Adaptive Network-Based Inference System Models On Multiband Patch Antenna Design
Adaptive Network-Based Inference System Models On Multiband Patch Antenna Design
Adaptive Network-Based Inference System Models On Multiband Patch Antenna Design
http://journals.tubitak.gov.tr/elektrik/
c TUB
doi:10.3906/elk-1301-159
Research Article
GLU
2,,
Erdem DEMIRC
IO
, Murat H
usn
u SAZLI1 , Orhan S
ENGUL
3
1
S
ehabeddin Taha IMECI , Hakk Alparslan ILGIN
1
Department of Electrical and Electronics Engineering, Ankara University,
Tando
gan, Ankara, Turkey
2
R & D Department, Prime Ministry of Turkey, Ankara, Turkey
3
K
ucu
kyal, Istanbul, Turkey
Received: 22.01.2013
Accepted: 05.03.2013
Printed: 09.02.2015
Abstract: In this study, an adaptive network-based fuzzy inference system (ANFIS) for multiband microstrip patch
antenna (MMPA) modeling is proposed. The MMPA includes a single patch with inverted L-shaped stubs on the edges.
The antenna production process needs iterative runs of the electromagnetic (EM) simulator, fabricating and testing the
simulated antenna to find the optimum geometry. Production processes (design, simulation, fabrication, and testing)
and the cost computation time add to the overall manufacturing expenses. A computer-aided modeling approach using
the ANFIS is developed and trained with data acquired from the EM simulators, Sonnet Suites and AWR AXIEM, and
measurement data. The overall data set includes 7777 input-output pairs, of which 3338 are used for the testing of the
ANFIS model. The trained ANFIS model provides accurate and reliable results compared with the EM simulators and
measured data.
Key words: Adaptive network-based fuzzy inference system, multiband microstrip patch antennas, EM simulator,
modeling
1. Introduction
Microstrip patch antennas (MPAs) have advantageous properties, such as having a low profile, light weight,
conformal structure, low design cost, and ease of integration with other microwave circuits and devices [14].
Although MPAs have extensive applications in military and commercial fields, the narrow bandwidth of less
than 5% compared to the operating frequency is a major problem to be solved. Broadband and multiband
MPAs (MMPAs) are proposed to overcome the disadvantage of operating with a narrow bandwidth. MMPAs
are realized by placing a half-wavelength shorted stub or quarter-wavelength open-circuit stub on the MPA
edges. The stub offers capacitive impedance below the resonance frequency and inductive impedance above the
resonance frequency of the MPA and causes a multiresonance response [58].
MPAs have a 2-dimensional radiating patch on a thin dielectric substrate and can therefore be categorized
as 2-dimensional planar components for analysis purposes. The analysis methods can be grouped into 2 groups.
In the first group, the transmission line model (TLM), cavity model (CM), and multiport network model
(MNM) are based on the equivalent magnetic current distribution around the patch edges, similar to slot
Correspondence:
118
GLU
DEMIRC
IO
et al./Turk J Elec Eng & Comp Sci
antennas. In the second group, the full-wave numerical analysis methods, method of moments, finite-element
method (FEM), spectral domain technique, and finite-difference time-domain method are derived from the
electric current distribution on the patch conductor and ground plane, similar to dipole antennas [1].
In this study, the modeling of an inverted L-shaped stubbed MMPA using an adaptive network-based fuzzy
inference system (ANFIS) model is presented. The ANFIS model is trained with various training algorithms
and optimization methods to achieve the highest accuracy. The proposed computer-aided modeling aims to
avoid iterative design processes, which increase the computation time and fabrication expenses. For instance,
microwave engineers must redesign and simulate the entire structure repeatedly once a tuning process is applied
to the device under focus. As the device gets complicated, these production processes (design, simulation,
fabrication, and testing) carry intense loads in terms of time, engineering resources, and cost. The proposed
ANFIS model avoids repetitive simulations and fabrications once a comprehensive data library that includes
simulated and measured data is accomplished. Consequently, the computer-aided design models assist in
diminishing the overall production costs.
The proposed model can be used to analyze all types of configurations that cannot be handled by the
TLM due to variation of field in the orthogonal direction to the direction of propagation. In the CM and
MNM, the arbitrary shapes cannot be accurately analyzed and numerical techniques are applied for complex
geometries. However, numerical methods carry a high computation load and time to solve electromagnetic (EM)
problems. The proposed model can be applied to complex geometries through an affordable computation time
compared with numerical methods.
GLU
DEMIRC
IO
et al./Turk J Elec Eng & Comp Sci
the simulated and measurement results. The proposed antenna shape of the MMPA is depicted in Figure 1.
c
W =
2fr
r + 1
2
)1/2
,
[
]1/2
r + 1 r 1
12h
ef f =
+
1+
,
2
2
W
(
)(
)
ef f + 0.3
W/h + 0.264
l = 0.412h
,
ef f 0.258
W/h 0.8
L=
c
2l,
2fr ef f
Rin =
ZoT =
452o
,
W2
Rin 50.
(1)
(2)
(3)
(4)
(5)
(6)
Here, W and L denote the width and length of the rectangular patch, respectively. ef f is the effective dielectric
constant, l denotes the uncertainty of the patch length, Rin is the input resistance of the patch, and ZoT is
the characteristic impedance of the microstrip feed line.
sw
GLU
DEMIRC
IO
et al./Turk J Elec Eng & Comp Sci
microstrip feed line is attached to the feeding line to simply solder the connector. The quarter-wave feed lines
are selected to achieve the highest possible gain values at the resonant frequencies. The antenna gain can be
adjusted for a fixed resonance frequency as the quarter-wave line alters.
GLU
DEMIRC
IO
et al./Turk J Elec Eng & Comp Sci
LAYER 1
LAYER 4
MF11
Die le ctric
c ons ta nt
(r)
LAYER 2
Inputs
LAYER 3
r1
MF12
w1
.
.
.
.
.
.
MF13
.
.
.
Inputs
MF21
Fre que ncy
(fr)
wn
MF22
MF23
S W to
thickne s s
ra tio
(s wtr)
MF32
LAYER 5
rnFn
rn
S UM
R e turn
l os s
.
.
.
.
.
.
.
.
.
MF31
r1F1
Inputs
r27F27
r27
w27
N
MF33
Dielectric consta nt
Stub width to substrate
thickness ratio (sw/h)
Freque ncy
ANFIS
MODEL
Scattering
parameters
(x)2
2 2
(7)
GLU
DEMIRC
IO
et al./Turk J Elec Eng & Comp Sci
The proposed ANFIS model is trained for 3 to 8 MFs and 1000 epochs. The number of rules is 512 (8 8 8)
for 8 MFs. The Gaussian MF is specified by 2 parameters; therefore, the ANFIS contains a total of 2096 fitting
parameters, where 48 (8 2 + 8 2 + 8 2) are the premise nonlinear parameters and 2048 (4 512) are
the consequent linear parameters. There are a total of 1078 nodes on the proposed ANFIS model. In Figure 5,
the estimated and actual measurements and the simulated return losses are depicted versus the frequency range.
The ANFIS models the return loss more accurately than the EM simulator, because the training set is acquired
among the measurement data. Since the number of rules and parameters increases as the number of MFs is
enhanced, the total ANFIS modeling time and required memory also increase with the accuracy. The modeling
accuracy can also be augmented by sampling additional training data around the resonance frequencies. Since
the substrate is thin and the MPAs naturally have narrow bandwidths, the sharp discontinuity in the resonance
frequencies causes errors in the modeling. This can be handled by taking more samples in these regions as the
simulator or measurement device capability allows.
RTd 5880 ANFIS estimation
0
5
10
15
20
25
ANFIS
MEASURED
AXIEM
30
35
8
10
Frequency (GHz)
12
14
The frequency, dielectric constant, and stub width to the substrate height are selected as the input
parameters of the proposed model, in which the input-output relation must be clearly defined. As new input
parameters are inserted into the model, it becomes complicated and is tougher to train. For instance, if the stub
number is taken as a variable parameter and altered, the number of resonances changes. Next, the modeling
problem converts to an optimization problem. In this study, a total of 8 inverted L-shaped stubs are attached to
the antenna. The proposed ANFIS model can be applied to 2-, 4-, and 6-stubbed antennas under a fixed stub
number and physical shape. This study aims to model as diverse as possible antennas with various substrates
and antenna dimensions to avoid repetitive production processes.
Figure 6 illustrates the antenna gains for some resonant frequencies. Nine resonance points are determined
among the overall frequency ranges, of which 4 provide sufficient gain and cross polarization. This still satisfies
the proposed multiband characteristic of the MPA. The cross polarization values are around 40 dB for each
frequency. The computed gain values alter between 6.11 dB and 11.44 dB.
The Table presents the ANFIS return loss outputs for some resonant frequencies and the computed error
percentage compared with the desired test outputs, which include both the simulation and measurement results.
The modeling accuracy can be improved as the number of sampling points around the resonant frequency is
123
GLU
DEMIRC
IO
et al./Turk J Elec Eng & Comp Sci
boosted. When the ANFIS optimization algorithms are compared, the hybrid learning provides better estimation
results than the back propagation method, which introduces 158.8% more error.
30
Gain (dB)
10
2.82 GHz
5.81 GHz
6.66 GHz
11.3 GHz
0
-60
60
-10
-20
-30
-90
90
Theta angle (deg)
Table. Comparison of ANFIS modeling outputs vs. the desired test values (* indicates the measured results).
Antenna parameters
(mil)
L
W
sw
h
950
790 115 31
950
790 115 31
950
790 115 31
950
790 115 31
700
570 120 20
700
570 115 20
700
570 110 20
700
570 105 20
1070 850 120 60
1070 850 120 60
1070 850 115 60
1070 850 110 60
570
470 65
70
570
470 65
70
570
470 70
70
570
470 75
70
950
730 100 125
950
730 105 125
950
730 110 125
950
730 115 125
Average percentage error
124
r
2.2
2.2
2.2
2.2
3.02
3.02
3.02
3.02
2.94
2.94
2.94
2.94
1.2
1.2
1.2
1.2
2.2
2.2
2.2
2.2
Resonant
frequency (GHz)
fr
2.8288
5.8213
6.6
11.2988
3.9575
8.3063
11.5088
13.1888
4.5175
7.8950
11.7188
14.7375
8.2538
9.89
12.3660
13.5630
4.3775
4.9988
6.7925
8.8575
ANFIS
12.3076
16.9678
13.6498
16.3448
17.1126
17.9980
17.6923
11.5628
13.8534
24.0645
13.6765
20.9589
16.2210
29.2283
11.9101
18.3192
11.0353
22.5825
12.7177
24.0371
Error
percentage (%)
ANFIS
4.76
11.55
2.10
10.32
4.97
4.78
2.50
7.49
12.52
9.05
0.76
1.58
6.30
7.77
2.39
4.30
0.15
2.10
0.62
3.65
4.98
GLU
DEMIRC
IO
et al./Turk J Elec Eng & Comp Sci
3. Conclusions
An ANFIS-based modeling technique was proposed to estimate the scattering parameters of MMPAs with
various substrates and dimensions. The ANFIS results were compared with the desired test outputs containing
the EM simulator and measurement results. The suggested method provides faster, lower-cost, highly accurate,
and reliable results, which proves that the complex nonlinear relations between the input and output parameters
in the field of antenna computer-aided design can be handled using ANFIS models. As the training data sets
get larger and statistically have enhanced sampling on the intended parameter ranges, the ANFIS modeling
capacity increases. ANFIS models can be easily adapted to various RF microwave modeling problems to shorten
the overall production processes.
References
[1] Kumar G, Ray KP. Broadband Microstrip Antennas. New York, NY, USA: Artech House, 2003.
[2] Kumar G, Ray KP. Tunable and dual band circular microstrip antenna with stubs. IEEE T Antenn Propag 2000;
48: 10361039.
[3] Garg R, Bhartia B, Bahl I, Ittipiboon A. Microstrip Antenna Design Handbook. New York, NY, USA: Artech
House, 2001.
[4] Deshmukh AA, Ray KP. Multiband rectangular microstrip antennas. IETE J Res 2011; 57: 437442.
[5] Daniel AE. Tunable dual band rectangular microstrip antennas and their arrays. PhD, Indian Institute of Technology, Bombay, India, 2006.
[6] Lu JH. Single feed dual frequency rectangular microstrip antenna with pair of step slots. Elect Lett 1999; 35:
354355.
[7] Ray KP, Kishna DD. Compact dual band suspended semicircular microstrip antenna with half u-slot. Microwave
Opt Tech Lett 2006; 48: 20212024.
[8] Zhong SS, Lo YT. Single element rectangular microstrip antenna for dual frequency operation. Elect Lett 1983; 19:
298300.
[9] Jang JSR. ANFIS: Adaptive-network-based fuzzy inference system. IEEE Trans Syst Man Cybern 1993; 23: 665
685.
[10] G
uney K, Sarikaya N. Adaptive neuro-fuzzy inference system for computing the resonant frequency of electrically
thin and thick rectangular microstrip antennas. Int J Elect 2007; 94: 833844.
[11] G
uney K, Sarikaya N. Multiple adaptive-network-based fuzzy inference system for the synthesis of rectangular
microstrip antennas with thin and thick substrates. Int J RF Microw Comput-Aid E 2008; 18: 359375.
[12] G
uney K, Sarikaya N. Adaptive-network-based fuzzy inference system models for input resistance computation of
circular microstrip antennas. Microw Opt Tech Lett 2008; 50: 12531261.
[13] G
uney K, Sarikaya N. Concurrent neuro-fuzzy systems for resonant frequency computation of rectangular, circular,
and triangular microstrip antennas. Prog Electromagnetics Res 2008; 84: 253277.
[14] G
uney K, Sarikaya N. Comparison of adaptive-network-based fuzzy inference systems for bandwidth calculation of
rectangular microstrip antennas. Exp Syst Appl 2009; 36: 35223535.
[15] T
urker N, G
unes F, Yldrm T. Artificial neural design of microstrip antennas. Turk J Electr Eng Co 2006; 14:
445453.
[16] Imeci T, Keskin N. Tri-band microstrip patch antennas. Int J RF Microw Comp-Aided Eng 2012; 22: 7584.
125