APPLICATIONS OF ARTIFICIAL INTELEGENCE TECHNIQUES IN MICROSTRIP

COMPONENT AND ANTENNA DESIGN AT JADAVPUR UNIVERSITY IN THE

PRESENT DECADE

B. Gupta, S. Chakraborty, P. Mukherjee and S. Biswas

E.T.C.E. Dept., Jadavpur University, Raja S.C. Mullick Road, Kolkata-700032, India
E-Mail: [email protected], [email protected]

Abstract: Artificial Intelligence Techniques viz. Artificial Neural Networks (ANN) and Genetic
Algorithms (GA) etc. are well-known in the field of soft computing to solve computationally intensive
problems. Researches at Jadavpur University by the authors focussed on applying these techniques to
overcoming the limitations of microstrip components and antennas in terms of several useful parameters
like band width, power handling capacity etc. Factors of influence experienced and innovations made in
these recent works are enlisted in brief.

1. INTRODUCTION resonant length of a patch antenna due to this

fringe effect is estimated by the proposed neural
Neural networks have recently drawn net model which helps to determine its
significant attention as powerful tools in artificial operational frequency accurately. Further, any
intelligence (AI) research. Their structural and open circuit type discontinuity in a microstrip
behavioral resemblance to systems of biological transmission line would have introduced stray
neurons has also proved useful in a wide variety radiation by the same mechanism, thereby
of engineering applications. In this paper, we affecting the system performance adversely.
Hence, another model is proposed for estimation
wish to exploit the advantages of using neural
of effective length extension of microstrip lines
networks for estimation of fringing effect in
at points of discontinuity [1].
microstrip antenna and open ended microstrip
The model consists of multi-layer perceptons
lines, for array failure correction and for design (with two hidden layers) where each hidden layer
and analysis procedure of microstrip lines. contains 40 hidden nodes and hyperbolic tangent
Genetic Algorithm is one of the fast growing function is used as activation function. The
search techniques in engineering problem thickness and dielectric constant of the substrate
solving in last 25 years. It is based on an idea of and characteristic impedance and operating
natural evolution, a genetic mechanism of the frequency of the corresponding microstrip
living organisms in the earth, which may be the transmission line form the four input nodes for
secret of development of life. It shows robust the model used to estimate the effect of
and highly efficient optimization capability in discontinuities in microstrip transmission lines.
various engineering fields and applications. In For the antenna model, all the input nodes are
addition, it has proven ability to address identical excepting the impedance node, which is
multimodal multi-object problems. Here we use absent. The output layer contains a single node
GA to design the feed for an innovative planar showing the corresponding length extension
dual band antenna with polarization diversity. It value.
is also used in detecting failed elements in large
antenna arrays.

2. RESEARCH WORK CARRIEDOUT

In this section, some of our different research
and development work performed are reported
under five subsections.

2.1 Development of neural net models for

fringing field at microstrip open circuits
When a microstrip patch radiates, it must be Fig. 1 Schematic view of the proposed neural net
from the fringing field at the open ended edges. model
The incremental enhancement of effective

978-1-4244-1864-0/07/$25.00 ©2007 IEEE

Authorized licensed use limited to: National Institute of Technology- Delhi. Downloaded on February 05,2021 at 11:23:01 UTC from IEEE Xplore. Restrictions apply.
 Relations obtained from standard literature and (8.3-8.7GHz) as compared to a maximum of
computer simulations are used to generate the 4.5% reported in the literature.
training data in all cases. After sufficient number
of iterations using few hundred sets of training 5

4.5

data, the back propagation algorithm used to 4

train both networks saturate and fairly accurate

3.5

Axial ratio(dB)
3

estimations for length extension are obtained 2.5

when the networks are tested with unknown test 1.5

data. 0.5

0
8.25 8.3 8.35 8.4 8.45 8.5 8.55 8.6 8.65 8.7 8.75 8.8

frequency(GHz)

Fig. 4 Measured axial ratio vs. frequency plot in

X band

The experimentally measured impedance

bandwidth for VSWR<1.5 is around 1.2 % (8.3-
8.4 GHz) with a very narrowband matching in C
band (VSWR ~2 at 4.312GHz). Further, for the
Fig. 2 Graph for the estimation of the length linearly polarized C band radiation, cross
increment due to fringe effect in microstrip line. polarization is found to be unusually low,
As pointed out already, accurate but fast typically 25dB down, as compared to 12-15 dB
estimation of the effective length extension is for standard patch antennas.
required for determining the resonant frequency This antenna may find exhaustive use in
of patch antennas. Also, similar estimation of communication systems and radars, where we
effective length at points of discontinuity in need dual frequency operation. Due to its planar
microstrip lines can help in compensating for the structure, this antenna offers less aerodynamic
power loss using properly designed stubs. These drag and easy integrability with MIC and
tasks are facilitated by the neural net model MMICs. Also because of its light weight, it can
developed. be easily used for airborne or spaceborne
applications. The structure of the antenna can be
2.2 Development of dual band coplanar modified to achieve different polarization
microstrip antenna with polarization diversity characteristics at the two frequency bands.
A new coplanar microstrip antenna has been
developed without stacking of elements for dual 2.3 Artificial neural network models for
band operation viz. in C and X bands [2]. The analysis and design of microstrip lines
antenna is intended to be operated as a circularly
polarized antenna in the X band and as a linearly Artificial neural networks (ANN) have been
polarized antenna in the C band. Very good formulated in the frequency range of 0.5GHz to
impedance matching has been achieved in both 30GHz to analyze and synthesize the impedance
bands with the help of an optimal feed network and dispersion characteristics of microstrip
designed using genetic algorithm. The antenna transmission lines and verified against software
with the feed is shown in fig.3. The outer ring MWI and experiments. Results show a good
acts both as a C band radiator and feed to the X parity for wide range of parameters [3,4].
band patch whose polarization may be controlled
varying the locations of connection between the
two radiating elements. W

r h
h=thickness of the dielectric material.
W= strip width.
εr=relative dielectric constant.

Figure 5 A typical microstrip line

Fig. 5 describes the line parameters and fig. 6
Fig.3 Structure of the (a) dual band antenna
shows the neural net for its analysis. To
(b) matching network
minimize error for the vast range of input
parameters three hidden layers are used. The
Fig. 4 shows measured axial ratio plot in X
network is trained with 950 data sets for w/h =
band. The 3dB axial-ratio bandwidth is ~ 4.8%
0.05 to 5 with the help of Backpropagation

Authorized licensed use limited to: National Institute of Technology- Delhi. Downloaded on February 05,2021 at 11:23:01 UTC from IEEE Xplore. Restrictions apply.
 Algorithm. After nearly 700,000 iterations with a 10
for h = 1.575mm. and w = 3.15mm. by MWI

variable learning rate of 0.00013 to 0.01, the 9

8 εr = 9
by ANN

effective relative
final network values yields an average of
7

permittivity.
6
5
εr = 6

absolute error of less than 2% with respect to 4

3
εr =3
MWI. 2
1

Almost the same network as shown in fig. 6 0

0 5 10 15 20 25 30 35

has been used for synthesis. In this case the value frequency in GHz.

of four inputs are proportional to the dielectric Fig. 8 Plot of effective relative permittivity vs.
thickness, characteristic impedance of the strip, frequency in GHz for different εr
relative dielectric constant and frequency of For thickness (h) = 1.575mm. and Zo = 30Ω By MWI

operation and a single output is proportional to 12

10
By ANN

Strip width (w) in mm.

εr = 3
the value of width of the strip. For normalization, 8

6 εr = 6
each variable is divided by the same factor as 4
εr = 9

before except the strip width, which is divided 2

0

by 15. This network is also trained with 950 data 0 5 10 15

Frequency in GHz.
20 25 30

sets for w/h = 0.05 to 5 with the help of

Backpropagation Algorithm. Here the range of Fig. 9 Plot of strip width vs. frequency in GHz
the strip width is taken to lie within 0.2mm. to for different εr to achieve a given Z0
13mm. After nearly 500,000 iterations with a It is seen from the plots that the results
variable learning rate of 0.00013 to 0.03, the obtained using ANN are in excellent agreement
final network values yields an average of with results obtained using MWI. It is shown
absolute error of less than 2% with respect to that once ANNs are trained, then for different
MWI. value of input variables, within the specified
range, outputs may be obtained for both analysis
f(x)= 1/{ 1+ exp(-x)} and design purpose without further mathematical
calculation with a very small error.

2.4 Element failure correction in antenna

arrays using soft computing
Failure of some elements in an antenna array
degrades the radiation pattern significantly. A
technique, based on soft computing, has been
proposed for correction of element failure in
microstrip antenna arrays [5]. The network is
made to adjust the current through remaining
elements to get back the original radiation
pattern as far as practicable.
An asymmetric Hopfield network with self
loops in equilibrium is used to describe the array.
As some of the nodes get disconnected
representing failure of corresponding elements,
equilibrium is regenerated by supervised learning
Fig. 6 Four input two output three hidden- using the original radiation pattern.
layered feedforward network for analysis of
microstrip lines f(x) f(x)

f(x) f(x)

f or h =1.575mm. and w =3.15mm. by MWI

f(x) f(x)
by ANN
80
70 εr =3 f(x) f(x)
60
50 εr = 6
40 εr = 9 f(x)
f(x)
30
20
10 f(x) f(x)

0
f(x) f(x)
0 10 20 30 40 f(x)

f r equency in GHz.

f(x)=1/[1+ exp(-λx)]

Fig. 7 Plot of characteristic impedance in ohm Fig. 10 Asymmetric Hopfield network with
vs. frequency in GHz. for different εr self loop for fifteen units

Authorized licensed use limited to: National Institute of Technology- Delhi. Downloaded on February 05,2021 at 11:23:01 UTC from IEEE Xplore. Restrictions apply.
 The uncorrected array factors with and without
failed elements and the corrected one are plotted
and compared in fig. 11 for a 31×31 element
array with 75 missing elements.
0
-90 -75 -60 -45 -30 -15 0 15 30 45 60 75 90

angle θ in degree
-10

-20
Table 1 Parameters selected for GA
-30

The results indicate that GA provides a more

efficient and accurate alternative to direct search
-40

methods to detect failed elements in a large

-50
array factor in dB

-60

-70
antenna array. Maximum benefit of the technique
-80
can be obtained for real life problems involving
very large arrays, since the possible number of
----- without missing elements combinations of failed elements would have
 with missing elements before correction been extremely large in those cases. For
(a) example, failure of 4 elements in a 32 X 32 array
can be due to 4.698721671e+10 different
0
-90 -75 -60 -45 -30 -15 0 15 30 45 60 75 90

angle θ in degree

combinations. A direct computation in such

-10

cases would have been extremely time and

-20

computation resource consuming.

-30

-40

-50

-60
3. CONCLUDING REMARKS
A brief overview of the recent works
array factor in dB

-70

-80
performed at Jadavpur University on research
and development on application of Artificial
--------without missing elements
Intelligence Techniques in microstrip technology
 with missing elements after correction
has been presented. However, much more work
(b)
has been carried out and are being performed
Fig.11 Plot of array factor with 75 missing
presently. All cannot be reported within the small
elements for 31×31 array with and without extent of this communication. Of the works
correction described also, many useful data and figures
Testing of the proposed technique is carried carrying important information had to be omitted
out successfully in case of severe element failure due to space constraint.
in very large arrays. Such large problem spaces
aere chosen deliberately to ascertain the REFERENCES
effectivity and robustness of the proposed [1] Samik Chakraborty, Ayona Sarkar, Tathagata
scheme. Chakraborty and Bhaskar Gupta, “Neural network model
for the estimation of fringe effect in open-ended microstrip
lines and microstrip patch antennas”, International
2.5 Element failure detection in microstrip Conference on Microwaves, Antenna, Propagation and
antenna arrays using genetic algorithm Remote Sensing, Jodhpur, 2006.
In this work, element failure detection is [2] P. Mukherjee, B. Gupta and R. Bhattacharjee, “Dual
treated as an optimization problem. The failed band coplanar microstrip antenna with polarization
diversity”, Journal of Electromagnetic Waves and
element(s) is/are detected by minimizing the
Applications, vol. 17, No 9, 2003, pp.1323-1330.
difference between the degraded pattern and the [3] S. Biswas, P.P. Sarkar and B.Gupta, “Artificial neural
pattern generated by the optimal constrained network model for microstrip dispersion characteristics”,
solution obtained through genetic algorithm Proc. 2002 IEEE Symposium on Antennas & Propagation,
San Antonio, Texas, pp.448-451.
(GA) [6]. [4] S. Biswas, P.P. Sarkar and B.Gupta, “An artificial
The method is applied for planar arrays with neural network model for analysis and design of
two or more elements failed. Table 1 gives the microstrip lines”, Indian Journal of Physics, vol. 78 B(2),
parameters selected in the process. The selection April 2004, pp.239-243.
[5] S. Biswas, P.P. Sarkar and B.Gupta, “Array factor
of chromosomes was made according to
correction using neural network model”, International
tournament selection criterion. The inter-element Journal of Electronics, vol. 91(5), pp. 301-308,
spacing chosen is 0.9λ (in both x and y May 2004.
directions) and the radiation patterns are [6] Pinaki Mukherjee, Bhaskar Gupta and K. Yasumoto,
specified between -90° to +90° at intervals of 1°. “Element failure detection in antenna arrays using genetic
algorithm”, Proc. International Symposium on Antennas
In every case, the failed elements are detected and Propagation(ISAP 2007), Niigata, Japan, pp. 330-
correctly without any mistake. 333.

Authorized licensed use limited to: National Institute of Technology- Delhi. Downloaded on February 05,2021 at 11:23:01 UTC from IEEE Xplore. Restrictions apply.