freedom aeroelastic > xn1 xn  t
>
> f n f n1 x n1 ;tt
model, with the dimensionless time   !

t, the preceding : 2

matrices and vectors are evaluated as xn  t2

A
x n  A

x t
n1  2 B
Qn B
Qn1 x  n1 ;tt
(8)
" # " #
1 x
!h =!
2 0
M ; K Equations (7) and (8) show that the explicit scheme needs to solve
x
r
2 0 r
2 the aerodynamic loads once, whereas the implicit scheme needs to do
    so twice at each physical time step.
1 2 Cl h=b
For the corrector step in Eq. (8), the aerodynamic loads of the
Q V ; 
 2Cm
n  1th time step (Qn1 ) are extrapolated by Qn1 and Qn :
m
V   V=!

b
1=2 ; and   Q n1  2Qn  Qn1 (9)
1 b2
Substituting Eq. (9) in Eq. (8), we obtain the following second-
An artificial neural network is a kind of nonlinear black-box order hybrid linear multistep scheme:
model. The time simulation method is usually applied to solve the 8
> x n1  xn  t 3f n  f n1   xn  t 3A
xn  A
xn1 
>
> 2 2
<  t 3B
Q  B
Q 
2 n n1
b b
>
> x  x  t
f  f n1 x n1 ; t  t
x >
:
n1 n 2 n
O  xn  t A

t
2
xn A 
x n1   2 3B
Qn  B
Qn1 
Kh (10)
h
The error analysis of the hybrid linear multistep scheme is
α E Kα documented in [36]. The superior accuracy, stability, and compu-
V G tational efficiency of the scheme are also demonstrated by time
domain flutter simulation of Isogai Wing and AGARD Wing 445.6 in
ab xα b
this reference.
The effect of the time step on the structural responses has been
z examined in our previous studies. If a cycle has 30 to 50 samples, a
Fig. 4 Aeroelastic model with 2 degrees of freedom. satisfactory accuracy can be achieved. The time step t used for
 ZHANG ET AL. 1023

0.2 0.2
0.1
α /rad CFD CFD
ROM ROM
0 0.1 0.05

Cm

Cm
0 0
-0.2
50 100 150 200
T -0.1 -0.05
A= 0.2 A= 0.1
1
-0.2 0 0.2 -0.1 0 0.1
α /rad α /rad
Cl

0
a) b)

50 100 150 200 0.02 0.005

CFD CFD
0.2 T ROM ROM
0.01
Cm

Cm

Cm
0 0 0

-0.01
Downloaded by Northwestern Polytechnic University on May 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/1.J050581

-0.2
50 100 150 200 A= 0.02 A= 0.005
T -0.02 -0.005
Fig. 5 Training signal. -0.02 0 0.02 -0.005 0 0.005
α /rad α /rad
c) d)
Fig. 8 Comparison between the phase diagrams of Cm obtained using
-20 ROM and CFD.
Power frequency (dB/Hz)

-30

-40
far more computationally efficient, because RBF neural networks
involve only algebraic equations to determine the unsteady aero-
-50 dynamic loads.
-60

-70 IV. Results and Discussion

0 1 2 3 4 To assess the validity and accuracy of the nonlinear ROM based on
k the RBF neural network, we first apply it to compute the unsteady
Fig. 6 Power spectrum density of 
T. aerodynamic loads of typical airfoil at various vibration amplitudes
and compare the results with those by the direct CFD method. Then,
we investigate the application of the nonlinear ROM to airfoil limit
ROMs must be the same as that used for training the neural network, cycle flutter problems and compare the results with those by the
because this method works in discrete time domain. direct CFD method.
The validation of the time domain simulation method with the
direct CFD solver for flutter predictions was documented in A. Computation of Unsteady Loads
[15,18,37]. The same coupled time domain simulation method A forced pitch motion of a NACA 64A010 airfoil at transonic flow
with nonlinear ROM as that of the CFD flow solver, was adopted. It is was used to assess the validity and accuracy of the preceding

0.06 Present (CFD)

CFD CFD
1 ROM Thomas et al. [28] (HB)
ROM 0.5

0.04
α /rad
Cl

Cl

0 0

0.02
A= 0.2 -0.5 A= 0.1
-1

-0.2 0 0.2 -0.1 0 0.1 0

0.55 0.6 0.65 0.7 0.75 0.8
α /rad α /rad
V∗
a) b)
a) Velocity

CFD
0.04 CFD Present (CFD)
0.06
ROM ROM Thomas et al. [28] (HB)
0.1 0.02
0.04
Cl

Cl

0
α /rad

-0.02 0.02
-0.1 A= 0.02 A= 0.005
-0.04
-0.02 0 0.02 -0.005 0 0.005
0
α /rad α /rad 0.64 0.66 0.68 0.7
c) d) ω f /ω α
Fig. 7 Comparison between the phase diagrams of Cl obtained using b) Frequency ratio
ROM and CFD. Fig. 9 Limit cycle behaviors compared with those of the method in [28].
 1024 ZHANG ET AL.

0.5 number of 0.8, at which case a strong shock is observed on the airfoil
surface. The pitching motion of the airfoil
T (T is the
dimensionless time) and direct CFD results of Cl and Cm are shown
h/b
0
in Fig. 5. The power spectrum density of
T is shown in Fig. 6.
From these figures, both large- and small-amplitude components
-0.5
were found in the training signal. Moreover, the power spectrum
0 500 1000 1500
density of
T is high, ranging from k  0 to k  1, where the
T
reduced frequency k is defined as !
b=V. The delay orders of the
a)
network are defined as m  2 and n  3. Some classical principles,
0.2 such as the error minimization, can be used to select the delay order of
the system. By varying the order of the model, one searches for a
0.1 model which minimizes the error between the CFD and predicted-
α /rad

0 model time histories. The delay orders in the current paper have been
carefully studied, acquiring a satisfactory level of accuracy while
-0.1
maintaining the stability of the model.
-0.2 Figures 7 and 8 show the nonlinear ROM predictions at reduced
0 500 1000 1500
T
frequency k  0:1635 with various vibration amplitudes. Both the Cl
and Cm computed by the nonlinear ROM agree with the direct CFD
b)
Downloaded by Northwestern Polytechnic University on May 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/1.J050581

results.
Fig. 10 Designed training signal.
B. Computation of Limit Cycle Oscillations
For the dynamic linear reduced-order modeling, the training
1 CFD signals are often designed as forced inputs, such as the impulse,
ROM
0.5 multilevel, and 3211 signals. The modeling of the nonlinear multi-
input/multi-output (MIMO) system requires a high-quality design of
Cl

0 the input signal, because of the invalidation of the superposition for a

nonlinear system. It requires not only the frequency characteristics
-0.5
but also the amplitude ratios and phase angles among the structural
-10 modes of the input signals to match those in the test. Unlike the
500 1000 1500
traditional design method of the input signals, the signals of the self-
T excited vibration of the aeroelastic system are designed as the input
a) signals in the current paper. By coupling structural equations and the
CFD solver, a typical aeroelastic response can be solved in the time
0.2 CFD
ROM
domain. The generalized structural displacements correspond to the
0.1 input vector, and the generalized aerodynamic force coefficients
correspond to the output vector.
Cm

0 Unlike the training process of the linear system, the characteristics

-0.1
of the nonlinear ROM depend on the amplitude of the training
signals. Thus, the amplitude of the initial perturbation will affect the
-0.2 quality of the ROM. The amplitude ranges of the self-excited
0 500 1000 1500
vibration signal should cover the amplitude ranges of LCO. The
T
designed self-excited vibration is a stable process at subcritical
b) velocity if the amplitude of the initial perturbation is small and
Fig. 11 Identified results for training data. unstable otherwise.
The test cases first show the identification results of unsteady
aerodynamics and then the LCO behavior for two flutter models with
nonlinear reduced-order modeling method. For this case, the pitching linear structural parameters in transonic flow.
angle is the input data and the aerodynamic loads (Cl and Cm ) are the The first aeroelastic test case is taken from [28]. The relevant
outputs. This system can be seen as a single-input multi-output parameters are the NACA64A010 airfoil, M  0:8,
0  0 deg,
system. The airfoil pitches around its 20% chord position at a Mach a  0:6, x
 0:25, ra2  0:75, !h =!
 0:5, and   75. The

0.4 CFD
CFD
0.4 ROM
ROM
0.2 0.2
h/b

h/b

0 0

-0.2 -0.2
-0.4
-0.4
0 500 1000 1500 0 200 400 600 800 1000
T T
0.1 CFD
0.05 CFD
ROM ROM
0.05
α /rad

α /rad

0 0

-0.05
-0.05
-0.1
0 500 1000 1500 0 200 400 600 800 1000
T T
a) Small amplitude of initial perturbation b) Large amplitude of initial perturbation
Fig. 12 Compared responses of LCOs at V   0:739.
 ZHANG ET AL. 1025

0.08
CFD CFD
ROM ROM
0.4
0.06

0.3

α /rad
h/b 0.04
0.2

0.02
0.1

0 0
0.65 0.7 0.75 0.8 0.65 0.7 0.75 0.8
V* V*
a) Velocity

0.08
Downloaded by Northwestern Polytechnic University on May 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/1.J050581

CFD CFD
ROM ROM
0.4
0.06

0.3

α /rad
h/b

0.04
0.2

0.02
0.1

0 0
0.64 0.66 0.68 0.7 0.72 0.64 0.66 0.68 0.7 0.72
ω f /ω α ω f /ω α
b) Frequency ratio
Fig. 13 Compared LCO behavior with M  0:8 and 0  0 deg.

second aeroelastic test case is the benchmark active controls responses become LCOs). Usually, when the amplitude of the initial
technology (BACT) model [38]. The relevant parameters are the perturbation is large, the vibration is designed as a convergent
NACA0012 airfoil,
0  0 deg, a  0, x
 0:00350, r
2  1:036, process under the subcritical speed. This speed is lower than the LCO
and !h =!
 0:6539. The second aeroelastic case has been studied onset speed, which can be calculated by the other previous methods
experimentally for flutter boundaries. (such as the flutter analysis methods based on linear ROM, such as
Figure 9 shows the LCO behavior of case 1. The results computed the ARX models used in [20]). When the amplitude of the initial
by the time domain direct CFD method are compared with those by perturbation is small, the vibration is designed as a divergent process
the harmonic balance method in [28]. The two different approaches under the supercritical speed. This speed is higher than the LCO
were found to offer a better agreement of the frequency behaviors onset speed by 10 to 20%. In this example, both the convergent and
than that of velocity behaviors. Figure 10 shows the designed training divergent processes are contained in the training signal.
signals. They are achieved by self-excited vibrations under reduced Figure 11 shows the identified results for the training data, which
velocity V   0:67 and V   0:8 (when 0 < T < 1200 and V   agree well with the direct CFD results. The same coupled time
0:67, the responses are stable; when T > 1200 and V   0:8, the domain simulation method as the ROM for different V  at a fixed

1 CFD 1.2 CFD

0.9 ROM ROM
Experiment 1.1 Experiment
0.8
1
0.7
0.6 0.9
ω f /ω α
Vf *

0.5 0.8
0.4 0.7
0.3
0.6
0.2
0.1 0.5

0 0.4
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
M M
a) Flutter velocity boundary b) Flutter frequency boundary
Fig. 14 Computed flutter boundary for the BACT model compared with experimental data.
 1026 ZHANG ET AL.

h/b
0.5
0.5

h/b
0
0
ξ

-0.5
α
-0.5
0 200 400 600 800
0 20 40 60 80 100 T
T Fig. 18 Typical response of the LCO (120 cycles, time cost of 82 h).
Fig. 15 Designed training signal (V   0:5).

Fig. 13 are determined from the visual inspection of structural

CFD responses such as those in Fig. 12. The behaviors (both the velocity
1 ROM
and frequency ratio) of LCOs computed by the nonlinear ROM
almost go through the predictions of the direct CFD method in the full
range of the amplitude.
Cl

0
Downloaded by Northwestern Polytechnic University on May 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/1.J050581

Figure 14 shows the computed flutter boundary of the BACT

model using the CFD-based ARX ROM [20]. The results almost
-1
exactly match those by the direct CFD method and are close to the
experimental data. Using the same method in the first case, we
0 20 40 60 80 100
computed the LCO behavior of the second case at M  0:8. From
T Fig. 14a, the flutter velocity index at a Mach number of 0.8 is about
a) 0.56. A convergent process under the subcritical velocity (V   0:5)
was designed at a high amplitude of initial perturbation, as shown in
CFD Fig. 15.
0.1 ROM Figure 16 shows the identified results for the training data, which
agree well with the direct CFD results for both lift and moment
coefficients. Using the nonlinear ROM, the time domain responses of
Cm

0
the airfoil for different V  values at a Mach number of 0.8 can be
simulated efficiently. The same conclusions can be made in Fig. 17.
-0.1 The LCO behaviors predicted by the nonlinear ROM maintain higher
accuracies compared with the direct CFD method.
0 20 40 60 80 100 Compared with the direct CFD method, which needs to compute
T hundreds of cycles for every velocity (dynamic pressure), the
b) nonlinear ROM method only requires a typical response solution to
Fig. 16 Identified results for training data. construct the nonlinear reduced-order aerodynamic model. When the
nonlinear ROM is prepared, the time cost of the ROM-based time
domain simulation may be neglected by comparing with those of the
direct CFD method. A typical case shows the predominance of the
Mach number was adopted. It is far more computationally efficient, nonlinear ROM method as follows. For the BACT model at
because it does not need to solve the unsteady flow. Figure 12 shows V   0:65, a stable LCO contains about 120 cycles of time domain
the responses of LCOs at V   0:739. The results of the nonlinear simulation, which needs 82 h on a personal computer (CPU:
ROM agree with those of the direct CFD. Figure 13 compares the 2.4 GHz, Memory: 1.5 GB). The computed response of the plunge
LCO behavior of the first case computed by the nonlinear ROM displacement is shown in Fig. 18. The time cost of the results of the
method and that by the direct CFD method. The LCO amplitudes in first case using the two methods is listed in Table 1.

0.8 ROM 0.8

ROM
CFD CFD
0.6 0.6
h/b

h/b

0.4 0.4

0.2 0.2

0 0
0.56 0.58 0.6 0.62 0.64 0.66 0.72 0.74 0.76 0.78
V* ω f /ω α

0.06 0.06
ROM ROM
CFD CFD
0.04 0.04
α /rad

α /rad

0.02 0.02

0 0
0.56 0.58 0.6 0.62 0.64 0.66 0.72 0.74 0.76 0.78
ω f /ω α
*
V
a) Velocity b) Frequency ratio
Fig. 17 Comparison of LCO behaviors for the BACT model.
 ZHANG ET AL. 1027

Table 1 Comparison between the time costs of direct CFD and nonlinear ROM
Method Time cost Total time
Direct CFD simulation 1) Time cost for each cycle is about 40 min; 2) there are TCFD  10 100 40 min 40; 000
100 cycles for each velocity; and 3) 10 response computations min 666:7 h
at 10 different velocities are needed to get the LCO behavior.
Nonlinear ROM 1) Time cost for train signal computation is about TROM  40 20  2:17 20  843
40 20 min (about 20 cycles); 2) time cost of the time domain min 14 h
simulation of 100 cycles is about 2.17 min; and 3) responses were
computed at 20 different velocities.

V. Conclusions doi:10.2514/1.21051
[11] Bendiksen, O., “High-Altitude Limit Cycle Flutter of Transonic
Based on the preceding comparison, the following conclusions Wings,” Journal of Aircraft, Vol. 46, No. 1, 2009, pp. 123–136.
can be obtained: doi:10.2514/1.36413
1) The nonlinear ROM constructed by the responses at typical [12] Vio, G. A., Dimitriadis, G., and Cooper, J. E., “Bifurcation Analysis and
velocity can be used for the dynamic analysis of the aeroelastic Limit Cycle Oscillation Amplitude Prediction Methods Applied to the
system at other velocities at a fixed Mach number.
Downloaded by Northwestern Polytechnic University on May 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/1.J050581

Aeroelastic Galloping Problem,” Journal of Fluids and Structures,

2) The behaviors of the LCOs, as well as the responses of the Vol. 23, No. 7, Oct. 2007, pp. 983–1011.
structures calculated by the nonlinear ROM method, agree well with doi:10.1016/j.jfluidstructs.2007.03.006
those obtained by the direct CFD method. [13] Dowell, E. H., and Hall, K. C., “Modeling of Fluid Structure
Interaction,” Annual Review of Fluid Mechanics, Vol. 33, No. 1, 2001,
3) The computational efficiency of the nonlinear ROM method is pp. 445–490.
improved by one to 2 orders of magnitude compared with that of the doi:10.1146/annurev.fluid.33.1.445
direct CFD method. [14] Lucia, D. J., Beran, P. S., and Silva, W. A., “Reduced-Order Modeling:
New Approaches for Computational Physics,” Progress in Aerospace
Sciences, Vol. 40, No. 1, 2004, pp. 51–117.
Acknowledgments doi:10.1016/j.paerosci.2003.12.001
This work was supported by the National Natural Science [15] Zhang, W., and Ye, Z., “On Unsteady Aerodynamic Modeling Based on
Foundation (nos. 10802063 and 11172237), Aeronautics Science CFD Technique and its Applications on Aeroelastic Analysis,”
Foundation (no. 2009ZA53009), Fundamental Research Foundation Advances in Mechanics, Vol. 38, No. 1, 2008, pp. 77–86.
[16] Zhang, W., and Ye, Z., “Effect of Control Surface on Airfoil Flutter in
of Northwestern Polytechnical University, and Aoxiang Star Transonic Flow,” Acta Astronautica, Vol. 66, Nos. 7–8, 2010, pp. 999–
Foundation. The author would also like to thank Lingcheng Zhao and 1007.
Yongnian Yang. doi:10.1016/j.actaastro.2009.09.016
[17] Zhang, W., and Ye, Z., “Reduced-Order-Model-Based Flutter Analysis
at High Angle of Attack,” Journal of Aircraft, Vol. 44, No. 6, 2007,
References pp. 2086–2089.
[1] Dowell, E. H., Edwards, J., and Strganac, T., “Nonlinear doi:10.2514/1.32285
Aeroelasticity,” Journal of Aircraft, Vol. 40, No. 5, 2003, pp. 857–874. [18] Zhang, W., Ye, Z., Yang, Q., and Shi, A., “Gust Response Analysis
[2] Lee, B. H. K., Price, S. J., and Wong, Y. S., “Nonlinear Aeroelastic Using CFD-Based Reduced Order Models,” 47th AIAA Aerospace
Analysis of Airfoils: Bifurcation and Chaos,” Progress in Aerospace Sciences Meeting Including the New Horizons Forum and Aerospace
Sciences, Vol. 35, No. 3, 1999, pp. 205–334. Exposition, AIAA Paper 2009-895, Orlando, FL, 5–8 Jan. 2009.
doi:10.1016/S0376-0421(98)00015-3 [19] Zhang, W., Ye, Z., and Zhang, C., “ROM Based Aeroservoelastic
[3] Henshaw, M. J., Badcock, K. J., Vio, G. A., Allen, C. B., and Analysis in Transonic Flow,” Journal of Aircraft, Vol. 46, No. 6, 2009,
Chamberlain, J., Kaynes, I., Dimitriadis, G., Cooper, J. E., Woodgate, pp. 2178–2183.
M. A., Rampurawala, A. M., Jones, D., Fenwick, C., Gaitonde, A. L., doi:10.2514/1.45249
Taylor, N. V., Amor, D. S., Eccles, T. A., and Denley, C. J., “Non-Linear [20] Zhang, W., and Ye, Z., “Control Law Design for Transonic
Aeroelastic Prediction for Aircraft Applications,” Progress in Aeroservoelasticity,” Aerospace Science and Technology, Vol. 11,
Aerospace Sciences, Vol. 43, Nos. 4–6, May–Aug. 2007, pp. 65–137. Nos. 2–3, 2007, pp. 136–145.
doi:10.1016/j.paerosci.2007.05.002 doi:10.1016/j.ast.2006.12.004
[4] Zhang, Z., Yang, S., and Liu, F., “Prediction of Flutter And LCO by an [21] Dowell, E. H., Thomas, J. P., and Hall, K. C., “Transonic Limit Cycle
Euler Method on Non-Moving Cartesian Grids with Boundary-Layer Oscillation Analysis Using Reduced Order Aerodynamic Models,”
Corrections,” AIAA Paper 2005-833, 2005. Journal of Fluids and Structures, Vol. 19, No. 1, 2004, pp. 17–27.
[5] Bendiksen, O., “Transonic Limit Cycle Flutter of High-Aspect-Ratio doi:10.1016/j.jfluidstructs.2003.07.018
Swept Wings,” Journal of Aircraft, Vol. 45, No. 5, 2008, pp. 1522– [22] Beran, P. S., Lucia, D. J., and Pettit, C. L., “Reduced-Order Modelling
1533. of Limit-Cycle Oscillation for Aeroelastic Systems,” Journal of Fluids
doi:10.2514/1.29547 and Structures, Vol. 19, No. 5, June 2004, pp. 575–590.
[6] Patil, M. J., Hodges, D. H., and Cesnik, C. E. S., “Limit-Cycle doi:10.1016/j.jfluidstructs.2004.04.002
Oscillations in High-Aspect-Ration Wings,” Journal of Fluids and [23] Jones, D. P., Roberts, I., and Gaitonde, A. L., “Identification of Limit
Structures, Vol. 15, No. 1, 2001, pp. 107–132. Cycles for Piecewise Nonlinear Aeroelastic Systems,” Journal of Fluids
doi:10.1006/jfls.2000.0329 and Structures, Vol. 23, No. 7, Oct. 2007, pp. 1012–1028.
[7] Tang, D. M., and Dowell, E. H., “Experimental and Theoretical Study doi:10.1016/j.jfluidstructs.2007.03.007
on Flutter and Limit-Cycle Oscillations of High-Aspect-Ratio Wings,” [24] Marzocca, P., Silva, W. A., and Librescu, L., “Open/Closed-Loop
AIAA Journal, Vol. 39, No. 8, 2001, pp. 1430–1441. Nonlinear Aeroelasticity for Airfoils via Volterra Series Approach,”
doi:10.2514/2.1484 AIAA Journal, Vol. 42, No. 4, 2004, pp. 673–686.
[8] Attar, P., Dowell, E. H., and White, J., “Modeling the LCO of a Delta doi:10.2514/1.9552
Wing Using a High Fidelity Structural Model,” AIAA Paper 2004- [25] Raveh, D. E., “Reduced-Order Models for Nonlinear Unsteady
1692, 2004. Aerodynamics,” AIAA Journal, Vol. 39, No. 8, 2001, pp. 1417–1429.
[9] Tang, D. M., Kerry, J. K., and Dowell, E. H., “Limit Cycle Oscillations doi:10.2514/2.1473
of Delta Wing Models in Low Subsonic Flow,” AIAA Journal, Vol. 37, [26] Silva, W., “Identification of Nonlinear Aeroelastic Systems Based on
No. 11, 1999, pp. 1355–1362. the Volterra Theory: Progress and Opportunity,” Nonlinear Dynamics,
doi:10.2514/2.627 Vol. 39, Nos. 1–2, 2005, pp. 25–62.
[10] Parker, G. H., Maple, R. C., and Beran, P. S., “Computational doi:10.1007/s11071-005-1907-z
Aeroelastic Analysis of Store-Induced Limit-Cycle Oscillation,” [27] Balajewicz, M., Nitzsche, F., and Feszty, D., “Application of Multi-
Journal of Aircraft, Vol. 44, No. 1, Jan.–Feb. 2007, pp. 48–59. Input Volterra Theory to Nonlinear Multi-Degree-of-Freedom
 1028 ZHANG ET AL.

Aerodynamic Systems,” AIAA Journal, Vol. 48, No. 1, Jan. 2010, “A Surrogate Based Approach to Reduced-Order Dynamic Stall
pp. 56–62. Modeling,” AIAA Paper 2010-3042, 2010.
doi:10.2514/1.38964 [33] Haykin, S., Neural Networks: A Comprehensive Foundation, 2nd ed.,
[28] Thomas, J. P., Dowell, E. H., and Hall, K. C., “Nonlinear Inviscid Prentice–Hall, Upper Saddle River, NJ, 1999.
Aerodynamic Effects on Transonic Divergence, Flutter, and Limit- [34] Sundararajan, N., and Saratchandran, P., Radial Basis Function Neural
Cycle Oscillations,” AIAA Journal, Vol. 40, No. 4, 2002, pp. 638–646. Networks with Sequential Learning, World Scientific, London, 1999.
doi:10.2514/2.1720 [35] Buhmann, M. D., Radial Basis Functions: Theory and Implementa-
[29] Thomas, J. P., Dowell, E. H., and Hall, K. C., “Using Automatic tions, Cambridge Univ. Press, Cambridge, England, U.K., 2003.
Differentiation to Create a Nonlinear Reduced-Order-Model [36] Zhang, W., Jiang, Y., and Ye, Z., “Two Better Loosely Coupled Solution
Aerodynamic Solver,” AIAA Journal, Vol. 48, No. 1, Jan. 2010, Algorithms of CFD Based Aeroelastic Simulation,” Engineering
pp. 19–24. Applications of Computational Fluid Mechanics, Vol. 1, No. 4, 2007,
doi:10.2514/1.36414 pp. 253–262.
[30] Faller, W. E., and Schreck, S. J., “Unsteady Fluid Mechanics [37] Zhang, W., “Efficient Analysis for Aeroelasticity Based on CFD,” Ph.D.
Applications of Neural Networks,” Journal of Aircraft, Vol. 34, No. 1, Dissertation, Northwestern Polytechnical Univ., Xi’an, China, 2006.
1997, pp. 48–55. [38] Rivera, J. A., Jr., Dansberry, B. E., Bennett, R. M., Durham, M. H., and
doi:10.2514/2.2134 Silva, W. A., “NACA0012 Benchmark Model Experimental Flutter
[31] Marques, F. D., and Anderson, J., “Identification and Prediction of Results with Unsteady Pressure Distributions,” NASA TM 107581,
Unsteady Transonic Aerodynamic Loads by Multi-Layer Functionals,” 1992.
Journal of Fluids and Structures, Vol. 15, No. 1, 2001, pp. 83–106.
doi:10.1006/jfls.2000.0321 B. Epureanu
Downloaded by Northwestern Polytechnic University on May 23, 2013 | http://arc.aiaa.org | DOI: 10.2514/1.J050581

[32] Glaz, B., Liu, L., Friedmann, P. P., Bain, J., and Sankar, L. N., Associate Editor

View publication stats