Adityo 2020 J. Phys. Conf. Ser. 1511 012074

Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

Journal of Physics: Conference Series

PAPER • OPEN ACCESS

Determination of robust weights hidden layers on backpropagation


algorithm to analyze coefficient drag high-speed train
To cite this article: Suksmono Adityo et al 2020 J. Phys.: Conf. Ser. 1511 012074

View the article online for updates and enhancements.

This content was downloaded from IP address 114.5.213.134 on 05/06/2020 at 15:55


International Conference on Science Education and Technology IOP Publishing
Journal of Physics: Conference Series 1511 (2020) 012074 doi:10.1088/1742-6596/1511/1/012074

Determination of robust weights hidden layers on


backpropagation algorithm to analyze coefficient drag high-
speed train

Adityo Suksmono, Hendrato, Sucipto, Irfan Ansori


Center of Technology for System and Infrastructure of Transportation, Agency for
The Assessment and Application of Technology, Serpong, Banten, Indonesia

Abstract. In this research, we developed a method using backpropagation to analyze the


coefficient drag of high-speed trains. Analyzing coefficient drag using a backpropagation
algorithm has more benefits, especially in cost and time, than using computational fluid
dynamics. Computational fluid dynamics need sophisticated software and hardware. It needs
much time to get a convergent result as well. We used 2D coordinates longitudinal profile nose
of the high-speed train as an input of the backpropagation algorithm. The weights between
hidden layers and input layers and between hidden layers and output layers, respectively, were
modeled as matrices that were formed from the iteration process. The coefficient drag
differences, between backpropagation algorithm and computational fluid dynamics analysis,
from each iteration, were used as a correction factor to form robust weights hidden layers
matrices. The results of this research showed that training in the backpropagation algorithm
can obtain robust weights of hidden layers that have been known from Mean Sum Square Error
in an exercise that is small enough. Because of the limited time to finish this research, we only
trained and exercised nine models instead of a thousand models. Robustness weights that are
resulted in this research are expected to contribute to accelerating a coefficient drag prediction
of high-speed train accurately. To improve this proposed method, 3D coordinates of the nose’s
surface of high-speed trains and many more 3D models are needed.

1. Introduction
Reducing energy consumption at a high-speed train to reduce its cost is the main target for the high-
speed train design phase recently. To achieve this, it must be started with proper design. Design
processes involve CFD analysis, which then it will be validated in wind tunnel testing. It is an iterative
process because based on wind tunnel testing it will give feedback to updated design to fulfill design
requirements. Thus design process needs more budget and time. For that reason, we need sophisticated
software to analyze accurately because of a lack of accusation in phase design which is validated using
wind tunnel testing, which is more cost and time to update the initial design.
To make a faster and cheaper phase design, in this research, by using the Backpropagation
algorithm, we developed robust weights between input layers and hidden layers and weights between
hidden layers and output layers. Ideally, this process needs more models as a sample but because of
many constraints, we just used just nine models. This research involved three main subjects, which are
a design that needs a mathematical formula to describe the profile and is generated to become 3D
high-speed train models, Computational fluid dynamics to compute coefficient drag model, and the
last Artificial Neural Network using the Backpropagation algorithm. The backpropagation algorithm is

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
International Conference on Science Education and Technology IOP Publishing
Journal of Physics: Conference Series 1511 (2020) 012074 doi:10.1088/1742-6596/1511/1/012074

prepared to substitute Computational Fluids dynamics that need high spec hardware, software, and
more time to the simulation process.
Table 1. Specification Hardware and Software.
Design Process aerodynamics Artificial Neural
Analysis Network Analysis
RAM 6 Gb 6 GB 6 GB
Processor Core i5 Core i7 Core i5
Capacity 1 TB 1 TB 1 TB
Software Inventor OpenFoam Matlab R2014a
OS Windows 7 Ubuntu Windows 7

2. Theoretical Background

2.1. Modeling Longitudinal High-Speed Train Nose Profile


In this research, we used two main mathematical formulas to model a longitudinal nose of a high-
speed train model. The formulas are ellipsoid formula and Hicks-Henne Function. The formulas are
shown respectively in equation (1) and (2).

x2 y2
 1 (1)
a2 b2

Where
a=length of the semi-major axis
b=length of the semi-minor axis
x=absis of coordinates of any point on the ellipse
y= ordinate of coordinates of any point on the ellipse

N
Y  Ybase  Wi f i (2)
i 1

Where 𝑌𝑏𝑎𝑠𝑒 is the baseline function.𝐹𝑖 and 𝑊𝑖 stand for the Hick-Henne shape function and the
weights factor, respectively. The Hick-Henne shape function 𝐹𝑖 is described in the following
Equations [1].

𝑥(1 − 𝑥)
𝐹1 = 0 < 𝑊1 < 0.12
𝑒 20𝑥

𝐹2 = sin(𝜋𝑥 0.25 )3 0.03 < 𝑊2 < 0.15

𝐹3 = sin(𝜋𝑥 0.5 )3 0.01 < 𝑊3 < 0.1

𝐹4 = sin(𝜋𝑥 0.8 )3 − 0.05 < 𝑊4 < 0.05

𝐹5 = sin(𝜋𝑥 1.357 )3 − 0.06 < 𝑊5 < 0.04

𝐹6 = sin(𝜋𝑥 3 )3 − 0.06 < 𝑊6 < 0.04 (3)

2
International Conference on Science Education and Technology IOP Publishing
Journal of Physics: Conference Series 1511 (2020) 012074 doi:10.1088/1742-6596/1511/1/012074

2.2. Computational Fluids Dynamics Analysis


Navier-Stokes equations are used to analyze aerodynamics phenomena like drag force, lift force,
aerodynamics moments, etc. There are no analytic solutions to solve the equations, so it is needed
Computational Fluids Dynamics to solve it numerically. Navier stokes equations consist of many
equations that describe conservation in continuity, momentum, and energy. Simplifying Navier-
Stokes equations in the general form is shown in equations (4) [2].
     
  U i      q (4)
t xi  xi 

In computational fluids dynamics, the Navier-Stokes equations are discretized so it can be solved
numerically using grid concepts.

Figure 1. Description of grid concepts in CFD simulations.

2.3. Backpropagation Algorithm


Backpropagation algorithms are a family of methods used to efficiently train artificial neural networks
(ANNs) following a gradient-based optimization algorithm that exploits the chain rule. The main
feature of backpropagation is its iterative, recursive and efficient method for calculating the weights'
update to improve the network until it can perform the task for which it is being trained.
In the human being, neural networks are activation function to process input signal which can be
modeled mathematically as Sigmoid bipolar activation and Sigmoid multipolar equation. In this
research we used Sigmoid bipolar because it is more accurate, that is shown in equation (5) [3]

1  e x
f ( x)  (5)
1  ex
Ending iteration process in Backpropagation algorithms is when the mean sum squared error is less
than tolerance value. Mean sum squared error is a parameter that shows the sum of the total deviation
between output and target in each iteration. Mean sum squared error is shown in equation (6)[4]
N
 T arg et  Output2
1
MSSE  (6)
N i 1

Where N is severaldata

3. Methods
Methods have been used in this research shown in figure 2.

3
International Conference on Science Education and Technology IOP Publishing
Journal of Physics: Conference Series 1511 (2020) 012074 doi:10.1088/1742-6596/1511/1/012074

Figure 2. Flowchart research

3.1. Design Model


The design of a high-speed train model was initialized from longitudinal nose high-speed train profile
models which consist of many coordinates that were derived from mathematical formulas. Three
mathematical formulas generate non-dimensional geometries. In each non-dimensional geometries
was variated in length, they are 3000 m, 5000m, and 7000 m. Based on 2D longitudinal nose high-
speed train models, the 3D models were drawn using inventor software. The head size of the high-
speed train model is 23 m x 3 m x 3 m (length x wide x height). However, the size of models are
completely similar but it can be variated into the different length of the nose and mathematical model
that construct nose.

4
International Conference on Science Education and Technology IOP Publishing
Journal of Physics: Conference Series 1511 (2020) 012074 doi:10.1088/1742-6596/1511/1/012074

Figure 3. Longitudinal Nose Profile

Figure 4. 2D Nose Modeling

Figure 5. 3D Nose Modeling

3.2. Computational Fluids Dynamics Analysis


Computational fluids analysis was used to compute coefficient drag in each model. In aerodynamics,
the parameter gives information about drag. Drag is caused by relative velocity between moving
objects and air molecules. When air molecules hit surface objects, it causes pressure drag and friction
between air molecule and the object's skin, that cause friction drag. The sum of two types of drag is a
total drag. Drag that is divided by multiplying results kinetic energy and surface is a coefficient drag.
So it could get two coefficient drags, that is coefficient drag pressure and coefficient drag friction. In
this research, coefficient drag was computed by computational fluids dynamics and simulated using
OpenFoam software.
One of the most important efforts to minimizing coefficient drag is by designing longitudinal
profile objects. In this research, we used three mathematical formulas to construct a high-speed train

5
International Conference on Science Education and Technology IOP Publishing
Journal of Physics: Conference Series 1511 (2020) 012074 doi:10.1088/1742-6596/1511/1/012074

nose in which each formula is used to construct the noise model which its length respectively are 3000
m, 5000 m, and 7000 m.

Figure 6. Description of drag due to pressure and friction

3.3. Backpropagation Analysis


Backpropagation analysis is one of the artificial neural network methods that adopt a human being
thinking process to find the pattern to identify the object. The objective of this research is to find
relation patterns between high-speed train coefficient drag and longitudinal two-dimensional
coordinates nose of high-speed train model. The pattern can be found using robust weights that
connect between input layers and hidden layers, and between hidden layers and output layers. Why
must the weight be robust? Because the backpropagation algorithm will substitute computational
fluids dynamics analysis to compute coefficient drag high-speed train. Ideally, backpropagation
analysis uses huge amounts of models but because of many constraints, we only used nine models.
Next time, the research will be planned using huge amount models and will be up to date that it will
use three-dimensional coordinates as input to compute high-speed train coefficient drag. The source
code of the Backpropagation algorithm is made by Matlab software and all computation involves
matrice operation.

4. Drawing Model and Computational Simulation


Drawing 3D head of high-speed train models was based on its longitudinal nose profiles. The
longitudinal nose profiles were generated from 2D coordinates that were extracted from mathematical
formulas which are variated in length in each formula. First, we determined constant value in each
formula that is shown in part 2.1 before. Based on the formula, we can plot its contour. Then the 2D
contour is generated to become the 3D surface of the head of the high-speed train. The contour is very
simple which is on the lateral side just follow a longitudinal nose profile. Actually, the nose of the
high-speed train is more complex. The longitudinal nose profiles, its formula, and the value of
constant are shown in Table 2.

6
International Conference on Science Education and Technology IOP Publishing
Journal of Physics: Conference Series 1511 (2020) 012074 doi:10.1088/1742-6596/1511/1/012074

Table 2. Mathematical Formulas Derive Longitudinal Nose of High-Speed Train Models


Length of
Model Mathematical Formula Constant Value
the nose(m)
Model 1 N
Y  Ybase  Wi f i f2 = 0.1 3
i 1

Model 2 N
Y  Ybase  Wi f i f2 = 0.1 5
i 1

Model 3 N
Y  Ybase  Wi f i f2 = 0.1 7
i 1

Model 4 N
f2 = 0.4 and f6=
Y  Ybase  Wi f i 0.1
3
i 1

Model 5 N
f2 = 0.4 and f6=
Y  Ybase  Wi f i 0.1
5
i 1

Model 6 N
f2 = 0.4 and f6=
Y  Ybase  Wi f i 0.1
7
i 1

Model 7
x2 y2 a = 3000
 1 3
a2 b2 b=1500

Model 8
x2 y2 a = 5000
 1 5
a 2
b 2 b=1500

Model 9
x2 y2 a = 7000
 1 7
a 2
b 2
b=1500

The nine models of the head of the high-speed train, which were converted to the obj file format,
were inputted to aerodynamics analysis using OpenFoam. Before the simulation was started, it must
be set up many aerodynamics parameters that are shown in table 3. The output of aerodynamics
analysis is the coefficient drag of each model that becomes the target of Backpropagation simulation
in the training phase to get robust weights.

7
International Conference on Science Education and Technology IOP Publishing
Journal of Physics: Conference Series 1511 (2020) 012074 doi:10.1088/1742-6596/1511/1/012074

Table 3. Aerodynamics Parameters


Parameter Value
Wind Speed 44.44 m/s
Frontal Area 9 m2
Air Density 1.225 kg/m3
Viscosity 1.789 x 10-5 Kg/m/s
Total Pressure 1.01 x 105 Pa

Coefficient drag values are fluctuating before they reach convergent values. We took a coefficient
drag from each model as the target of backpropagation simulation when it has been convergent.

Table 4. Architecture ANN Backpropagation


Characteristics Specification
Architecture 1 hidden layer
Input neuron 4000
Hidden Neuron 50
Input neuron 1
Activation Function Sigmoid
Initialization weights Random
MSSE Less than 10-8
Maximum Epoch 10000
To obtain robust weights between input layers and hidden layers and between hidden layers and
output layers using Backpropagation simulation, it must determine many parameters, which are shown
in table 4, that it is more like trial and error process.

5. Results and Analysis


In backpropagation analysis in this research using 4 configurations that are be variated in composition
number model which include training or exercise respectively. The robustness of weights each
configuration is shown in parameter MSSE. The results are shown in Table 5. Based on table 5 it can
be concluded that the more models simulated in training the more robust weights between input layers
and hidden layers and between hidden layers and output layers. It shows that the more models were
included in training the smaller MSSE in exercise. We hope the research will be continued using a lot
of models and to accommodate the complex shape of the head of the high-speed train and many
aerodynamics parameters, the profile of the nose must be in 3D coordinates.

Table 5. Configuration in Backpropagation Analysis


Model Training Excercise MSSE Excercise
Configuration I 9 models 9 models 6.73 x 10-9
Configuration II 8 models 1 model 8.92 x 10-7
Configuration III 7 models 2 models 8.92 x 10-5
Configuration IV 6 models 3 models 9.63 x 10-5

Because we only used nine models in this simulation, the time to get the convergent result was very
fast. It only took 10.000 epoch to reach the convergent result that is shown in figure 7, which the
iteration will be stoped if Mean Sum Square Error below than 10-8. The weakness of using a few
models is that the robustness is achieved if we use the same models for both training and exercise.

8
International Conference on Science Education and Technology IOP Publishing
Journal of Physics: Conference Series 1511 (2020) 012074 doi:10.1088/1742-6596/1511/1/012074

0,001
Mean Sum Square error

0,0008

0,0006

0,0004

0,0002

0
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Epoch
Figure 7. Mean Sum Square Error in configuration I.

Comparison between CFD analysis results and Backpropagation results in the configuration I am
shown in table 6.

Table 6. Coefficient Drag Analysis Comparison


Model Coefficient Drag Coefficient Drag Back
CFD Results Propagation Results
Model 1 0.78 0.779841
Model 2 0.68 0.680018
Model 3 0.65 0.650124748
Model 4 0.76 0.760261
Model 5 0.68 0.680088
Model 6 0.64 0.639936
Model 7 0.72 0.720002
Model 8 0.8 0.800002
Model 9 0.88 0.879989479

6. Conclusion
Based on results from aerodynamics and Backpropagation simulation, it can be concluded below.
1. Based on nine head of high-speed train models simulation using Backpropagation analysis, it is
concluded that this research can result in robust weights to compute coefficient drag, which inputs
are 4000 coordinates longitudinal nose of high-speed train model.
2. Based on aerodynamics analysis using OpenFoam software, it is known that in a similar shape the
longer the long nose of high-speed train the smaller the coefficient drag.
3. Backpropagation algorithm can be used to determine the coefficient drag of high-speed train, thus
tom guarantee robustness of weights, it must train using many models.

References
[1] Wang, Ruibin., Jianjun, Zhang., Shaojun Bianand., Lihuan, You. (2018).“A survey of
parametric modeling methods in high-speed train head design”.Journal of Rail and Rapid
Transit. Vol. 232(7), PP. 1965-1983.
[2] Bistafa, Sylvio R. (2018). “On the development of the Navier–Stokes equation by Navier”.
History of Physics and Related Sciences. Revista Brasileira de Ensino de F´ısica, vol. 40, nº
2, e2603 (2018).
[3] Julpan., Erna Budhiarti Nababan., Muhammad Zarlis. (2015). “ANALISIS FUNGSI
AKTIVASI SIGMOID BINER DAN SIGMOID BIPOLAR DALAM ALGORITMA
BACKPROPAGATION PADA PREDIKSI KEMAMPUAN SISWA”. Jurnal Teknovasi.
Volume 02, Nomor 1, 2015, PP. 103 – 116, ISSN: 2355-701X.

9
International Conference on Science Education and Technology IOP Publishing
Journal of Physics: Conference Series 1511 (2020) 012074 doi:10.1088/1742-6596/1511/1/012074

[4] Chai, T., Draxler, R.R. (2014). “Root mean square error (RMSE) or mean absolute error
(MAE)? – Arguments against avoiding RMSE in the literature”. Geoscientific Model
Development Discussion., 7, 1247–1250, 2014.
[5] Balaji, S. Arun. dan Baskaran, K. (2013). “Design and Development of Artificial Neural
Networking (ANN) System Using Sigmoid Activation Function to Predict Annual Rice
Production in Tamilnadu”. International Journal of Computer Science, Engineering and
Information Technology(IJCSEIT). Vol. 3. No.1.
[6] Anike, Marleni., Suyoto, dan Ernawati. (2012). “Pengembangan Jaringan Syaraf Tiruan dalam
Memprediksi Jumlah Dokter Keluarga Menggunakan Backpropagation (Studi Kasus :
Regional X Cabang Palu)”. Seminar Nasional Teknologi Informasi dan Komunikasi
(SENTIKA).
[7] Devi, C. J., Reddy, B. S. P., Kumar, K.V., Reddy, B. M., Nayak, N.R. (2012).
“ANNApproach for Weather Prediction using Back Propagation”. International
Journal of Engineering Trends and Technology. Vol. 3. Issue 1.
[8] Abd Alkader, Shahlla. A., Al-Allaf, Omaima. N. A. (2011). “Backpropagation Neural
Network Algorithm for Forecasting Soil Temperatures Considering Many Aspects:
A Comparison of Different Approaches”. The 5th International Conference on
Information Technology.
[9] Rebello, S., Maheshwari, U., Safreena, Souza, R. V. D. (2011). “Backpropagation Neural
Network Method for Predicting Lac Gene Structures in Streptococcus Pyogenes M Group A
Streptococcus Strains”. International Journal for Biotechnology and Molecular Biology
Research. Vol. 2 (4), PP.61-72.
[10] Gaurang Panchal, Amit Ganatra, Parth Shah, Devyani Panchal. (2011). “Determination Of
Over-Learning And Over-Fitting Problem In Back Propagation Neural Network”.
International Journal on Soft Computing (IJSC). Vol.2. No.2.

10

You might also like