GRNN For Forecasting Resonance Frequency of Circular Patch Antenna

Science Academy
Publisher
International Journal of Research and Reviews in Soft and Intelligent Computing (IJRRSIC)
Vol. 1, No. 2, June 2011
ISSN: 2046-6412
Copyright Science Academy Publisher, United Kingdom
www.sciacademypublisher.com

GRNN for Forecasting Resonance Frequency of Circular Patch
Antenna
Shuwen Chen
1
and Quan Hua
2
1
School of Information Science and Engineering, Southeast University, Nanjing, P. R. China
2
School of Information and Communication Engineering, Nanjing University of Post and Telecommunication, Nanjing, P. R. China
Email: chenshuwen@126.com
Abstract Microstrip patch antennas have been studied extensively. There is a close relationship between circular patch
antenna geometrical parameters and its resonance frequency. In this work, the Generalized Regression Neural Network (GRNN)
is implemented to forecast resonance frequency of circular patch antenna according to its radius parameter. After training, the
GRNN can predict resonance frequencies of circular patch antennas successfully with satisfying accuracy, which have different
radius parameters in a local region. The experimental result demonstrates the high degree of accuracy of the GRNN forecasting
method.

1. Introduction
Traditional analytical methods for the resonance frequency
of circular patch antenna include theoretical computation [1-3]
and actual measurement. With the development of computer
technology, a number of numerical methods (time domain or
frequency domain) have been presented, such as MOM, FEM,
BEM, FDTD, to apply to solve microwave radiation and
scattering problem. These methods usually consume much
computing time due to large computing quantity and storage
quantity. However, using empirical formula or looking-up
table methods may limit accuracy of processing problem. So,
the Generalized Regression Neural Network (GRNN) method
is presented to simulate nonlinear relationship between circular
path antenna radius and its resonance frequency.
As a new kind of computing model, artificial neural
network (ANN) has two advantages which traditional
numerical methods do not have:
- A perfect non-linear mapping performance.
- A lower demand for much empirical knowledge of the
modeled objects.
The GRNN network has some strong advantages in
approximation quantity, classification quantity and learning
speed compared with BP network and RBF network [7-9]. The
network finally converges to optimal regression surface of
samples. Furthermore, the prediction effect is satisfying even
lacking in sample data, too. In addition, the GRNN method
also can handle the instable data.
The structure of the rest of this paper is organized as
follows: next section 2 gives detailed description of the general
principles of GRNN; experiments in section 3 demonstrate the
effectiveness of the GRNN model, final section 4 is devoted to
conclusions.

2. General Principle of GRNN
2.1. Basic Structure of GRNN
GRNN [4, 5] consists of a radial basis function network
layer and a linear network layer, shown in figure 1.

Figure 1. Basic structure of GRNN neuron

In this network, P represents input vector; R represents the
input dimension of input vector; S represents the neurons
number of each network layer and the number of training
samples; b
1
represents the threshold of hidden layer and output
layer do not have a threshold factor. The product of and
b
1
use the symbol in radial basis layer.
2.2. The Principle of GRNN Model
The basis function of hidden layer nodes in network takes
Gaussian function,
2 2
2
( )
i
x c
R x e
o
= (1)
as the network transfer function, where represents the
smooth factor [6], which is the only parameter of GRNN
network to be adjusted artificially. The learning process of
network primarily depends on the data sample.
dist
i
o
International Journal of Research and Reviews in Soft and Intelligent Computing (IJRRSIC) 24
The GRNN network is a kind of forward neural network
[7], which can be divided into four layers: the input layer,
pattern layer, summation layer and output layer, shown in
figure 2. The input and output vectors are
1 2
[ , , , ]
T
n
x x x x = and
1 2
[ , , , ]
T
n
y y y y = ,
respectively.

Figure 2. Topology of GRNN network

The number of neurons of input layer is equal to the
dimension m of input vector of training samples. The input
layer directly delivers the various elements of the input vector
to the pattern layer. The weight value function is
( )
2
1
1
1
, 1, 2, ,
R
i ji
j
i
dist P w j S
=
= =
(2)
The number of neurons of pattern layer is equal to the
quantity n of learning samples. Each neuron corresponds to
different sample. The transfer function of neuron i of pattern
layer is
( ) ( )
2
, 1, 2, ,
2
T
i i
i
x x x x
P i m
o

= =

(3)
where
i
x is the learning sample corresponding to neuron i of
pattern layer. The number of neurons of output layer is equal
to the dimension l of output variable of learning samples. The
final output value of the jth neuron of output layer results from
(4):
, 1, 2, ,
j
j
D
S
y j l
S
= = (4)
The theoretical base of GRNN is nonlinear regression
analysis. The joint probability density function of variable x
and y is ( ) , f x y , and when the measured value x is X , the
regression of y with respect to X is:
( )
( )
,
[ / ]
,
y f X y dy
Y E y X
f X y dy
+
= =

(5)
Where
Y is the prediction output of Y when input is X .

3. Experiment and Result
3.1. Modeling
In this work, the dimension parameters of circular path
antenna are shown in figure 3.
Parameter R represents the radius of the circular patch; h is
the thickness of the path; d donates the distance between the
circular patch center and joint of feed line and circular patch;
and
r
c is the dielectric constant of substrate.

Patch
Substrate
PEC
Ground
R
d
h
r
c
Feed Point

Figure 3. A circular patch antenna model
3.2. Acquiring Training Data
According to different values of parameter d, h,
r
c and
R, corresponding resonance frequencies as desired output
value could be got through HFSS software.
3.3. Test Result of Case 1
In numerical case 1, parameter d and h are equal to 9.5mm
and 5mm, respectively,
r
c is 3.38, radius R varies from 16mm
to 26mm, a total of 21 corresponding resonance frequencies as
training data could be got, shown in Table 1. In table 2, another
22 test samples, which are different from the training data, are
selected to confirm the effectiveness of GRNN network.

Table 1. Training Data
No.
R
(mm)
Resonance
Freq. (GHz)
No.
R
(mm)
Resonance
Freq. (GHz)
1 16 2.9 12 21.5 2.1
2 16.5 2.75 13 22 1.95
3 17 2.66 14 22.5 1.9
4 17.5 2.64 15 23 1.95
5 18 2.53 16 23.5 1.9
6 18.5 2.45 17 24 1.85
7 19 2.4 18 24.5 1.83
8 19.5 2.35 19 25 1.8
9 20 2.3 20 25.5 1.75
10 20.5 2.25 21 26 1.7
11 21 2.15

Table 2. Test Data
No.
R
(mm)
Resonance
Freq. (GHz)
No.
R
(mm)
Resonance
Freq. (GHz)
1 16.2 2.85 12 21.7 2.0
2 16.7 2.75 13 22.2 2.05
3 17.2 2.67 14 22.7 1.95
4 17.7 2.6 15 23.2 1.92
5 18.2 2.5 16 23.7 1.9
6 18.7 2.45 17 24.2 1.85
7 19.2 2.4 18 24.7 1.8
8 19.7 2.3 19 25.2 1.78
9 20.2 2.25 20 25.7 1.75
10 20.7 2.19 21 26.2 1.7
11 21.2 2.14 22 26.7 1.68

The test data is shown in Table 2, and the curve of the
prediction compared with test data is shown in figure 4. In
addition, relative frequency error curve is shown in figure 5.
The relative frequency errors of test samples are all less than
5%, and the maximum absolute value of relative frequency
errors is only 4.95%, which satisfies the real industrial need.

16 18 20 22 24 26
1
1.5
2
2.5
3
3.5
Radius of Circular Patch (mm)
F
r
e
q
u
e
n
c
y

(
G
H
z
)

Training data from simulation
Compared test data from simulaton
Predicted value from GRNN output
Extrapolation

Figure 4. Comparing simulation and GRNN output results

16 18 20 22 24 26 28
-0.05
-0.04
-0.03
-0.02
-0.01
0
0.01
0.02
0.03
0.04
0.05
Radius of Circular Patch (mm)
R
e
l a
t
i v
e

F
r
e
q
u
e
n
c
y

E
r
r
o
r

Relative Frequency Error

Figure 5. Relative frequency error curve
3.4. Test Result of Case 2
In numerical case 2, parameter R is changed to 20mm and
other parameters are as same as those of case 1. The distance d
varies from 4mm to 18mm, and a total of 18 corresponding
resonance frequencies as training data could be got, shown in
table 3. In table 4, another 15 test samples are used to verify
capability of GRNN network.

Table 3. Training data
No.
d
(mm)
Resonance
Freq. (GHz)
No.
d
(mm)
Resonance
Freq. (GHz)
1 4 2.09 8 11 2.31
2 5 2.2 9 12 2.35
3 6 2.21 10 13 2.37
4 7 2.22 11 14 2.4
5 8 2.24 12 15 2.41
6 9 2.28 13 16 2.44
7 10 2.29 14 18 2.45

Table 4. Test data
No.
d
(mm)
Resonance
Freq. (GHz)
No.
d
(mm)
Resonance
Freq. (GHz)
1 4.5 2.15 9 12.5 2.34
2 5.5 2.2 10 13.5 2.36
3 6.5 2.2 11 14.5 2.4
4 7.5 2.24 12 15.5 2.41
5 8.5 2.26 13 16.5 2.42
6 9.5 2.28 14 17.5 2.45
7 10.5 2.31 15 18.5 2.46
8 11.5 2.31

Comparison result and relative frequency error curve are
shown in figure 6 and 7. The relative frequency errors of test
samples are all less than 1.2%, and the maximum absolute
value of relative frequency errors is only 1.06%.
According to the aforementioned analysis of two cases, the
simulation values of test data and GRNN prediction are very
close to each other. The GRNN network has good
generalization ability, and the error meets the requirement of
prediction accuracy. Therefore, GRNN can be used to forecast
resonance frequency of circular patch antenna.

4 6 8 10 12 14 16 18 20
2.05
2.1
2.15
2.2
2.25
2.3
2.35
2.4
2.45
2.5
2.55
Distance between Circular Patch Center and Feed Line (mm)
F
r
e
q
u
e
n
c
y

(
G
H
z
)

Training data frome simulation
Compared tested data from simulation
Tested data from GRNN output

Figure 6. Comparing simulation and GRNN output results

4 6 8 10 12 14 16 18 20
-6
-4
-2
0
2
4
6
8
10
12
x 10
-3
Distance between Circular Pathc Center and Feed Line (mm)
R
e
l
a
t
i
v
e

F
r
e
q
u
e
n
c
y

E
r
r
o
r

Relative frequency error

Figure 7. Relative frequency error curve

Conclusion
Based on the analysis of influencing factors of resonance
frequency of circular patch antenna, this paper proves the
effectiveness and accuracy of GRNN network in predicting
resonance frequency of circular patch antenna. The mechanism
of GRNN model is simple, even for multi-dimensional input
variables; it behaves excellent smooth characteristic.
Moreover, the number of parameters in GRNN network need
to be adjusted is only a smooth factor, relatively smaller
compared with those of BP neural network, , Therefore, the
GRNN network can find its optimal weights/biases more easily.
The future work focuses on applying the GRNN model to other
industrial applications, including the stock price prediction [8],
pattern classification [9], and image processing.

References
[1] Kumar G, Ray K P, Broadband Microstrip Antennas, Dedhanm:
Artech House, 2003.
[2] Bahl I J, BHartia, Microstrip Antennas, Dedham: Artech House, 1980.
[3] Balanis C A, Antenna Theory-Analysis and Design, New York: John
Wiley & Sons, 1997.
[4] Specht D F, A general regression neural network, IEEE J. Trans
Neural Networks, 1991(2), pp. 568-576.
[5] Specht D F, The general regression neural network rediscovered, IEEE
Trans Neural Networks, 1993(6), pp. 1033-1034.
[6] Specht D F, Romsdahl H, Experience with adaptive probabilistic
neural networks and adaptive general regression neural networks [C],
in Proceedings of the IEEE World Congress on Computational
Intelligence, Orlando, USA, 1994, pp. 1203-1208.
[7] Yudong Zhang, Lenan Wu Weights Optimization of Neural Network
via Improved BCO Approach [J], Progress in Electromagnetics
Research, 2008(83), pp.185-198.
[8] Yudong Zhang, Lenan Wu, Stock Market Prediction of S&P 500 via
combination of improved BCO Approach and BP Neural Network [J],
Expert systems with applications, 2009, 36(5), pp.8849-8854.
[9] Yudong Zhang, Shuihua Wang, Lenan Wu, A Novel Method for
Magnetic Resonance Brain Image Classification based on Adaptive
Chaotic PSO [J], Progress in Electromagnetics Research. 2010, 109:
325-343.

Shuwen Chen Ph.D in electromagnetic field and
microwave technology, Southeast University, Nanjing,
China. Major field of study is in microwave and
millimeter wave theory and technology.

Quan Hua Master in information and communication
engineering, Nanjing University of Post and
Telecommunication, Nanjing, China. Major field of study
is in signal and information processing.

GRNN For Forecasting Resonance Frequency of Circular Patch Antenna

Uploaded by

Copyright:

Available Formats

GRNN For Forecasting Resonance Frequency of Circular Patch Antenna

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

GRNN For Forecasting Resonance Frequency of Circular Patch Antenna

Uploaded by

Copyright:

Available Formats

Science Academy

Y is the prediction output of Y when input is X .

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.