Research Article

Advances in Mechanical Engineering

2021, Vol. 13(11) 1–15
Ó The Author(s) 2021
Flutter speed prediction by using deep DOI: 10.1177/16878140211062275
journals.sagepub.com/home/ade
learning

Yi-Ren Wang and Yi-Jyun Wang

Abstract
Deep learning technology has been widely used in various field in recent years. This study intends to use deep learning
algorithms to analyze the aeroelastic phenomenon and compare the differences between Deep Neural Network (DNN)
and Long Short-term Memory (LSTM) applied on the flutter speed prediction. In this present work, DNN and LSTM are
used to address complex aeroelastic systems by superimposing multi-layer Artificial Neural Network. Under such an
architecture, the neurons in neural network can extract features from various flight data. Instead of time-consuming
high-fidelity computational fluid dynamics (CFD) method, this study uses the K method to build the aeroelastic flutter
speed big data for different flight conditions. The flutter speeds for various flight conditions are predicted by the deep
learning methods and verified by the K method. The detailed physical meaning of aerodynamics and aeroelasticity of the
prediction results are studied. The LSTM model has a cyclic architecture, which enables it to store information and
update it with the latest information at the same time. Although the training of the model is more time-consuming than
DNN, this method can increase the memory space. The results of this work show that the LSTM model established in
this study can provide more accurate flutter speed prediction than the DNN algorithm.

Keywords
Flutter analysis, deep learning, deep neural network, long short-term memory

Date received: 15 August 2021; accepted: 1 November 2021

Handling Editor: James Baldwin

Introduction Bisplinghoff et al.2 not only studied flutter phenomena

and explained the physical meaning of flutter, but also
The aeroelastic flutter is a dynamic instability of a made great contributions to flutter analysis, prediction
flight vehicle associated with the interaction of aerody- of structural divergence and control of aeroelastic
namic, elastic, and inertial forces. The flutter phenom- structures. Based on the complexity of aeroelastic cou-
enon is a self-excited destructive oscillatory instability. pling, Pitt and Haudrich3 used the artificial neural net-
As the aerodynamic force exerted on the flexible body work to analyze various flutter phenomena. Due to the
is coupled with its natural vibration mode, the vibra- difficulty in establishing computer instruction cycles
tion amplitude increases. Flutter speed is the critical and data at that time, the accuracy was not high, but it
speed when the flutter just occurs. Nearly 80% of aero- has inspired the follow-up research direction. In recent
elastic flutter analysis in the industry is based on classi-
cal flutter analysis (CFA).1 The objective of the CFA is
to determine the flight conditions that correspond to Department of Aerospace Engineering, Tamkang University, New Taipei
the flutter boundary. The flutter boundary corresponds City, Taiwan
to the conditions for which the aircraft is sustaining a
Corresponding author:
simple harmonic motion. In the past decades, among Yi-Ren Wang, Department of Aerospace Engineering, Tamkang
many methods for analyzing flutter phenomena, P, K, University, 151, Ying-Chuang Road, New Taipei City 25137, Taiwan.
and P-K methods are more practical and well known.1 Email: [email protected]

Creative Commons CC BY: This article is distributed under the terms of the Creative Commons Attribution 4.0 License
(https://creativecommons.org/licenses/by/4.0/) which permits any use, reproduction and distribution of the work
without further permission provided the original work is attributed as specified on the SAGE and Open Access pages
(https://us.sagepub.com/en-us/nam/open-access-at-sage).
2 Advances in Mechanical Engineering

years, machine learning has developed vigorously. For Hochreiter and Schmidhuber14 improved the recur-
example, Google recently defeated top Go players with rent neural network (RNN) algorithm and proposed a
AlphaGo4 artificial intelligence Go software. Besides, long short term memory (LSTM) model. This model
deep learning methods are also widely used in image has a longer-term memory ability than the recurrent
recognition, voice recognition, and data analysis and neural network. In recent years, Li et al.9 and Halgan
processing. In fact, the concept of ‘‘artificial intelli- et al.15 have used computational fluid dynamics (CFD)
gence’’ was put forward by American scholar John method to build long short term memory network mod-
McCarthy in 1955. Around the 21st century, the devel- els and analyze the aeroelastic phenomena of bridges
opment of artificial intelligence has undergone major and airfoils. In the framework of deep learning, long
changes. Hao5 showed the design concept of artificial term short term memory models can process a large
intelligence, which transits from establishing a large amount of data and learn more hidden information
number of rules and system knowledge to machine from non-linear systems.16
learning. Machine learning is a branch of artificial In order to analyze flutter speeds under various flight
intelligence. The concept of machine learning is to use conditions, this study adopted K method of flutter
different algorithms to establish a set of systems that analysis method to analyze the occurrence of flutter
can enable computers to learn automatically, and the speeds and build a large amount of data. The K method
cognitive model trained by this can predict and judge is a semi-analytical and numerical method. This method
unknown data. Deep learning (DL) is a branch of needs to establish the mathematical model and also
machine learning, which was first proposed by Hinton needs computer coding to find flutter speeds. Most
et al.6 in 2006. The concept of DL is to superimpose importantly, the flutter speed can only be predicted by
multiple hidden layers to simulate the neural network K method case by case each time. Once the data bank
of the human brain for learning, but it also has the risk was established, using the deep learning method can be
of over-fitting. Therefore, Hinton et al. also proposed efficiency and easier in predicting the results. The pres-
the concept of dropout7,8 to improve the phenomenon ent study uses the DNN algorithm of machine learning
of over-fitting. The concept of dropout is to randomly to process complex aeroelastic model by superimposing
discard some neurons during the training process of the a multi-layer neural network architecture. Referring to
neural network, thereby reducing the risk of over-fit- Hagan et al.15’s research on neural network architec-
ting. Recently, both Li et al.9 and Halder et al.10 used ture, the present study designed a set of machine learn-
the deep long short-term memory (LSTM) networks to ing methods for computers to find rules from tens of
analyze the aeroelastic effects on the bridge and airfoil, thousands of flight data and obtain a set of classifica-
respectively. They both used the high-fidelity computa- tion methods. At the same time, this research also uses
tional fluid dynamics (CFD) to establish their fluid- the LSTM algorithm to establish a deep learning model,
solid coupled models. CFD uses computers to perform and compares the advantages and disadvantages of
the calculations required to simulate the free-stream DNN and LSTM. The supervised learning was used to
flow of the fluid. Most of the computing time is used to guide the machine to recognize the learning target with
calculate the interaction of the fluid with the solid sur- Labels. Among many programming languages, Python
face defined by the boundary conditions. With high- has the highest support at present.17,18 Python has a
speed supercomputers, better solutions can be achieved. large number of third-party modules and a powerful
It is commonly known that the CFD method is time standard library in order to use extended modules of
consuming11 and is difficult to build the big data for other programming languages. This study used this tool
machine learning. Ziaei et al.11 used the machine learn-
as the basis. Under such analysis, the deep learning net-
ing method to predict non-uniform steady turbulent
work can extract the features from various flight data,
flows in a 3D domain. Their proposed method provides
and find the flutter speed of the aircraft under different
an easy way for the designers and engineers to generate
flight conditions. Finally, the theoretical predictions
immense amounts of design alternatives without facing
from the K method were applied to verify the learning
the time-consuming task of evaluation and selection.
status of deep learning and analyze the effectiveness of
In the development of modern deep neural networks
deep learning.
(DNNs), feature learning and classification applica-
tions are even more outstanding. AlexNet12 won the
ImageNet LSVRC title in 2012. GoogLeNet13 won the Introduction to the basic theory
championship in the ILSVRC competition in 2014.
They achieved an error rate of only 6.67% using a 22- Among many flutter analysis methods, P-K method is
layer neural network architecture, which indirectly often used for analysis. Compared with P-K method,
proves that increasing the number of layers of neural the K method occasionally misjudges the degree of free-
networks can describe complex models more accurately dom of structure flutter when analyzing aeroelastic
and bring better accuracy. problems.1,19 However, this study only considered
Wang and Wang 3

€ + mbxa €h + Ka a = M
Ip a ð2Þ
where, m is the mass of airfoil, Ip is the moment of iner-
tia of the airfoil, xa = e  a is expressed as static unba-
lance parameter. If this parameter is positive, it means
that the position of the elastic axis is forward from the
position of the center of mass. The L and M are the lift
and aerodynamic moment generated by the aerody-
namic action of the airfoil. L and M can be expressed
in the following form:

L = 2pr‘ bU 2 a ð3Þ
Figure 1. Schematics of the airfoil. 1
M = M1 + b( + a)L ð4Þ
4 2
M1 = 0 ð5Þ
flutter speed, and did not focus on the flutter related to 4
qffiffiffiffi qffiffiffiffi
different degrees of freedom. Therefore, using K Kh Ka
Then vh = m , va = Ip , after simplification, com-
method will not affect the prediction of flutter speed,
and can save much time when generating a large bining equations (1) and (2) with equations (3)–(5), the
amount of flutter data. matrix form of the equation of motion of the airfoil can
be obtained as follows:
! !
2 2 €
Aeroelastic equations of motion for two-dimensional mb mb xa h

b
2
airfoils mb xa Ip a
€
0 2 2 2 1
Referring to the theory of flutter analysis proposed by mb vh 2pr‘ b U 2 !  
h 0
Hodges and Pierce,1 we analyzed the flutter phenom- B 2 C 
b
+@ (  1  a)pr‘ b U 2 A =
enon of an airfoil, in which this model includes two 0 a 0
degrees of freedom: plunge, and pitch. We assumed that + Ip va 2
these two degrees of freedom are subjected to springs ð6Þ
Kh and Ka respectively, and this model is shown in
Figure 1.
Here, Q in Figure 1 is the aerodynamic center. The
K method
aerodynamic center is the point that the pitching K method is to add the artificial damping term to the
moment for the airfoil does not vary with lift, and usu- right of the original aeroelastic equation of motion. By
ally defines as dCm/dCL = 0, where Cm is the pitching observing the variation of artificial damping, we can
moment coefficient, CL is the lift coefficient. C is the judge the divergence trend of the structure. Because K
center of mass, and P is the elastic axis b is the half method is not like the general classical flutter analysis
chord length of the airfoil, e is the location of elastic method, which needs a lot of iteration to calculate flut-
axis, a is the location of C.G., a is airfoil pitching ter speed, the K method is also the most efficient among
motion. The dimensionless parameters e and a are many classical flutter analysis methods. Therefore, we
between 1 and 21. If e is smaller than 0, it means that chose K method to establish and verify the deep learn-
the centroid is located closer to the leading edge (LE). ing model in this study.
If e is larger than 0, it means that the center of mass is After adding artificial damping term (FDh and FDa )
located closer to the trailing edge (TE). If e is 0, it to the right of equations (1) and (2), we can obtain the
means that the centroid position is at the position of following equations:
half chord; if a is smaller than 0, it means that the elas-
tic axis is located closer to LE. If a is larger than 0, it m(€h + bxa a
€ ) + K h h =  L + F Dh ð7Þ
means that the elastic axis is located closer to TE. If a € + mbxa €h + Ka a = M + FDa
Ip a ð8Þ
is 0, it means that the position of the elastic axis is on
the half chord position. We used simple harmonic motion to express the plunge
Through Euler-Lagrange equation, we can get the motion (h) and pitch motion (a) of this model, and its
basic Airfoil equation of motion as follows1: lift (L) and aerodynamic moment (M). They are
expressed as follows: h(t) = heivt , a(t) = a  eivt ,
m(€h + bxa a
€ ) + Kh h =  L ð1Þ  ivt  ivt
L(t) = Le , M(t) = Me . FDh and FDa can be
expressed in the following forms:
4 Advances in Mechanical Engineering

Figure 2. Artificial damping of K method prediction. Figure 3. Frequency of K method prediction.

herein is based on supervised learning, and a set of

gh heivt
FDh =  i ð9Þ neural networks was designed to predict the results after
repeated operations by a large number of artificial neu-
FDa =  i  eivt
ga a ð10Þ rons. In this study, supervised learning in machine learn-
qffiffiffiffiffiffi ing was adopted, so in building a learning model, besides
Ip vh
After substituting m = m 2 , r = 2
, s
^= va into the features of data, data labels should also be the input.
pr‘ b mb
Therefore, we needed to mark huge data, and the data
equations (7) and (8), we can obtain the following
marking in this study was based on K method of flutter
equation:
analysis. We set the parameters used in the airfoil equa-
0 h  2 i 1 tion of motion in Section II as data features, and the
m 1s ^ 2 vva ð1 + igh Þ flutter speed obtained by using K method was the data
B mxa + la (k, M‘ )C
B + l (k, M ) C label. We used computer processing to carry out a large
B h ‘ C
B h   iC
B number of labels to avoid spending too much time on
@ mr 1  v ð1 + iga Þ C
2 va 2
A
mxa + mh (k, M‘ ) data collection. The parameters used in the learning
+ ma (k, M‘ ) model are the location of center of mass, the location of
!  

h elastic axis, mass ratio, radius of gyration, frequency
 0
b =
a
 0 ratio, and static unbalance parameter. Table 1 is the
parameter data used by the dataset. Finally, there are
ð11Þ 350,892 data sets established in this study.
 In the feature scaling of data, the data normalization
where, k = bv U
U is the reduced frequency, M‘ = C‘ is the
method we used is Min-Max normalization, which
Mach number.
scales the data to the [0, 1] interval in equal proportion
After solving the determinant of equation (11), the
to avoid the excessive contribution of a certain para-
flutter speed can be obtained. Figures 2 and 3 are the
meter, which affects the convergence speed of model
results of using K method to obtain flutter speed.
building and further reduces the training efficiency. The
From the change of artificial damping in Figure 2, it
expression of Min-Max normalization is as follows:
can be seen that artificial damping will change from
negative value to positive value at the position of 4.6, x  xmin
which means that flutter will occur at the position of x0 = ð12Þ
xmax  xmin
dimensionless flutter speed (1/k) = 4.6. Meanwhile,
comparing with Figure 3, we can see the change trend where, x# denotes normalized data, x is raw data, xmin
of system frequency before and after flutter generation. is the smallest data value in the raw data set, and xmax
is the largest data value in the original data set.

Deep learning Deep neural network (DNN)

This study used the deep learning algorithm to perform DNN is a kind of Artificial Neural Network (ANN),
flutter speed analysis. The DNN architecture adopted also known as the Neural network. The concept of
Wang and Wang 5

Table 1. Parameters used by the dataset.

Variables in the flutter speed dataset

Features Label
Location of c.g. Location of Mass ratio Radius of Frequency Static
elastic axis gyration ratio unbalance
a e m r s
^ xa

Range [20.9, 0.9] [20.9, 0.9] [5, 60] [0.1, 0.9] [0.1, 0.9] [21.8, 1.8] [0, 100]

Figure 4. The schematic diagram of the neural network-like structure.

deep neural network is a neural network formed by a by multiplying each input value and weight value and
combination of a large number of neurons, which simu- summing it up with a given deviation value, s is the
lates the behavior of the human brain to transmit and activation function (AF). After obtaining the summa-
process information through neuronal connections to tion of weighted inputs and deviation value, it needs to
respond or solve problems. In the basic architecture of be converted by an activation function. The AFs used
a deep neural network, there are multiple hidden layers in this paper are Sigmoid function, Rectified Linear
between the input layer and the output layer, and each Unit (ReLU) function, and Softmax function, where
layer is composed of a large number of neurons, and the above AFs are respectively represented as follows,
z
all the neurons in the input layer are individually con- s(z) = 1 +1ez , s(z) = max (0, z),s(z)i = Pne i zj .
e
nected to the neurons in the hidden layer, and the neu- j=1

rons in the hidden layer are also individually connected

to the output layer. The schematic diagram of the Long short term memory (LSTM)
neural network-like structure is shown in Figure 4. LSTM uses three control gates to learn and achieve the
Neural network is composed of many artificial neu- effect of long-term memory, and these three control
rons, in which each artificial neuron is connected with gates will determine the storage and use of memory,
a weight value, and each artificial neuron has a devia- which are Input gate, Forget gate, and Output gate.
tion value. The relationship between the input value Figure 5 is a schematic diagram of the LSTM model.
and the output value is generally expressed by the fol- When a long-term and short-term memory model is
lowing form, operating, the hidden state at the previous time state
X
n  and the current input value will be entered into the
y=s wi xi + b ð13Þ long-term and short-term memory for calculation at
i=1 the same time, which will be multiplied by a weighting
value and added to a deviation value. Then, the conver-
where x is the input value of the artificial neuron, w is sion is performed through the activation functions of
the weight, b is the deviation value, y is a value obtained the input gate, the forget gate, and the output gate. The
6 Advances in Mechanical Engineering

At this point, the long and short-term memory unit has

been completely trained. The LSTM method uses mem-
ory units and hidden states to increase the dependence
of data training. Therefore, we use the LSTM method
to establish the deep learning model to predict the flut-
ter speed.

Loss function
The loss function we used in deep neural networks is
categorical cross-entropy. It is applicable to a variety of
classification problems. We also use Softmax function
Figure 5. LSTM model. as the activation function of the output layer. The cal-
culation of categorical cross entropy is shown in equa-
activation functions are all sigmoid functions, and the tion (20).
updated cell state uses the hyperbolic tangent function
X
n
as the activation function. The calculations are shown ED =  yi log (pi ) ð20Þ
as follows:. i

It = sðwIX ½Xt  + bI + wIh ½ht1  + bI Þ ð14Þ where n is the number of classification categories, yi is
the label of the data, and pi is the accuracy of the classi-
Ft = sðwFX ½Xt  + bF + wFh ½ht1  + bF Þ ð15Þ fier in predicting the occurrence of flutter speed. In
Ot = sðwOX ½Xt  + bO + wOh ½ht1  + bO Þ ð16Þ LSTM method, we used the mean square error (MSE)
to evaluate the error of the training model. This method
 t = tanhðwCX ½Xt  + bC + wCh ½ht1  + bC Þ
C ð17Þ is a commonly used regression loss function. This is a
where It is the information passing through the input commonly used regression loss function. The calcula-
gate, Ft denotes the information passing through the tion of the mean square error is shown in equation (21).
forgotten gate, Ot represents the information passing 1 Xq
through the output gate, C  t is the updated unit state, Xt EL = i=1
ðyi  ^yi Þ2 ð21Þ
q
represents the current input value, and ht-1 denotes the
previous hidden state. Then the result of the input gate where q is the total number of data, yi is the label of the
and the forgetting gate will determine how to update data, and ^yi is the predicted value.
the cell state transferred to the next long and short-term
memory. If the value converted by the activation func-
tion of the forgotten gate is 0, it means that the unit Back propagation
state at the last time point is cleared. If the value con- Back propagation is a sequential transmission from the
verted by the activation function of the input gate is 0, output layer to the input layer, and the difference
it means that the current information has not passed between the training data converted by the activation
the input gate. Therefore, the cell state calculated from function and the corresponding input target value is
the hidden state at the last time point and the current obtained, thus obtaining the loss function gradient
input value will not be used, and the updated cell state related to each weight parameter. Here, input training
calculation is shown in equation (18). data is converted through AF and we can obtain
y = s(z). The back propagation calculates and stores
t
Ct = Ft  Ct1 + It  C ð18Þ the loss function gradient in each neural network layer
through the chain rule in calculus and the difference
Ct is not only the updated cell state, but also the cell
between the output value trained by the neural network
state transferred to the next long and short-term mem-
and the exact output value, and repeatedly trains and
ory. Finally, the result of the output gate calculation
adjusts until the relevant weight value tends to be sta-
will determine how to update the hidden state. If the
ble. The gradient of the loss function in each neural
value of the activation function converted by the output
network layer is as follows,
gate is 0, it means that the current unit state cannot
pass the output gate, so the hidden state will not be ∂SL ∂z ∂y ∂SL
recorded. The updated hidden state expression is shown = ð22Þ
∂w ∂w ∂z ∂y
in equation (19).
After obtaining the gradient of the loss function in each
ht = Ot  tanhðCt Þ ð19Þ hidden layer, the weight can be updated by using the
Wang and Wang 7

gradient of the loss function to obtain the best weight

value. The loss function used by us in the DNN was
categorical cross-entropy. Finally, the Softmax function
was adopted as the AF of the output layer. The loss
function of classification cross entropy is as follows.

X
n
SL =  yi log (^yi ) ð23Þ
i

where, n is the number of classification categories, yi is

the volume label of the data, ^yi is the accuracy of flutter
generation predicted by the classifier.

Other parameter setting

In the weight update of the DNN, we choose Adam Figure 6. Accuracy of 1-layer deep learning architecture
method20 to calculate and adjust the weight value of training.
the updated model. We chose Adam method for weight
update because compared with other weight update
methods, this method is easier to execute, has relatively
high calculation efficiency, requires less memory, and is
suitable for problems with a large amount of data and
parameters. Adam related expressions are as follows.

mt = b1 mt1 + (1  b1 )gt ð24Þ

2
vt = b2 vt1 + (1  b2 )gt ð25Þ
where, new representations m^ t and ^vt can be obtained
by performing deviation correction on equations (24)
and (25). The representations are as follows,
mt
^t =
m ð26Þ
1  b1 t
vt
^vt = ð27Þ
1  b2 t Figure 7. Loss of 1-layer deep learning architecture training.

In equations (24) and (25), gt is expressed as loss func-

∂S
tion gradient, gt = ∂wLtt . Then, new weights can be cal- predict flutter speed, and carried out deep learning on
culated by using equations (26) and (27), and the 350,892 flight data. Among them, 70% of the flight
operation mode is as follows. data are regarded as training set, 20% as validation set,
and the remaining 10% as prediction data. In the design
hm^t of architecture, we first established the most basic
wt + 1 = wt  pffiffiffiffiffiffiffiffiffiffiffiffi ð28Þ three-layer deep learning framework for test training.
^vt + e
Among them, the starting function in the deep learning
where m and v are parameters used to adjust the weight architecture uses ReLU function for training except for
values of neural networks. Among them, the default Softmax function in the output layer. After 10,000
value of b1 is 0.9, and that of b2 is 0.999, and e is Epoch training, the results obtained are not satisfied.
1e208. The learning rate in this study h is set as 0.003. The training accuracy and testing accuracy are stuck at
82%. The training results are shown in Figures 6 and 7.
Modeling and Analysis After the first test, we decided to increase the num-
ber of artificial neurons and the number of layers of
DNN model deep learning architecture to achieve better accuracy.
Based on the deep learning theory proposed in last When a 7-layer deep learning architecture was used,
Section and referring to Hagan and Schmidhuber’s14 after 10,000 Epoch training, the training accuracy has
research on neural network architecture, this Section reached 97%, and the accuracy of verification reaches
constructed a set of deep learning architecture that can 94%. However, in the loss results, there is an increasing
8 Advances in Mechanical Engineering

Table 2. Training loss results with different activation function combinations.

DNN training and testing loss with different activation function arrangement
Activation Layer 1–5: Layer 1: ReLU, Layer 1–2: ReLU, Layer 1–3: ReLU, Layer 1–4: ReLU,
function sigmoid layer 2–5: sigmoid layer 3–5: sigmoid layer 4–5: sigmoid layer 5: sigmoid

Training loss 0.2261 1.6089 1.6181 1.6229 0.1363

Testing loss 0.2508 1.6091 1.6257 1.6231 0.1803
Layer 1–5: ReLU Layer 1: sigmoid, Layer 1–2: sigmoid, Layer 1–3: sigmoid, Layer 1–4: sigmoid,
layer 2–5: ReLU layer 3–5: ReLU layer 4–5: ReLU layer 5: ReLU
Training loss 0.1636 0.1475 0.1508 0.2814 0.2921
Testing loss 0.2234 0.2041 0.2712 0.3474 0.3444

Figure 8. Accuracy of 7-layer deep learning architecture Figure 9. Loss of 7-layer deep learning architecture training.
training.

trend with the increase of Epoch. The reason is that the

over-fitting phenomenon is due to the deep learning
architecture stacking too many layers or using too
many artificial neurons, which causes the noise in each
data to be amplified when verifying the results, result-
ing in an increasing trend of loss. The training results
are shown in Figures 8 and 9.
Figure 10 shows the training loss and testing loss
results of different layers of neural network architec-
ture. After many attempts (see Figure 10), we chose a
6-layer deep learning architecture and used Dropout7
to improve the over-fitting phenomenon. Although its
accuracy is not as high as that of the 7-layer deep learn-
ing architecture, it avoids the phenomenon of over-fit-
ting. We adopted Sigmoid function, ReLU function,
and Softmax function as the activation function of the
Figure 10. Training loss and testing loss of different layers of
model in this study. The Categorical cross-entropy was
neural network architecture.
selected as the loss function of this model, and Softmax
function as the starting function of the output layer.
The activation function of each other layer uses
As shown in Table 2, when the ReLU function is
Sigmoid function and ReLU function. Table 2 shows
used as the activation function in Layer 1 to Layer 4,
the results of model training loss under different combi-
and the Sigmoid function is used as the activation
nations of Sigmoid and ReLU activation functions.
Wang and Wang 9

Figure 11. Training loss results of different number of neurons.

function in Layer 5, the loss of training results is the rel-

atively lowest compared with other combinations, and
its training loss is 0.1363, and the loss during verifica-
tion is 0.1803. Therefore, this group of combination
was chosen as the basis for establishing the deep learn- Figure 12. Deep learning model.
ing model activation function of this study. After deter-
mining the activation function of the model in this
study, we decided to train the model with different
number of neurons and find out the number of neurons
suitable for this model. Figure 11 is the result of train-
ing losses using different number of neurons. From the
training loss results of different number of neurons, it
can be seen that when the number of neurons increases,
the training loss tends to decrease with the increase of
the number of neurons. However, the loss of test tends
to increase after 120 neurons, we then use 120 neurons
to establish the deep learning model of this study.
Based on the above results, we have established a 6-
layer deep learning architecture, in which layers 1–5 use
ReLU function as activation function, layer 6 uses
Sigmoid function as activation function, the output
layer uses Softmax function as activation function, the
loss function uses Categorical cross-entropy method, Figure 13. Precision of deep learning architecture.
and each layer is established with 120 neurons. Figure
12 is the framework of the deep learning model in this remaining 10% as the predicted data and import them
study, and the training results are shown in Figures 13 into the LSTM model for training. The input para-
and 14, and the accuracy of the trained model is 95.6%. meters of the LSTM model in this study are consistent
Although it is 1.4% lower than the previously estab- with the DNN model. In the design of the LSTM archi-
lished 7-layer deep learning architecture (97%), this tecture, we first established a basic single-layer deep
model has higher robustness. learning architecture for testing and training, and then
increased the number of layers of long-term and short-
term memory. Finally, we chose to use five layers of
LSTM model long- and short-term memory as the LSTM model. The
We regard 70% of the 350,892 flight data as the train- LSTM architecture test training results of different
ing set, 20% of the data as the validation set, and the layers are shown in Figure 15.
10 Advances in Mechanical Engineering

Figure 14. Loss of deep learning architecture.

Figure 16. LSTM training loss using different numbers of cells.

Figure 15. LSTM testing and training results of different layers.

Since the Sigmoid function and the hyperbolic tan-

Figure 17. Loss and accuracy of different epoch training.
gent function have been used as the activation function
in the LSTM model, we no longer use the activation
function between the hidden layers to avoid excessive establish this deep learning model. Next step is to
non-linear conversion between the hidden layers caus- choose an appropriate number of epochs to train the
ing the output value to be distorted. In the training LSTM model in this study. Figure 17 shows the loss
process, we also added a dropout layer between the and accuracy results of different epoch training. It can
hidden layers to prevent over-fitting. This study uses a be seen from Figure 18 that the loss result after 3000
5-layer LSTM layer. Since the previous training has epochs training is low; however, when it is increased to
used a fixed number of LSTM cells, we will tune the 4000 epochs, the accuracy does not increase signifi-
number of LSTM cells to obtain the best number of cantly. Therefore, we choose to use 3000 epochs to
long and short-term memory cells. Figure 16 shows the build the deep learning model of this research.
training loss results using different numbers of LSTM Based on the above results, we have established a 5-
cells. From the training loss results of different num- layer LSTM architecture, in which the loss function
bers of long and short-term memory cells, it can be seen uses the mean square error method, and each layer is
that when the number of cell is 80, the training loss and built using 80 LSTM cells, and the number of training
test loss are small, so we chose to use 80 cells to is 3000 epochs. Figure 17 is the architecture of the
Wang and Wang 11

Figure 20. Loss of the LSTM training model.

Figure 18. Architecture of the LSTM model.

Figure 21. DNN prediction results.

Table 3. DNN model data.

Variables information Setting point

Figure 19. Accuracy of the LSTM training model. Algorithm DNN
Number of samples 350,892
Classification 1–400
LSTM model of the present study. The average accu- Average accuracy 95.6%
racy of the trained model is 96.8%. The accuracy and
loss results of the training are shown in Figures 19
and 20. flutter speed versus the predicted flutter speed, while
the dashed line represents the ideal case of DNN
(Figure 21) or LSTM (Figure 22) model prediction
Model prediction and result discussion
result. The verification results by LSTM model are
In this Section, the deep learning model obtained from shown in Figure 22. In the verification result, an accu-
last Section was used to predict flutter speed of various racy of 96.8% is achieved. Again, the red dots in
flight parameters. The verification results by DNN Figure 22 indicate the results of the true flutter speed
model are shown in Figure 21. Table 3 is the relevant versus the predicted flutter speed, while the dashed line
data of this DNN model. In the verification result, an represents the ideal case of LSTM model prediction
accuracy of 95.6% is achieved. It is noted that the red result. The results show that the LSTM model has bet-
dots in Figures 21 and 22 indicate the results of the true ter accuracy than the DNN model.
12 Advances in Mechanical Engineering

Figures 23 to 27 are the prediction results of flutter

speed after input and training of various parameters.
Vertical axis is flutter speed. The higher the value is, the
higher the flutter speed is. A higher flutter speed means
that it is less likely to cause structure divergence. For
each figure, (a) is the predicted results by DNN, and
(b) is the predicted results by LSTM model. Figure 23
shows that this model has high stability if the center of
mass locates from quarter chord to half chord, the esti-
mated value is consistent with aeroelastic point of view.
That is, when the center of mass is located in the front
half of airfoil (quarter chord), the model has higher sta-
bility than the rear half. Figure 24 shows that the aeroe-
lastic model is more stable when the position of elastic
axis is placed near the quarter chord. It is true that
Figure 22. LSTM prediction results. from aerodynamic point of view, when the position of
elastic axis is placed near the quarter chord, the system
will have higher stability. Figure 25 is the relationship

Figure 23. Relationship between location of center of mass and flutter speed, predicted by (a) DNN model and (b) LSTM model.

Figure 24. Relationship between location of elastic axis and flutter speed, predicted by (a) DNN model and (b) LSTM model.
Wang and Wang 13

Figure 25. Relationship between mass ratio and flutter speed, predicted by (a) DNN model and (b) LSTM model.

Figure 26. Radius of gyration and flutter speed, predicted by (a) DNN model and (b) LSTM model.

between mass ratio and flutter speed, where the mass stability when it is placed in front of the center of mass.
ratio is expressed as m = m=pr‘ b2 . The mass ratio is a This is also consistent with aeroelastic point of view,
dimensionless parameter of air density. When the mass that is, the position of the elastic axis is placed in front
ratio is larger, it means that the air density is lower, and of the center of mass, which has higher stability than
the energy obtained by the elastic structure from the air when it is placed behind the center of mass. So far, the
is greatly reduced, so the possibility of the occurrence 6-layer DNN and LSTM methods we have established
of flutter decreases and the stability is higher. Figure 26 can provide accurate prediction of flutter speed.
is the relationship between the radius of gyration and From Figures 23 to 27, we can see the trend of the
flutter speed, and the increase of rotation radius has a flutter speed predicted by LSTM is consistent with the
relatively stable trend, where the radius of gyration is
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi trend of the prediction result of the DNN model, and
expressed as r = Ip =mb2 . Thus, the radius of gyration the prediction conditions of both can be explained by
is related to the moment of inertia. The larger the the physical meaning of aerodynamics and aeroelasti-
radius of gyration, the larger the moment of inertia, city. As far as the results of the current stage of this
and the more stable the pitching motion is. Figure 27 is research are concerned, the prediction of flutter speed
the relationship between static unbalance parameter has reached a good level, among which the average pre-
and flutter speed. This value represents the distance diction accuracy of the DNN model and the LSTM
between the center of mass position and the elastic axis model have reached more than 95%. Figures 21 and 22
position, and it is relatively stable when it is 0 to 21, are the predictions by importing the same data set.
which means that the elastic axis has relatively high From the two figures, it can be clearly judged that the
14 Advances in Mechanical Engineering

Figure 27. Relationship between static unbalance and flutter speed, predicted by (a) DNN model and (b) LSTM model.

deep learning model established by using the LSTM 1. The number of layers and the number of neu-
method is better than the deep learning model estab- rons for the DNN and the LSTM will affect the
lished by the DNN method. Moreover, most of the pre- accuracy of the model. Not that the more layers
dictions of the LSTM model are closer to the the better the effect. The number of neurons is
theoretical flutter speeds than the DNN model. From not the more the better the effect. Too many
Figures 23 to 27, the fluctuations of the flutter speeds layers or too many neurons will cause the model
predicted by the DNN model are larger than that of to produce over-fitting phenomenon, resulting
the LSTM model, and the robustness is also lower than in a decrease in the accuracy of the prediction
that of the LSTM model. We believe that the reason results. A case analysis based on this research is
for this result is that the LSTM model has a cyclic necessary.
architecture, which enables it to store information and 2. The number of epochs of the deep learning
update it with the latest information at the same time. model in this study determines whether the
It also uses three control gates to adjust and select model is fully trained, but if the epoch is set too
memory storage and access. Although the training of large, it may cause the model to be over-fitted
the model is more time-consuming than DNN, this during the training process, which will affect the
method can increase the memory space. The DNN accuracy. At the same time, too many number
model is a feed-forward neural network. The data sets of epochs will increase the calculation time and
imported into this model are not well related. Although reduce the efficiency of training. Choosing the
the accuracy of the DNN model has reached 95.6%, appropriate number of epochs will improve the
the prediction robustness of the DNN model is rela- efficiency of training.
tively low. 3. The prediction results of the DNN and the
LSTM deep learning models conform to the
theoretical explanation of aeroelasticity and
Conclusions aerodynamics.
4. This study uses DNN and LSTM methods to
This study uses deep learning algorithms to analyze the
build deep learning models, with an average
aeroelastic phenomenon and compare the differences accuracy of more than 95%. The LSTM model
between Deep Neural Network (DNN) and Long performs better than the model established by
Short-term Memory (LSTM) applied on the flutter the DNN method, and its robustness is also
speed prediction. Instead of time-consuming high-fide- higher.
lity computational fluid dynamics (CFD) method, this
study uses the K method to build the aeroelastic flutter
speed big data for different flight conditions. The flutter In this study, DNN and LSTM are employed to predict
speeds for various flight conditions are predicted by the the occurrence of flutter speed, both of the deep learn-
deep learning methods and verified by the K method. ing methods are used to learn the relation between each
The detailed physical meaning of aerodynamics and flight data and its associated flutter speed by consider-
aeroelasticity of the prediction results are studied. The ing this problem as a multi-class classification, so as to
conclusions are listed below: achieve the goal of this study. The results show that the
Wang and Wang 15

LSTM architecture can effectively predict the flutter 12. Krizhevsky A, Sutskever I and Hinton GE. ImageNet
speed, and it is a potential model in the application of classification with deep convolutional neural networks.
aerospace field. In: Advances in neural information processing systems,
Lake Tahoe, NV, 2013. Curran Associates, Inc.
Declaration of conflicting interests 13. Szegedy C, Liu W, Jia Y, et al. Going deeper with convo-
lutions. In: Neural information processing systems confer-
The author(s) declared no potential conflicts of interest with ence, Montreal, Canada, 2015.
respect to the research, authorship, and/or publication of this 14. Hochreiter S and Schmidhuber J. Long short-term mem-
article. ory. Neural Comput 1997; 9: 1735–1780.
15. Hagan MT, Demuth HB, Beale M, et al. Neural network
Funding design. 2nd ed. Martin Hagan, 2014.
The author(s) disclosed receipt of the following financial sup- 16. LeCun Y, Bengio Y and Hinton G. Deep learning.
port for the research, authorship, and/or publication of this Nature 2015; 521: 436–444.
article: This research was supported by the Ministry of 17. Chollet F. Keras, https://github.com/fchollet/keras (2015).
Science and Technology of Taiwan, Republic of China (grant 18. Abadi M, Barham P, Chen J, et al. Tensorflow: a system
number: MOST 110-2221-E-032-026). for large-scale machine learning. In: 12th USENIX sym-
posium on operating systems design and implementation,
ORCID iD Savannah, GA, 2016, pp.265–283.
Yi-Ren Wang https://orcid.org/0000-0001-5641-2491 19. Hassig HJ. An approximate true damping solution of the
flutter equation by determinant iteration. J Aircr 1971; 8:
References 885–889.
20. Kingma DP and Ba J. Adam: a method for stochastic
1. Hodges D and Pierce GA. Introduction to structural optimization. arXiv:1412.6980v9 [cs.LG], 2017.
dynamics and aeroelasticity. Cambridge Aerospace Series.
2011. Cambridge: Cambridge University Press.
2. Bisplinghoff RL, Ashley H and Halfman RL. Aeroelasti- Appendix
city (Dover Books on Aeronautical Engineering). New
York, NY: Dover Publications, 2013. Notations (for the aeroelastic system)
3. Pitt D and Haudrich D. Development of an artificial
a location of C.G.
neural aeroelastic network (AN^2) for the prediction of b half chord length of the airfoil
multiple flutter crossing. In: 46th AIAA/ASME/ASCE/
AHS/ASC structures, structural dynamics and materials
C center of mass
conference, 18–21 April 2005, Austin, TX, USA. e location of elastic axis
4. Silver D, Schrittwieser J, Simonyan K, et al. Mastering gh amplitude of the plunge artificial
the game of go without human knowledge. Nature 2017; damping
550: 354–359. ga amplitude of the pitching artificial
5. Hao K. We analyzed 16,625 papers to figure out where damping
AI is headed next. MITS Technol Rev 2019. https:// h airfoil plunge motion
www.technologyreview.com/2019/01/25/1436/we-analyzed- h amplitude of h
16625-papers-to-figure-out-where-ai-is-headed-next/ IP moment of inertia of the airfoil
6. Hinton GE, Osindero S and Teh YW. A fast learning Kh spring constant in the plunge D.O.F.
algorithm for deep belief nets. Neural Comput 2006; 18: Ka spring constant in the pitching D.O.F.
1527–1554.
L lift
7. Hinton GE, Srivastava N, Krizhevsky A, et al. Improv-
ing neural networks by preventing co-adaptation of fea-
M aerodynamic moment
ture detectors. arXiv:1207.0580, 2012. m mass of airfoil
8. Srivastava N, Hinton GE, Krizhevsky A, et al. Dropout: P the elastic axis
a simple way to prevent neural networks from overfit- Q aerodynamic center
ting. J Mach Learn Res 2014; 15: 1929–1958. U flight speed
9. Li T, Wu T and Liu Z. Nonlinear unsteady bridge aero- xa = e  a static unbalance parameter
dynamics: reduced-order modeling based on deep LSTM a airfoil pitching motion
networks. J Wind Eng Ind Aerodyn 2020; 198: 104116. a amplitude of a
10. Halder R, Damodaran M and Khoo BC. Deep learning r‘ air density
based reduced order model for airfoil-gust and aeroelas- v airfoil natural frequency
tic interaction. AIAA J 2020; 58: 4304–4321. va airfoil natural frequency in the pitching
11. Ziaei D, Hekmatiathar S and Goudarzi N. Assessment of a
D.O.F.
CFD-based machine learning approach on turbulent flow
vh airfoil natural frequency in the plunge
approximation. In: ASME 2019 13th international conference
on energy sustainability collocated with the ASME 2019 heat D.O.F.
transfer summer conference, Bellevue, WA, 2019.