Vol.2 Issue.

4,
INTERNATIONAL JOURNAL OF RESEARCH IN AERONAUTICAL AND MECHANICAL ENGINEERING April 2014.
Pgs: 127-137

ISSN (ONLINE): 2321-3051

INTERNATIONAL JOURNAL OF RESEARCH IN

AERONAUTICAL AND MECHANICAL ENGINEERING
Flutter prediction using neural networks-based hybrid optimization
scheme

Prabhu L1, Dr J Srinivas2

1
Graduate student, Department of Mechanical Engineering, NIT Rourkela-769008,
[email protected]
2
Associate Professor, Department of Mechanical Engineering, NIT Rourkela-769008,
[email protected]
Abstract

In this paper, an unsteady aerodynamic model is used to study the influence of geometrical parameters on the
flutter velocity of an airfoil section. The most affecting geometric variable on flutter speed is predicted using
theory of ANOVA. A single hidden layer back-propagation neural network model is trained with different sets
of geometric parameters and flutter velocities. The optimal parameters of the section corresponding to a known
flutter velocity are obtained through the trained neural network in combination with hybrid optimization
scheme formulated based on micro genetic algorithm (µGA) with micro differential evolution (µDE). Results
are presented with a benchmark airfoil geometry.

Keywords: Airfoil, Flutter, Artificial neural network, Hybrid optimization, Aeroelasticity

1. Introduction
The study of the effect of aerodynamic forces over elastic bodies is known as aeroelasticity. The aeroelastic
problems are different from theory of elasticity. The theory of elasticity is concerned about the stress and
deformation of an elastic body subjected to external loading, which is independent of deformation of the body.
In aeroelasticity problems the aerodynamic loading purely depends on posture of the body relative to flow;
unless the value of elastic deformation is not known the aerodynamic loading cannot be predicted. Aeroelastic
problems are solved mainly to get the stability regions. The flutter of an airfoil is a dynamic aeroelastic
instability which involves the interactions of aerodynamic, elastic and inertia forces. When an airfoil is
subjected to the air flow whose speed increase gradually, the rate of damping of the oscillation of airfoil
initially increases and then rapidly falls down. At particular speed where, the oscillation of airfoil could be
maintained with steady amplitudes is known as critical flutter speed. Even for small disturbances of airfoil
above the critical speed leads to violent oscillation and this state is known as flutter. The effect of airfoil
parameters on flutter speed were empirically given by Theodorsen and Garrick [1] in early 1940s.
In order to predict and control the flutter conditions, different methods of aerodynamic and structural
modelling have been used in practice. The finite element modelling is most commonly used approach [2].

Prabhu L, Dr J Srinivas 127

 Vol.2 Issue.4,
INTERNATIONAL JOURNAL OF RESEARCH IN AERONAUTICAL AND MECHANICAL ENGINEERING April 2014.
Pgs: 127-137

While predicting flutter in subsonic wing, Moosavi et al. [3] employed a procedure based on Galerkin’s
method to determine the critical speed with and without considering compressibility effect using quasi-steady
aerodynamic model. Haddapour et al. [4] compared the aeroelastic behaviour of wing structure using quasi-
steady aerodynamic model with that of unsteady aerodynamic model to know the best one for analysis. Riccy
[5] presented a numerical study of flutter by using Leishman’s state space model under unsteady aerodynamic
loads to analyze the stability region and the effectiveness was verified by comparing with Theodorsen and
Garricks’ approach. Brain et al. [6] discussed various tools to examine the aeroelastic stability of wind turbine
blade and suggested a design tool. Park et al. [7] utilized such unstable oscillations and converted them as the
useful energy in the form of an electromagnetic energy harvester. In more recent works, frequency-domain
identification for prediction of flutter in cantilevered wing [8] and experimental study of flutter behavior of
flexible airfoils supported with laminated composite plates [9] were also illustrated in this line.
In order to understand primarily the flutter concept, an implicit model relating the object geometry with
flutter velocity is required. Neural networks and optimization techniques are being recently used in several
applications of aerodynamics. The critical flutter load and boundary conditions of tapered beam were obtained
by Takahashi [10] by multilayer perceptron network trained with the back propagation algorithm. Chen et al.
[11] used two different back-propagation neural networks to identify the flutter derivatives. Mannarino and
Mntegazza [12] extended ANN application to find the aeroelastic limit cycle oscillations. Use of neural
networks for function approximation and utilize this further for optimizing the outputs is a usual task. A hybrid
algorithm holds the best properties of various optimization techniques. Recently, several works proposed
hybrid optimization schemes [13-16].
In this paper first part deals with identification of flutter speed using unsteady aerodynamic model and the
influence of aeroelastic parameters on flutter speed is analyzed. In the second part a back-propagation neural
network is trained with aeroelastic parameters and flutter speed and the weights are acquired to identify the
relationship between flutter speed and airfoil parameters. In the last part, the parameters at a desired flutter
speed are obtained by using hybrid optimization technique based on micro genetic algorithms and differential
evolution schemes. The content of the paper is organized as follows: section 2 presents related equations of
motion including lift and pitching moment expressions. Section 3 deals with parametric studies and section 4
explains training of neural network and hybrid optimization methodology along with results and discussions.

2. Equations of motion
A two-degree of freedom of airfoil is considered as shown in Fig.1.

Kh

α Kα

Fig. 1: Two degree of freedom model

Prabhu L, Dr J Srinivas 128

 Vol.2 Issue.4,
INTERNATIONAL JOURNAL OF RESEARCH IN AERONAUTICAL AND MECHANICAL ENGINEERING April 2014.
Pgs: 127-137

Here, h is plunging translation and α is a pitching rotation about elastic axis. It is subjected to a lift force L and
a pitching moment Mα. Here, E is elastic axis, M is mid chord, and c.g is centre of mass. The equations of
motion are written as:
.. ..
m h + S α + hK h = − L (1)
.. ..
S h+ I α α + αK α = M α (2)
where m, S, Iα, Kh and Kα are mass, static moment, mass moment of inertia, plunge coefficient and pitch
coefficient respectively.

2.1. Aerodynamic model

In the unsteady aerodynamics, the lift and moment can be divided into two parts; one is apparent mass which is
not associated with creation of vorticity and other is circulatory part, purely depends on the motion and
influence of wake. The lift and moment expression for a sinusoidal oscillating airfoil is given by Theodorsen’s
unsteady aerodynamics as:

h  1  
L = πρ b 3 ω 2  Lh + α  Lα −  + a  Lh   (3)
 b   2  
 1   h  1   1 
2  
M α = πρ b 4 ω 2   M h −  + a  L h  +  M α −  + a  (L α + M h ) +   L h α  (4)

 2   b  2  2+a  

where
1
L h = 1 − i 2C ( k ) (5i)
k
1 1 + 2C ( k ) 2C ( k )
Lα = − i − (5ii)
2 k k2
1 3 1
Mh = ,Mα = − i (5iii)
2 8 k

ωb
Here i denotes the imaginary part of complex number and k is the reduced frequency k = with V as flow
V
velocity and b as semi-chord length. By considering harmonic motion of airfoil, Eqs. (1-4) can be simplified
and written in matrix form as a characteristic equation as follows:

   2 2  h 
µ 1 −  ω h   ωα   + Lh µx + L − L
1
h + a

  
   ωα   ω   α α
2   b 
     = 0 (6)
 1    ω 2  1  1   
2
 µxα + M h − Lh  + a  µrα 1 − 
2 α
  + M α − (Lα + M h ) + a  + Lh ( + a    
2  ω   2  2    α 
   
This has non-trivial roots if determinant is zero and it results in two sets of complex conjugate roots. By
varying the flow velocity V over a range and obtaining different roots, we can plot real and imaginary terms.
The intersection these terms occurs at a particular value of V (corresponding to 1/k) and is theoretically given
by [17]:
bω α
U cr = (7)
k X

Prabhu L, Dr J Srinivas 129

 Vol.2 Issue.4,
INTERNATIONAL JOURNAL OF RESEARCH IN AERONAUTICAL AND MECHANICAL ENGINEERING April 2014.
Pgs: 127-137

3. Influence of geometric parameters on flutter velocity

3.1. Conventional method

The flutter velocity is strongly affected by distance between mid-chord and elastic axis (ah), distance between
elastic axis and center of mass (xα) and bending torsion frequency ratio (ωh/ωα) as well as many other
parameters. Table-1 shows the fixed parameters of airfoil employed in present work.

Table: 1 Aeroelastic parameters

Parameters Value
µ 76
ah -0.15
xα 0.25
rα2 0.388
b 0.127 m
ωα 64.1 rad/s
ωh 55.9 rad/s

In order to find the influence of each parameter on the flutter speed, three levels are considered as depicted
in Table 2. By considering one of the parameters in level 1 as constant and combining with other parameters in
all level results in 9 combinations, by doing in similar fashion it gives 27 combinations. Other parameters for
aeroelastic analysis are considered as constant.

Table: 2 Aeroelastic parameter levels

Level ah xα ωh/ωα
1 -0.2 0.1 0.4
2 -0.175 0.2 0.6
3 -0.15 0.25 0.87

Keeping ah as constant, flutter velocities are obtained as shown in Fig 2. From Fig 2 it is clear that flutter
speed is almost equal with very small variation for different ah value (-0.2, -0.175, and -0.15) with different xα
and ωh/ωα combinations. Increase in xα and ωh/ωα leads to an increase in flutter velocity. When ωh/ωα is 0.4 and
0.6 the flutter velocity decreases with corresponding to increase in xα but when ωh/ωα is 0.87 the flutter velocity
increases with corresponding to increase in xα.

Fig. 2: Variation of flutter velocity with ah as constant

Prabhu L, Dr J Srinivas 130

 Vol.2 Issue.4,
INTERNATIONAL JOURNAL OF RESEARCH IN AERONAUTICAL AND MECHANICAL ENGINEERING April 2014.
Pgs: 127-137

Now xα is taken as constant, then flutter velocity is plotted against ah and ωh/ωα in Fig 3. Flutter velocity is
varying with different xα values (0.1, 0.2, and 0.25) can be seen from Fig 3. Increase in ah (-0.15 to -0.2) and
ωh/ωα gives increased value of flutter velocity. At xα is 0.1, ωh/ωα is 0.4 and for various ah the flutter velocity is
high and it is low when ωh/ωα is 0.87 comparing to xα (0.2 and 0.25). When ωh/ωα is 0.6 the combination of
various ah and xα, the flutter velocities are with small difference.

Fig. 3: Variation of flutter velocity with xα as constant

At last ωh/ωα is taken as constant with various values and variation of flutter velocity is plotted in Fig 4.
The increase in ωh/ωα results in decrease of flutter velocity. Decrease and increase of xα and ah (-0.15 to -0.2)
leads to increase in flutter velocity. The combination of least values of ωh/ωα, xα and ah gives the high flutter
velocity.

Fig. 4: Variation of flutter velocity with ωh/ωα as constant

3.2. Design of Experiments

In order to minimize the number of experiments, the method of design of experiments (DOE) is employed. The
first step before using the DOE is to identify number of factors and levels. Once it is done orthogonal array is
used to know the possible number of experiments. Taguchi method is employed for three level and three
parameters given in Table 2, Taguchi L9 (3*3) orthogonal array is used to predict the influence of parameters
such as distance between mid-chord and elastic axis (ah), distance between elastic axis and center of mass (xα)
and bending torsion frequency ratio (ωh/ωα) on the output factor namely flutter speed (Ucr). The experimental

Prabhu L, Dr J Srinivas 131

 Vol.2 Issue.4,
INTERNATIONAL JOURNAL OF RESEARCH IN AERONAUTICAL AND MECHANICAL ENGINEERING April 2014.
Pgs: 127-137

layout of an L9 orthogonal array is shown in Table 3. As the essential objective is to maximize the flutter
velocity, one has to predict the best level of each factor. Minitab 16 is used to carry out the Taguchi method;
flutter speed of 9 combinations is predicted by using MATLAB code and fed into Minitab to get the Means
graphs for analysis.

Table: 3: L9 Orthogonal array

Level ah xα ωh/ωα Vf
1 -0.2 0.1 0.4 85.31
2 -0.2 0.2 0.6 43.74
3 -0.2 0.25 0.87 27.83
4 -0.175 0.1 0.6 45.84
5 -0.175 0.2 0.87 25.19
6 -0.175 0.25 0.4 74.12
7 -0.15 0.1 0.87 19.72
8 -0.15 0.2 0.4 75.02
9 -0.15 0.25 0.6 42.37

Fig. 5 shows the variation of parameters against mean flutter speed. The parameter with higher value of
Mean has very good influence over flutter speed. It graphically compares the level of an output variable for
different input parameters, to gain an understanding of the effect of a parameter on the output. A relatively flat
line shows it has little effect upon the output, while a line that has a lot of up and down movement indicates
that as the factor changes; it has a greater mean effect. From Fig 5, it is observed that the parameter highly
influencing the flutter velocity is frequency ratio ωh/ωα and other two have less influence. In the case of flutter
speed (Ucr), the optimum parameters are ah =-0.2, xα= 0.10 and frequency ratio ωh/ωα=0.40 as shown in Fig. 5,
this same as conventional one.

Main Effects Plot for Means

Data Means

ah xalpha
80

60

40
Mean of Means

20
-0.200 -0.175 -0.150 0.10 0.20 0.25
freq ratio
80

60

40

20
0.40 0.60 0.87

Fig. 5: Effect of parameters on flutter velocity

Prabhu L, Dr J Srinivas 132

 Vol.2 Issue.4,
INTERNATIONAL JOURNAL OF RESEARCH IN AERONAUTICAL AND MECHANICAL ENGINEERING April 2014.
Pgs: 127-137

The aeroelastic system is statistically analyzed by Analysis of Variance (ANOVA). The output of ANOVA
from Minitab is given in Table 4 and it also shows the frequency ratio has high influence (97.43%) compared
to others.

Table: 4 ANOVA for aeroelastic analysis

Analysis of Variance for Means
Source DF Seq SS Adj SS Adj MS F P
ah 2 65.90 65.90 32.95 1.58 0.387
xalpha 2 10.10 10.10 5.05 0.24 0.805
freq ratio 2 4462.47 4462.47 2231.23 107.19 0.009
Residual Error 2 41.63 41.63 20.82
Total 8 4580.10

4. Artificial neural networks

The feed-forward back propagation (BP) network is a very popular model in neural networks and the
architecture is shown in Fig. 6. As the name implies the computations are passed forward from the input to
output layer, calculated errors are propagated back order to update the weights. Twenty seven training sets
from three level parameters shown in Table 2 are used to train the 3-layer neural network with 4 neurons in
hidden layer. Here ah, xα, and ωh/ωα are given as inputs and Ucr as output parameter. In-house code is written in
MATLAB to train the network and to acquire the weights. Convergence pattern of error is shown in Fig. 7 and
weights are shown in Table 5.

Fig. 6: Neural network structure

Prabhu L, Dr J Srinivas 133

 Vol.2 Issue.4,
INTERNATIONAL JOURNAL OF RESEARCH IN AERONAUTICAL AND MECHANICAL ENGINEERING April 2014.
Pgs: 127-137

Fig. 7: Convergence of error

Table: 5 Weights
Weights
Between input and hidden layer -5.6978 0.0774 -1.5806 -1.9556
-9.5137 -0.6191 -3.2522 -4.7738
-5.4921 -4.8774 -4.9607 -0.5389
Between hidden and output layer 20.9597 12.7619 14.8446 -4.8813

4.1. Hybrid-optimization

4.1.1. Genetic algorithms

GA is a stochastic optimization technique based on principle of ‘survival of the fittest’. GA has perfect balance
between the robustness and the efficiency of survival in different environment. Applications of GA are
increasing day by day in many engineering fields. The principle behind GA is Darwin’s theory of evolution.
The population members are strings similar to chromosomes, which are usually in binary representation of
solution vectors. At the beginning, initial population is evaluated for fitness value. Starting with this
information, the new generation of the population can be deduced using selection, crossover and mutation
operators. This process is iterated up to convergence. Every gene (sub-string) in the chromosome consists of
fixed number of binary digits and represents a single value.
Micro GA (µGA) operates on a population similar to GA as explained above. Unlike the GA, the
mechanics of µGA allows for a very small population size with crossover alone. Though the population is very
small the chances of becoming stagnant is high, once algorithm enters this situation reinitialization is the only
solution to take back the algorithm in right way.

4.1.2. Hybrid µGA with µDE

In this paper a hybrid µGA is proposed which uses µDE for aeroelastic optimization studies. In this technique
population initialization takes place randomly, the fitness values are evaluated and the ranking will be given,
and new set of population is created by crossover. The new set of population is evaluated and if stagnant
occurs, µDE is activated and it will be on till the best population is achieved comparing to population from
µGA or twenty cycles, whichever is earlier. Once µDE is completed the improved populations are then passed

Prabhu L, Dr J Srinivas 134

 Vol.2 Issue.4,
INTERNATIONAL JOURNAL OF RESEARCH IN AERONAUTICAL AND MECHANICAL ENGINEERING April 2014.
Pgs: 127-137

to crossover. Crossover value taken is 0.99 and 0.05 for µGA and µDE correspondingly and mutation value for
µDE is 0.8. Flowchart for hybrid µGA with µDE is shown in Fig. 8.

Fig. 8: Flowchart of hybrid µGA with µDE

The above said hybrid optimization technique is used to predict the influencing parameters for a flutter
speed of 60 m/s. Each population consist of three parameters namely ah, xα, and ωh/ωα which are randomly
initialized and the output Ucr is evaluated by the weights acquired through the neural network training. The
fitness value of each population is obtained by:

Fitnessvalue = (U T arg et − U Optim (a h , xα , ωh / ωα ))2 (8)

The trends followed by the various optimization techniques are shown in Fig. 9 and the influencing parameters
by optimization technique using ANN is shown along with the evaluation of parameters by weights and
MATLAB code in Table 6.

Prabhu L, Dr J Srinivas 135

 Vol.2 Issue.4,
INTERNATIONAL JOURNAL OF RESEARCH IN AERONAUTICAL AND MECHANICAL ENGINEERING April 2014.
Pgs: 127-137

Fig. 9: Comparison of optimization techniques

Table: 6 Parameters by optimization

Parameters by Flutter velocity
optimization with
ANN ANN-Weights MATLAB code

ah=-0.1062 59.98 57.50

xα=0.2661
ωh/ωα=0.4650

5. Conclusion

This paper presented a parametric study based on unsteady aerodynamic model and various influencing
parameters on flutter speed by Taguchi and ANOVA and the prediction is compared with conventional
approach. It is observed that the frequency ratio is a highly influencing parameter. A 3-4-1 back-propagation
neural network model was trained with twenty seven training sets to obtain the flutter velocities at unknown
geometric data. Hybrid of µGA with µDE optimization tool was employed and it shows better performance
comparing to µGA and µDE.

References
[1]. Theodorsen T, Garrick I E, 1940, Mechanism of flutter a theoretical and experimental investigation of
the flutter problem, NACA report 685.
[2]. Mushim J J, Shokat A T, Tariq S T, 2012, Flutter speed limits of subsonic wings, J Eng, 18, 2, 163-
183.
[3]. Moosavi M R, Naddaf Oskouei A, Khelli A, 2005, Flutter of subsonic wing, Thin wall Struct, 43, 4,
617-627.
[4]. Haddadpour H, Firouz –Abadi R D, 2006, Evaluation of quasi-steady aerodynamic modelling for
flutter prediction of aircraft wings in incompressible flow, Thin wall Struct, 44, 9, 931-936.
[5]. Riccy K, 2013, Numerical study of flutter of a two dimensional aeroelastic system, In Proc.World
congress on engineering, London, U.K.
[6]. Brian O C, Griffith D T, Brian R R, John E H, 2013, Impact of modelling approach on flutter
predictions for very large wind turbine blade designs, In: Proceedings AHS 69th annual forum,
phoenix., Arizona.

Prabhu L, Dr J Srinivas 136

 Vol.2 Issue.4,
INTERNATIONAL JOURNAL OF RESEARCH IN AERONAUTICAL AND MECHANICAL ENGINEERING April 2014.
Pgs: 127-137

[7]. Park J, Morgenthal G, Kim K, Kwon S, Kincho L, Power evaluation for flutter-based electromagnetic
energy harvester using CFD simulations, J Intell Mat Syst Str (In press) doi:
10.1177/1045389X14526954.
[8]. B.Gjerek, R.Drazumeric, F.Kosel, Flutter behavior of a flexible airfoil: Multiparameter experimental
study, Aerosp Sci Technol, http://dx.doi.org/10.1016/j.ast.2014.04.002
[9]. J.Ertveldt, J.Lataire, R.Pintelon, S.Vanlanduit, Frequency-domain identification of time-varying
systems for analysis and prediction of aeroelastic flutter, Mech Syst Signal Pr,
http://dx.doi.org/10.1016/j.ymssp.2013.08.020.
[10]. Takahashi I, 1999, Identification for critical flutter load and Boundary conditions of a beam using
neural network, J Sound Vib, 228, 4, 857-870.
[11]. Chen C H, Wu J C, Chen J H, 2008, Prediction of flutter derivatives by artificial neural networks, J
Wind Eng Ind Aerod, 96, 925-1937.
[12]. Mannarino A, Mantegazza P, Nonlinear aeroelastic reduced order modeling by recurrent neural
networks. J Fluid Struct (In Press), http://dx.doi.org/10.1016/j.jfluidstructs.2014.02.016i
[13]. Santana-Quintero L V, Hernandez-Diaz A G, Molina J, Coello C A, Caballero R, 2010, DEMORS: A
hybrid multi-objective optimization algorithm using differential evolution and rough set theory for
constrained problems, Comput Oper Res, 37, 470-480.
[14]. Zammer A, Mizra S M, Mizra N M, 2014, Core loading pattern optimization of a typical two-loop
300 MWe PWR using Simulated Annealing (SA) novel crossover Genetic Algorithms (GA) and
hybrid GA(SA) schemes, Ann Nucl Energy, 65, 122-131.
[15]. Wang Y, 2014, Hybrid genetic algorithm with two local optimization strategies for travelling
salesman problem, Comput Ind Eng, 14, 124-133.
[16]. Yuan X, Zhao J, Yang Y, Wang Y, 2014, Hybrid parallel chaos optimization algorithm with harmony
search algorithm, Appl Soft Comput,17, 12-22.
[17]. Fung Y C, 1993, An introduction to the theory of Aeroelasticity, Dover publications. Inc, New York,
Chap.6.

Prabhu L, Dr J Srinivas 137