Analysis of Microstrip Finger on Bandwidth of Interdigital Band Pass Filter using Artificial Neural Network
1 like444 views
This document discusses using artificial neural networks to estimate the bandwidth of an interdigital band pass filter based on variations in the finger length. An ANN model was developed using data from electromagnetic simulations of filters with finger lengths ranging from 34mm to 24mm. Both multi-layer perceptron and radial basis function networks were tested, with the RBF network providing more accurate results with a mean squared error of 1.13173e-005. The proposed ANN approach allows estimating the filter bandwidth without complex calculations and provides a fast design method for interdigital band pass filters.
![E-ISSN: 2349-7610
INTERNATIONAL JOURNAL FOR RESEARCH IN EMERGING SCIENCE AND TECHNOLOGY, VOLUME-1, ISSUE-3, AUGUST-2014
1
Analysis of Microstrip Finger on bandwidth of Interdigital band Pass Filter
using Artificial Neural Network
1
Robert Mark, 2
Vandana V. Thakare
1
Dept of Electronics, Madhav Institute of Technology and Science, Gwalior, India
1
markrobert7@gmail.com
2
Dept of Electronics, Madhav Institute of Technology and Science, Gwalior, India
2
vandanavt_19@rediffmail.com
ABSTRACT
In this paper, a novel method of bandwidth estimation using variation of finger length on interdigital band pass filter has been
presented using artificial neural networks for desired frequency range between 1.5--3.5GHz. Interdigital filter is multifinger
periodic structure which offers compact filter design space. An ANN model has been developed and tested for estimating the cut
off frequency of band pass filter and performance is evaluated in terms of mean square error and concluded that RBF network is
more accurate than MLPFFBP. The proposed method of design provides the exact bandwidth for particular finger length of filter
without using painstaking calculation.
Keyword - Artificial neural networks (ANN), Multi layer Perceptron feed Forward back propagation (MLPFFBP), Interdigital
capacitor (IDC), Mean square error (MSE), Radial Basis Function (RBF), Artificial neural network (ANN), Levenberg Marquardt
(LM).
1. INTRODUCTION
In modern communication systems, high performance and
small size band pass filters are essentially required to enhance
the system performance and to reduce cost of fabrication.
Existing Parallel coupled microstrip filters have been widely
used in the RF front end of microwave and wireless
communication systems for decades. But their large size was
problematic with the RF system where size is important
consideration [1-2]. To overcome existing drawback,
Interdigital capacitor (IDC) geometry is used in which gaps
meanders back and forth in a rectangular area forming two sets
of fingers that are interdigital. IDC is a multifinger periodic
structure which offer very high Q value in compact design
space with microstrip finger coupled between the input and
output ports across the gaps, hence increase performance of
filter. [3]
In present work, an attempt has been made to utilize the
potential of neural network to calculate the cut off frequency
of filter using data from design parameter. ANN model has
been established by learning from microwave data which
was acquired by EM simulation of filter design in CST
(computer simulation technology) Microwave environment
and results are used as standard to train the network through a
process called training. An ANN model based on Radial basis
function and feed forward back propagation algorithms are
used to train the network and determine Fu (upper cut off) and
Fl (lower cut off) and hence bandwidth (Fu - Fl) of filter for
desired frequency range of 1.5-3.5 GHz.
2. PROPOSED FILTER DESIGN AND DATA
GENERATION
Interdigital capacitor (IDC) is an element for producing high
pass characteristic using microstrip finger lines used for filter
design. Number of Fingers (N) provide coupling between the
input and output ports across the gaps. Typically, the gaps
between finger and end of the fingers are equal to reduce
parasitic capacitance. [3]
The length (L) and width (W) of the finger determined total
capacitance and capacitance per unit length along the width of](https://image.slidesharecdn.com/pid-13201401-140926134836-phpapp02/85/Analysis-of-Microstrip-Finger-on-Bandwidth-of-Interdigital-Band-Pass-Filter-using-Artificial-Neural-Network-1-320.jpg)
![E-ISSN: 2349-7610
INTERNATIONAL JOURNAL FOR RESEARCH IN EMERGING SCIENCE AND TECHNOLOGY, VOLUME-1, ISSUE-3, AUGUST-2014
2
the finger. The total capacitance of an interdigital structure(C)
expressed as:
(1)
For a finite substrate (ℇr), the effect of h (height of substrate)
must be included in A1 and A2. In the final design, usually S
(gap between finger) =W and, l ≤ ʎ /4. Approximate
expressions for A1 and A2 (design constant) are as [2, 5]:
A1= 4.409 tanh [ ] ×10-6
(2)
A2= 9.92 tanh [0.52( )] ×10-6
(3)
(4)
(5)
The filter with Number of Fingers (N) =10, was designed to
work for frequency range of 1.5-3.5 GHz with L=34mm
a n d W=S=2mm respectively. The length and width of the
filter design are calculated by the given equation (1), (2), (3),
(4) & (5).
As total capacitance value depends on length of finger strip
,Fig 1 shows the proposed geometry with finger length varied
between L = 34mm to L”= 24mm to extract response such that
it can be used to train the neural network and estimate the
bandwidth of filter. MATLAB neural tool box has been used
to calculate the cut off frequency of band pass filter i.e. Fu
(upper cut off frequency) and Fl (lower cut off frequency) for
different training algorithm and results are compared with EM
simulation on basis of mean square error of ANN.
Fig.1: Proposed Interdigital filter structure
3. NEURAL MODEL FOR FILTER
The ANN model developed for Microstrip band pass filter is
shown in Fig 2. The feed forward back propagation and radial
basis network has been utilized to calculate the lower (Fl) and
upper cut off frequency (Fu) of the band pass filter by
providing Length (L), Width (W) and substrate (ℇr) as input to
network. To analyze neural model for filter design 180
different inputs for desired frequency range of operation has
been utilized for training of neural model using Levenberg
Marquardt (LM) and radial basis function and 20 test patterns
are used to validate the accuracy of neural model for best
performance.
Fig-2: ANN model for Band pass filter
3.1. Multi Layer Feed Forward Back Propagation
(MLFFBP) Neural Network
MLP network are supervised networks, they required a
desired response to be trained. With one or two hidden layers
they can approximate virtually any input output map. The
weights of the network are usually computed by training the
network using the back propagation algorithm. [8] In present,
work three layers multilayer Perceptron feed-forward back
propagation neural network with one hidden layer are trained
by LM algorithm. The network is analysed with different
number of hidden layers and also the numbers of processing
elements are also varied to acquire the accuracy. It is
concluded that three layer
MLPFFBP with one hidden layer and 20 processing elements
in hidden layers is the optimum structure for proposed
problem. There are three input neurons in the input layer, 20
hidden neurons in the hidden layer and two output neurons in](https://image.slidesharecdn.com/pid-13201401-140926134836-phpapp02/85/Analysis-of-Microstrip-Finger-on-Bandwidth-of-Interdigital-Band-Pass-Filter-using-Artificial-Neural-Network-2-320.jpg)
![E-ISSN: 2349-7610
INTERNATIONAL JOURNAL FOR RESEARCH IN EMERGING SCIENCE AND TECHNOLOGY, VOLUME-1, ISSUE-3, AUGUST-2014
3
the output layer. The training of network performs in 71
epochs so as to reduce the average MSE close to 4.41e-05.
The network has been trained for different IDC finger
length but in a specified range (34mm-24mm). During the
training process the neural network automatically adjusts its
weights and threshold values such that the error between
predicted and sampled outputs is minimized [9]. The
adjustments are computed by the back propagation algorithm.
The transfer function preferred is tansig for neural network. In
table 1 lower cut off frequency obtained using CST software
and using LM algorithm for different test patterns are
compared and mean square error has been calculated.
Table 1: Comparisons of results obtained using CST and LM
algorithm on basis of mean square error for analysis of Band
pass filter.
Fig-3: Number of epochs to achieve minimum mean square
error level in case of MLFFBP with LM training algorithm
3.2. RBF Network
Radial basis function network is a feed forward neural
network with a single hidden layer that use radial basis
activation functions for hidden neurons. The RBF neural
network has both a supervised and unsupervised component to
its learning.
It consists of three layers of neurons—input, hidden and
output. The hidden layer neurons represent a series of center in
the input data space. Each of these centres has an activation
function, typically Gaussian [8-9].
In present work the spread value for RBF network was
chosen as 1.1, which gives the best accuracy. The network was
trained with 180 samples and tested with 20 samples. RBF
network with accuracy of 99.3% is more accurate compared to
MLPFFBP for filter design. The RBF network automatically
adjusts the number of processing elements in the hidden layer
till the defined accuracy is reached. Fig.4 shows the training
performance of the neural network model is trained in 175
epochs with MSE = 1.13173e-005. Table 2 shows results
obtained using CST software and RBF network for different
test patterns are compared.
Finger
length
(mm)
Freq lower
(GHz)
CST
Freq lower
(Ghz)
LM
MSE
(LM)
33.8 1.820 1.8087 0.000135
33.45 1.825 1.8219 1.44E-05
33.2 1.832 1.8277 2.02E-05
32.85 1.842 1.8351 5.93E-05
32.3 1.862 1.8642 4.41E-06
32 1.880 1.8939 0.00018
31.55 1.923 1.9331 9.8E-05
31.15 1.948 1.9506 3.61E-06
30.8 1.961 1.9589 4.84E-06
30.4 1.973 1.9671 4.62E-05
30.15 1.981 1.9732 6.24E-05
29.75 1.990 1.9862 1.68E-05
28.35 2.074 2.0809 4.36E-05
27.85 2.113 2.1248 0.000121
27.1 2.182 2.1797 5.76E-06
26.3 2.238 2.2354 9.61E-06
25.9 2.268 2.2703 2.25E-06
25.35 2.304 2.2994 2.4E-05
24.95 2.353 2.3589 2.7E-05
24.35 2.422 2.42 5.29E-06](https://image.slidesharecdn.com/pid-13201401-140926134836-phpapp02/85/Analysis-of-Microstrip-Finger-on-Bandwidth-of-Interdigital-Band-Pass-Filter-using-Artificial-Neural-Network-3-320.jpg)
![E-ISSN: 2349-7610
INTERNATIONAL JOURNAL FOR RESEARCH IN EMERGING SCIENCE AND TECHNOLOGY, VOLUME-1, ISSUE-3, AUGUST-2014
5
Madhav Institute of Technology and Science Gwalior, India in
preparation of this paper.
7. REFERENCES
[1]. R. S. Chen, X. Zhang, K. F. Tsang, and K. N. Yung
,“Modelling and design of Interdigital Capacitor based on
Neural Networks and Genetic algorithm.
[2]. Lumped Elements for RF and Microwave Circuits, Inder
Bahl.
[3] Interdigital Capacitor Design, “Agilent EEs of EDA”.
[4] W. P. du Plessis, “Interdigital Filter Design,” Microwave
and Optical Technology Letters, vol. 51, no. 10, pp. 2269-
2272, October 2009.
[5] Microstrip Filters for RF/Microwave Applications, “Jia-
Sheng Hong, M. J. Lancaster”.
[6] CST (Computer Simulation Technology) Microwave
Studio Suit Software 2010.
[7] MATLAB Software, version 7.10.
[8] Vandana Vikas Thakare, Pramod Singhal. “Artificial
Intelligence in the Estimation of Patch Dimensions of
Rectangular Microstrip Antennas”. Circuits and Systems,
2011, 2, 330-337.
[9] Vandana Vikas Thakare and Pramod Singhal, “Neural
network based CAD model for the design of rectangular patch
antennas”, Journal of Engineering and Technology Research
Vol. 2(7), pp. 126-129, July 2010.
[10] FitriYuli Zulkifli “Implementation of single cell
composite right–left handed transmission line for ultra wide
band pass filter”.](https://image.slidesharecdn.com/pid-13201401-140926134836-phpapp02/85/Analysis-of-Microstrip-Finger-on-Bandwidth-of-Interdigital-Band-Pass-Filter-using-Artificial-Neural-Network-5-320.jpg)
Analysis of Microstrip Finger on Bandwidth of Interdigital Band Pass Filter using Artificial Neural Network
- 1. E-ISSN: 2349-7610 INTERNATIONAL JOURNAL FOR RESEARCH IN EMERGING SCIENCE AND TECHNOLOGY, VOLUME-1, ISSUE-3, AUGUST-2014 1 Analysis of Microstrip Finger on bandwidth of Interdigital band Pass Filter using Artificial Neural Network 1 Robert Mark, 2 Vandana V. Thakare 1 Dept of Electronics, Madhav Institute of Technology and Science, Gwalior, India 1 markrobert7@gmail.com 2 Dept of Electronics, Madhav Institute of Technology and Science, Gwalior, India 2 vandanavt_19@rediffmail.com ABSTRACT In this paper, a novel method of bandwidth estimation using variation of finger length on interdigital band pass filter has been presented using artificial neural networks for desired frequency range between 1.5--3.5GHz. Interdigital filter is multifinger periodic structure which offers compact filter design space. An ANN model has been developed and tested for estimating the cut off frequency of band pass filter and performance is evaluated in terms of mean square error and concluded that RBF network is more accurate than MLPFFBP. The proposed method of design provides the exact bandwidth for particular finger length of filter without using painstaking calculation. Keyword - Artificial neural networks (ANN), Multi layer Perceptron feed Forward back propagation (MLPFFBP), Interdigital capacitor (IDC), Mean square error (MSE), Radial Basis Function (RBF), Artificial neural network (ANN), Levenberg Marquardt (LM). 1. INTRODUCTION In modern communication systems, high performance and small size band pass filters are essentially required to enhance the system performance and to reduce cost of fabrication. Existing Parallel coupled microstrip filters have been widely used in the RF front end of microwave and wireless communication systems for decades. But their large size was problematic with the RF system where size is important consideration [1-2]. To overcome existing drawback, Interdigital capacitor (IDC) geometry is used in which gaps meanders back and forth in a rectangular area forming two sets of fingers that are interdigital. IDC is a multifinger periodic structure which offer very high Q value in compact design space with microstrip finger coupled between the input and output ports across the gaps, hence increase performance of filter. [3] In present work, an attempt has been made to utilize the potential of neural network to calculate the cut off frequency of filter using data from design parameter. ANN model has been established by learning from microwave data which was acquired by EM simulation of filter design in CST (computer simulation technology) Microwave environment and results are used as standard to train the network through a process called training. An ANN model based on Radial basis function and feed forward back propagation algorithms are used to train the network and determine Fu (upper cut off) and Fl (lower cut off) and hence bandwidth (Fu - Fl) of filter for desired frequency range of 1.5-3.5 GHz. 2. PROPOSED FILTER DESIGN AND DATA GENERATION Interdigital capacitor (IDC) is an element for producing high pass characteristic using microstrip finger lines used for filter design. Number of Fingers (N) provide coupling between the input and output ports across the gaps. Typically, the gaps between finger and end of the fingers are equal to reduce parasitic capacitance. [3] The length (L) and width (W) of the finger determined total capacitance and capacitance per unit length along the width of
- 2. E-ISSN: 2349-7610 INTERNATIONAL JOURNAL FOR RESEARCH IN EMERGING SCIENCE AND TECHNOLOGY, VOLUME-1, ISSUE-3, AUGUST-2014 2 the finger. The total capacitance of an interdigital structure(C) expressed as: (1) For a finite substrate (ℇr), the effect of h (height of substrate) must be included in A1 and A2. In the final design, usually S (gap between finger) =W and, l ≤ ʎ /4. Approximate expressions for A1 and A2 (design constant) are as [2, 5]: A1= 4.409 tanh [ ] ×10-6 (2) A2= 9.92 tanh [0.52( )] ×10-6 (3) (4) (5) The filter with Number of Fingers (N) =10, was designed to work for frequency range of 1.5-3.5 GHz with L=34mm a n d W=S=2mm respectively. The length and width of the filter design are calculated by the given equation (1), (2), (3), (4) & (5). As total capacitance value depends on length of finger strip ,Fig 1 shows the proposed geometry with finger length varied between L = 34mm to L”= 24mm to extract response such that it can be used to train the neural network and estimate the bandwidth of filter. MATLAB neural tool box has been used to calculate the cut off frequency of band pass filter i.e. Fu (upper cut off frequency) and Fl (lower cut off frequency) for different training algorithm and results are compared with EM simulation on basis of mean square error of ANN. Fig.1: Proposed Interdigital filter structure 3. NEURAL MODEL FOR FILTER The ANN model developed for Microstrip band pass filter is shown in Fig 2. The feed forward back propagation and radial basis network has been utilized to calculate the lower (Fl) and upper cut off frequency (Fu) of the band pass filter by providing Length (L), Width (W) and substrate (ℇr) as input to network. To analyze neural model for filter design 180 different inputs for desired frequency range of operation has been utilized for training of neural model using Levenberg Marquardt (LM) and radial basis function and 20 test patterns are used to validate the accuracy of neural model for best performance. Fig-2: ANN model for Band pass filter 3.1. Multi Layer Feed Forward Back Propagation (MLFFBP) Neural Network MLP network are supervised networks, they required a desired response to be trained. With one or two hidden layers they can approximate virtually any input output map. The weights of the network are usually computed by training the network using the back propagation algorithm. [8] In present, work three layers multilayer Perceptron feed-forward back propagation neural network with one hidden layer are trained by LM algorithm. The network is analysed with different number of hidden layers and also the numbers of processing elements are also varied to acquire the accuracy. It is concluded that three layer MLPFFBP with one hidden layer and 20 processing elements in hidden layers is the optimum structure for proposed problem. There are three input neurons in the input layer, 20 hidden neurons in the hidden layer and two output neurons in
- 3. E-ISSN: 2349-7610 INTERNATIONAL JOURNAL FOR RESEARCH IN EMERGING SCIENCE AND TECHNOLOGY, VOLUME-1, ISSUE-3, AUGUST-2014 3 the output layer. The training of network performs in 71 epochs so as to reduce the average MSE close to 4.41e-05. The network has been trained for different IDC finger length but in a specified range (34mm-24mm). During the training process the neural network automatically adjusts its weights and threshold values such that the error between predicted and sampled outputs is minimized [9]. The adjustments are computed by the back propagation algorithm. The transfer function preferred is tansig for neural network. In table 1 lower cut off frequency obtained using CST software and using LM algorithm for different test patterns are compared and mean square error has been calculated. Table 1: Comparisons of results obtained using CST and LM algorithm on basis of mean square error for analysis of Band pass filter. Fig-3: Number of epochs to achieve minimum mean square error level in case of MLFFBP with LM training algorithm 3.2. RBF Network Radial basis function network is a feed forward neural network with a single hidden layer that use radial basis activation functions for hidden neurons. The RBF neural network has both a supervised and unsupervised component to its learning. It consists of three layers of neurons—input, hidden and output. The hidden layer neurons represent a series of center in the input data space. Each of these centres has an activation function, typically Gaussian [8-9]. In present work the spread value for RBF network was chosen as 1.1, which gives the best accuracy. The network was trained with 180 samples and tested with 20 samples. RBF network with accuracy of 99.3% is more accurate compared to MLPFFBP for filter design. The RBF network automatically adjusts the number of processing elements in the hidden layer till the defined accuracy is reached. Fig.4 shows the training performance of the neural network model is trained in 175 epochs with MSE = 1.13173e-005. Table 2 shows results obtained using CST software and RBF network for different test patterns are compared. Finger length (mm) Freq lower (GHz) CST Freq lower (Ghz) LM MSE (LM) 33.8 1.820 1.8087 0.000135 33.45 1.825 1.8219 1.44E-05 33.2 1.832 1.8277 2.02E-05 32.85 1.842 1.8351 5.93E-05 32.3 1.862 1.8642 4.41E-06 32 1.880 1.8939 0.00018 31.55 1.923 1.9331 9.8E-05 31.15 1.948 1.9506 3.61E-06 30.8 1.961 1.9589 4.84E-06 30.4 1.973 1.9671 4.62E-05 30.15 1.981 1.9732 6.24E-05 29.75 1.990 1.9862 1.68E-05 28.35 2.074 2.0809 4.36E-05 27.85 2.113 2.1248 0.000121 27.1 2.182 2.1797 5.76E-06 26.3 2.238 2.2354 9.61E-06 25.9 2.268 2.2703 2.25E-06 25.35 2.304 2.2994 2.4E-05 24.95 2.353 2.3589 2.7E-05 24.35 2.422 2.42 5.29E-06
- 4. E-ISSN: 2349-7610 INTERNATIONAL JOURNAL FOR RESEARCH IN EMERGING SCIENCE AND TECHNOLOGY, VOLUME-1, ISSUE-3, AUGUST-2014 4 Fig-4 Number of epochs to achieve minimum mean square error level in RBF Network Table 2: Comparisons of results obtained using CST and RBF on basis of mean square error for analysis of Band pass filter 4. RESULT The paper provides the effects of variation of finger length on bandwidth of interdigital band pass filter. It has been observed that the both Fu & Fl cut off frequency start shifting toward high frequency end by reducing the finger length during simulation, which provides better understanding that for given filter design by reducing length of finger strip, we can shift operating frequency towards higher frequency. The mean absolute percentage error (MAPE) for LM training is 1.512 % whereas for RBF network is 0.6532%, which shows that experimental results generated by radial basis method are more fast and accurate. Figure-5 shows comparison of LM and RBF algorithm based on mean square error of output cut off frequency and observed that RBF algorithm provide low value of performance goal indicating that trained ANN model is an accurate model for filter design. Fig-5: Graph showing comparison of Average mean square error for Levenberg-Marquardt and RBF ANN Algorithm 5. CONCLUSION An efficient and time saving method for calculation of bandwidth of filter was discussed using ANN. The results obtained with the present technique are closer to the experimental results generated by simulating interdigital filter using CST software on the FR4 (Lossy) substrate. The paper concludes that results obtained using present ANN techniques are quite satisfactory and followed the experimental drift and also that RBF is giving the best approximation to the design as compared to MLFFBP. In the analysis, one can select the exact value of finger length to obtain desire set of cut off frequency over working frequency range without using laborious calculation. 6. ACKNOWLEDGMENT Authors acknowledge the support and valuable guidance received from Dr. V. V. Thakare, Department of Electronics, Finger length (mm) Freq lower (GHz) CST Freq lower (GHz) RBF MSE (RBF) 33.8 1.820 1.812 0.00016 33.45 1.825 1.8221 3.14E-05 33.2 1.832 1.8327 1.21E-06 32.85 1.842 1.8425 8.41E-06 32.3 1.862 1.8614 0.000016 32 1.880 1.8833 1.52E-05 31.55 1.923 1.9226 8.1E-07 31.15 1.948 1.946 9E-08 30.8 1.961 1.9623 4.9E-07 30.4 1.973 1.9739 8.1E-07 30.15 1.981 1.9797 6.4E-07 29.75 1.990 1.9921 9E-08 28.35 2.074 2.0726 6.25E-06 27.85 2.113 2.1176 8.41E-06 27.1 2.182 2.1823 5.76E-06 26.3 2.238 2.2382 2.21E-05 25.9 2.268 2.27 1.68E-05 25.35 2.304 2.2996 3.84E-05 24.95 2.353 2.3594 1.3E-05 24.35 2.422 2.4205 2.92E-05
- 5. E-ISSN: 2349-7610 INTERNATIONAL JOURNAL FOR RESEARCH IN EMERGING SCIENCE AND TECHNOLOGY, VOLUME-1, ISSUE-3, AUGUST-2014 5 Madhav Institute of Technology and Science Gwalior, India in preparation of this paper. 7. REFERENCES [1]. R. S. Chen, X. Zhang, K. F. Tsang, and K. N. Yung ,“Modelling and design of Interdigital Capacitor based on Neural Networks and Genetic algorithm. [2]. Lumped Elements for RF and Microwave Circuits, Inder Bahl. [3] Interdigital Capacitor Design, “Agilent EEs of EDA”. [4] W. P. du Plessis, “Interdigital Filter Design,” Microwave and Optical Technology Letters, vol. 51, no. 10, pp. 2269- 2272, October 2009. [5] Microstrip Filters for RF/Microwave Applications, “Jia- Sheng Hong, M. J. Lancaster”. [6] CST (Computer Simulation Technology) Microwave Studio Suit Software 2010. [7] MATLAB Software, version 7.10. [8] Vandana Vikas Thakare, Pramod Singhal. “Artificial Intelligence in the Estimation of Patch Dimensions of Rectangular Microstrip Antennas”. Circuits and Systems, 2011, 2, 330-337. [9] Vandana Vikas Thakare and Pramod Singhal, “Neural network based CAD model for the design of rectangular patch antennas”, Journal of Engineering and Technology Research Vol. 2(7), pp. 126-129, July 2010. [10] FitriYuli Zulkifli “Implementation of single cell composite right–left handed transmission line for ultra wide band pass filter”.