You are on page 1of 8

Alexandria Engineering Journal (2019) 58, 1367–1374

H O S T E D BY
Alexandria University

Alexandria Engineering Journal


www.elsevier.com/locate/aej
www.sciencedirect.com

ORIGINAL ARTICLE

The ability of forecasting flapping frequency of


flexible filament by artificial neural network
M. Fayed a,b, M. Elhadary b,*, H. Ait Abderrahmane c, Bassem Nashaat Zakher d

a
College of Engineering and Technology, American University of the Middle East, Kuwait
b
Department of Mechanical Engineering, Faculty of Engineering, Alexandria University, Egypt
c
Khalifa University of Science and Technology, Masdar Institute, Masdar City, P.O. Box 54224, Abu Dhabi,
United Arab Emirates
d
Pharos University in Alexandria (PUA), Mechanical Department, Alexandria, Egypt

Received 28 August 2019; revised 6 November 2019; accepted 13 November 2019


Available online 27 November 2019

KEYWORDS Abstract Artificial Neural Networks (ANNs) are reliable and computationally inexpensive
Flapping frequency; compared to numerical methods such as CFD simulations and experimental investigations in
Artificial neural network aerodynamics research. In this article, an Artificial Neural Network (ANN) has been introduced to
(ANN); predict the flapping frequencies of a filament placed in a 2-D soap-film tunnel. The multi-layer
Flexible filament perception (MLP) networks have been used in developing the Artificial Neural Network while the
backpropagation Levenberg-Marquardt algorithm was used to perform the training of the ANN.
A part of the experimental data was considered for the training process while the rest for the prediction
test of the suggested ANN. The ANN results indicate that it can predict the frequencies of the periodic
flapping with good accuracy. However, it fails when the flapping presents amplitude modulation.
Ó 2019 The Authors. Published by Elsevier B.V. on behalf of Faculty of Engineering, Alexandria
University. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/
licenses/by-nc-nd/4.0/).

1. Introduction established that aerodynamic coefficients can be identify using


neural networks.
The capabilities of Artificial neural networks (ANNs) in pre- Artificial Neural Networks are applied for the flight control
dicting drag coefficients, unsteady development, separation applications and the identification of dynamical systems due to
and reattachment of the boundary layer, steady and unsteady their ability to approximate general continuous nonlinear
aero dynamic load and airfoil harmonics oscillations are pro- functions [3–5]. Rakotomamonjy et al. [6] used experimental
ven [1]. Linse and Stengle [2] used a Computational Neural results of hummingbird and insects flapping flight to develop
Network to derive aerodynamic coefficient models using the a flight dynamics simulation model for flapping-wing micro
simulated flight-test data for a twin-jet transport aircraft. They air vehicles. They introduced a neural network to model the
movements of the wing by imitating numerous function shape.
McCracken [7] aimed to solve the Navier-Stokes Equations for
* Corresponding author.
turbulence by applying the neural network to detect, based on
E-mail address: mostafaelhadary@yahoo.com (M. Elhadary).
3-dimensional inputs, ionospheric velocity fields. The neural
Peer review under responsibility of Faculty of Engineering, Alexandria
network succeeded to map the velocity fields within a 67%
University.
https://doi.org/10.1016/j.aej.2019.11.007
1110-0168 Ó 2019 The Authors. Published by Elsevier B.V. on behalf of Faculty of Engineering, Alexandria University.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
1368 M. Fayed et al.

Nomenclature
di target values for output node i
List of symbols oi model values for output node i
L filament length (mm) p input/output patterns training number
N the number of output nodes V flow speed (m/s)

precision of the testing data used. Khezzar et al. [8] designed 2. Experiments
an artificial neural network model to predict the reattachment
length for laminar flow. They demonstrated that, for a limited The experiments were conducted in vertical gravity-driven
choice of Reynolds numbers, computational fluid dynamics flowing soap film tunnel; Fig. 1a. A nozzle from a reservoir
(CFD) can be used in the first instance to obtain reattachment was designed to supply the soapy solution, while the planar
lengths and ANN will be used for other intermediate Reynolds soap film was obtained by spreading and guiding the soapy
number values. Awad et al. [9] applied the performance and water via two taut nylon threads of 1 mm in diameter, which
robustness of a proposed control system based on neural net- are spaced by 100 mm from each other.
work approach to active fins that takes into account roll The soapy water was pumped into the reservoir with aid of
response spectral characteristics transitions due to not only a gear pump. To fix film thickness and the flow rate in the test
varying sea state but also changing sailing conditions is section, the fluid height in the reservoir was kept constant. The
assessed. Babikir et al. [10] presented a new model that com- reservoir inhibits the flow perturbations due to the pump. The
posed of an Artificial Neural Network (ANN) integrated with filament of various lengths was placed within the soap film,
a hybrid algorithm consists of Cat Swarm Optimization (CSO) hold by a needle. The filament should be placed where the flow
and Firefly Algorithm (FA) algorithms to predict the noise of a reached its terminal velocity with constant thickness, so it was
submersible Axial Piston Pump (APP) for different valve seat placed sufficiently far downstream from the nozzle. The flow
materials. velocity was measured by tracking a small droplet of dye trans-
In the present study, we test the ability of an ANN model in ported by the flow; see Fig. 2a(i) and a(ii). This value of the
predicting the filament flapping frequencies in a two- velocity has been verified by tracking the center of vortices
dimensional soap-film tunnel. The Neural Networks Toolbox shed by the flapping filament when sodium light is used and
of the MATLAB software with backpropagation Levenberg- the very small air bubbles present in the soap film [13].
Marquardt algorithm [11] has been used to develop the To visualize the flow, the light source was placed in front of
ANN model. This is considered as a canonical experimental the test section. The filament flapping motion was logged at a
model of fluid-structures interactions problem, which occurs frame rate of 800 fps using a camera of high speed. A sample
in several industrial and natural situations. These include soft of series snapshots, depicting a half period of a 35-flapping
plate vibration in snoring, plant leaf’s oscillation, heat cycle of the flapping filament, is shown in Fig. 1b. The gray
exchangers with parallel-plate-type vibration, flags fluttering, represents the background of the owing soap film while the
web vibration in paper-making [12]. filament is in black. To extract the flapping dynamics of the

Fig. 1 (a) A Schematic of the experimental setup; (b) raw snapshots of the flapping filament in the soap film tunnel; (c) superposition of
the successive positions of the filament.
The ability of forecasting flapping frequency of flexible filament 1369

3. The Artificial neural network to forecast the flapping motions

3.1. Artificial neural networks

The main components of ANN are the input, the hidden and
output layers. The back-propagation training algorithm is
commonly used to iteratively minimize the following cost func-
tion concerning the interconnection weights and neurons
thresholds [14,15]:

Fig. 2 Experimental details: (a) flow-speed measurement tech-


nique, (b) flapping filament, and (c) wake behind a flapping
filament when using low-sodium light [13].

filaments we used an image processing algorithm, implemented


in MATLAB environment; Fig. 1c.
The flapping dynamics of the filament depends on two con-
trol parameters namely, the length of the filament and the flow
velocity. In this study we have considered filaments of different
lengths: 40, 50, 60, 80, and 100 mm. The flow speeds were
ranging from 4.2 m/s to 8.7 m/s. In these conditions, we did
not observe interference in the soap film illuminated by a
low-pressure sodium light. For higher speed values, fluctua-
tions will be generated in the soap film and the flow condition
will change from subcritical to supercritical [13]. On the other
hand at a speed less than 4.2 the filament will not fluctuate at
any filament length. A time series sample of the filament free
end displacement and its power spectral density is shown in
Fig. 3. These two figures indicate that the flapping of the fila-
ment is periodic with the fundamental frequency (f = 35.94
Hz). The other peaks in the power spectrum are the harmonics Fig. 4 Neural network architecture (test 1); NMSE for each
of the fundamental frequency. number of neurons in the hidden layer.

Fig. 3 Time series and fast Fourier transform of the filament flapping dynamic. The filament is 50 mm long placed in owing with a speed
of 4.25 m/s.
1370 M. Fayed et al.

L =40 mm L =80 mm

L =60 mm
L =100 mm

Fig. 5 FFNN Simulations for (a) L = 40 mm, (b) L = 60 mm (c) L = 80 mm (d) L = 100 mm.

Fig. 6 FFNN Simulations for L = 50 mm.

1 Xp XN 3.1.1. Tests
E¼ ðdi  oi Þ2 ð1Þ
2 i¼1 i¼1
The ability to forecasting the flapping motion was investigated
In the present study, when the normalized-mean-square- with two different tests and with two different techniques. In
error (NMSE) between the ANN outputs and the tested data the first test, the inputs of the neural network were the filament
for all elements in the training set has established the target length and flow speed while the flapping frequency was the sin-
value, the training process is finished. gle output of the network. In the second test, there were three
The ability of forecasting flapping frequency of flexible filament 1371

inputs: filament length, flow speed and time. The output was output of suggested Feed-Forward Neural Networks (FFNN)
the amplitude of the filament endpoint displacement. at L = 40, 60, 80 and 100 mm, respectively.
Test 1: Fig. 4a shows the inputs and outputs for the first Fig. 6 shows the comparison between the experimental data
test. The tested NN was trained at different filament lengths, and the expected NN values relation between the flow speed
L = 40, 60, 80 and 100 mm and expected NN values were and the flapping frequency for L = 50 mm.
obtained for filament length L = 50 mm. For each filament Test 2: Fig. 7a shows the suggested NN architecture for this
length, more than 10 different flow speeds within the interval test. It is a feed-forward NN with an input layer, which has 10
[4.2; 8.7] m/s were tested. The output was the flapping frequen- sigmoid neurons and one pure-linear neuron in the output
cies. These are compared with the one found using Fast Four- layer. These layers were tested with one hidden layer contain-
ier Transform (FFT) of the row time series. ing 5, 10, 12, 20 pure-linear neurons. In this test, we used time
The suggested NN architecture was feed-forward NN with as the third input in addition to the two inputs of the first test
an input layer, which has 10 sigmoid neurons and one pure- namely: the filament length and the flow speed. The output
linear neuron in the output layer. The hidden layer was tested here is the amplitude of the flapping motion. For certain fila-
5, 10, 12, 20 purelinear neurons. ment length and flow conditions, the filament flapping dynam-
Fig. 4b illustrates NMSE for each number of neurons of the ics is not always periodic, but it can be chaotic or quasi-
hidden layer. The best performance was for a hidden layer of periodic [13]. The goal of this second test is to test the ability
12 neurons. of the NN in predicting the complex flapping dynamics of
Fig. 5 shows the relation between the flow speed and the the filament. The output in this test is the displacement time
flapping frequency for both the experimental data and the series of the filament free end. This time series will be

Fig. 7 Neural network architecture for test 2; NMSE for each number of neurons in the hidden layer.
1372 M. Fayed et al.

V = 5.2 m/s V = 7.7 m/s

V = 6.3 m/s V = 8.48 m/s

Fig. 8 Comparison between the trained NN values and experimental for L = 100 mm and different velocity.

V = 4.376 m/s V = 7.681 m/s

V = 4.947 m/s V = 8.583 m/s

Fig. 9 Comparison between the trained NN values and experimental for L = 50 mm and different velocity.
The ability of forecasting flapping frequency of flexible filament 1373

V = 5.119 m/s V = 6.754 m/s

V = 7.592 m/s V = 8.737 m/s

Fig. 10 Comparison between the trained NN values and experimental for L = 80 mm and different velocity.

compared to the one measured experimentally. Fig. 7b illus- the suggested ANN model was a feed-forward neural network
trates NMSE for each number of neurons of the hidden layer. with three layers: one input layer, one output layer and a hid-
The best performance was for a hidden layer of 12 neurons. den layer in between. The backpropagation Levenberg-
The amplitude of the filament flapping was changing with Marquardt algorithm was used to perform the training of
time. The training was made for a period of 0.25 s which rep- the suggested ANN. In the first model, the flapping fundamen-
resents in average 300-time steps. The data was trained for tal frequency was successfully predicted when the filament
around 12 flapping periods. In the case of low flow speed length and the flow speed were given as inputs. In the second
and small length filament, the flapping is symmetric, and all model, time was used to be as a third input in addition to
the flapping cycles were identical. For high velocities and long the two inputs of the first model and the output was the dis-
filament, the filament flapping is not strictly periodic. The flap- placement time series of the filament free end. The neural net-
ping showed amplitude modulations. work was able to predict successfully the amplitude of the
The neural network was trained for four different filament filament oscillations when the flapping was periodic. However,
lengths and four flow velocities. The results of training for the as soon as small amplitude modulation of the periodic flapping
case of L = 100 and 50 mm are shown in Figs. 8 and 9. As starts to appear, some discrepancies between the predicted and
shown in the figures the comparisons between the experimental the observed oscillations were noticed. These discrepancies
and the trained data using neural agreed well for all the cases. grow with the increase of the amplitude modulations until
The prediction, made for L = 80 mm and four flow velocities, the prediction is completely lost.
are shown in Fig. 10. This figure shows a quite good prediction
as long as the filament flaps periodically. The predictions start Declaration of Competing Interest
to depart from the experimental data when relatively small
wave modulation is present when flow speeds get higher. The authors declared that there is no conflict of interest.

4. Conclusion References

[1] D.F. Kurtulus, Ability to forecast unsteady aerodynamic forces


The aim of this study was testing the ability of the neural net-
of flapping airfoils by artificial neural network, Neural Comput.
work to forecast the flexible filament flapping dynamics in 2D Appl. 18 (4) (2009) 359.
flow. It was found to be practically applicable to predict the [2] D.J. Linse, R.F. Stengel, Identification of aerodynamic
fundamental flapping frequency of the flexible filament. The coefficients using computational neural networks, J. Guid.
capability was tested by using two models. For both models, Control Dyn. 16 (6) (1993) 1018–1025.
1374 M. Fayed et al.

[3] K.S. Narendra, K. Parthasarathy, Identification and control of [10] H.A. Babikir, M.A. Elaziz, A.H. Elsheikh, E.A. Showaib, M.
dynamical systems using neural networks, IEEE Trans. Neural Elhadary, D. Wu, Y. Liu, Noise prediction of axial piston pump
Netw. 1 (1) (1990) 4–27. based on different valve materials using a modified artificial
_
[4] K.J. Hunt, D. Sbarbaro, R. Zbikowski, P.J. Gawthrop, Neural neural network model, Alexandr. Eng. J. (2019).
networks for control systems—a survey, Automatica 28 (6) [11] H. Demuth, M. Beale, Neural Network Toolbox for Use with
(1992) 1083–1112. Matlab–User’S Guide Verion 3.0, 1993.
[5] A.J. Calise, R.T. Rysdyk, Nonlinear adaptive flight control [12] M.J. Shelley, J. Zhang, Flapping and bending bodies interacting
using neural networks, IEEE Contr. Syst. Mag. 18 (6) (1998) 14– with fluid flows, Annu. Rev. Fluid Mech. 43 (2011) 449–465.
25. [13] H. Ait Abderrahmane, M.P. Paidoussis, M. Fayed, H.D. Ng,
[6] T. Rakotomamonjy, M. Ouladsine, T.L. Moing, Modelization Flapping dynamics of a flexible filament, Phys. Rev. E.84 (2011)
and kinematics optimization for a flapping-wing microair 1–8.
vehicle, J. Aircraft 44 (1) (2007) 217–231. [14] S.V. Kartalopoulos, S.V. Kartakapoulos, Understanding
[7] M. McCracken, Artificial neural networks in fluid dynamics: a Neural Networks and Fuzzy Logic: Basic Concepts and
novel approach to the Navier-Stokes equations, arXiv preprint Applications, Wiley-IEEE Press, 1997.
arXiv:1808.06604, 2018. [15] David M. Skapura, Building Neural Networks, vol. 304, ACM
[8] Lyes Khezzar, Saleh M. Al-Alawi, CFD and artificial neural Press/Addison-Wesley Publishing Co., New York, NY, USA,
networks analysis of plane sudden expansion flows, Int. J. Eng. 4 1995.
(4) (2010) 296–307.
[9] T. Awad, M.A.E. Elgohary, T.E. Mohamed, Ship roll damping
via direct inverse neural network control system, Alexandr. Eng.
J. 57 (4) (2018) 2951–2960.

You might also like