Você está na página 1de 5

ISSN 2249-6343

International Journal of Computer Technology and Electronics Engineering (IJCTEE)


Volume 2, Issue 4, August 2012

A Low Pass FIR Filter Design Using Genetic


Algorithm Based Artificial Neural Network
Sheenu Thapar, Parminder Kaur, Neha Aggarwal

(3)Realization: This involves converting the transfer


Abstract— For the design of Low pass FIR filters complex function obtained in 2 into suitable filter network or structure.
calculations are required to calculate the filter coefficients. (4)Analysis of finite word length effects: Here, we analyze
Mathematically, by substituting the values of pass band,
the effects of quantizing the filters coefficients and input data
transition width, pass band ripple, stop band attenuation and
sampling frequency in any of the methods like window method, as well as the effect of carrying out the filtering operation
frequency sampling method or optimal method we can get the using fixed word lengths on the filter performance.
values of filter coefficients h(n). (5)Implementation: This involves producing the software
In this paper a low pass FIR filter has been designed using code and/or hardware and performing the actual filtering.
artificial neural network. The optimization of the network has
been done using genetic algorithms. The proposed approach has
been compared with Kaiser window method. The result shows
that the filter designed using ANN optimized with GA requires
lesser iterations for the performance goal meeting.

Index Terms—FIR filter, ANN, GA

I. INTRODUCTION
With the technological evolution, great advances have been
made on design techniques for various digital filters. A filter
is essentially a system or network that selectively changes the
wave shape amplitude -frequency and or phase – frequency
characteristics of a signal in a desired manner. A digital filter
is a mathematical algorithm implemented in hardware and/or
software that operates on a digital input signal to produce a
digital output signal for the purpose of achieving a filtering
objective. A simplified block diagram of a real-time digital
filter, with analog input and output signals is given in Fig.1.

Fig.2: Design of digital Filter

Finally, because there is a tradeoff between filter complexity


and implementation feasibility, complexity and
Fig.1: Block diagram of digital filter implementation feasibility, complexity is a performance
criteria. Ideal filter characteristics are practically
II. FIR FILTER DESIGN unrealizable.
The design of a digital filter involves five steps: There are many methods to design FIR filter which are:-
(1)Filter Specification: This may include stating the type of (1) Fourier series method
filter, for example low pass filter, the desired amplitude. (2) Frequency Sampling method
and/or phase response and the tolerances (if any) we are (3) Window method
prepared to accept, the sampling frequency and the word
length of the input data. The most of these design techniques suffer from some kinds
(2)Coefficient calculation: At this step, we determine the of drawback, Some of them could not give optimal design in
coefficients of transfer function, H(z), which will satisfy the any sense, some is lacking of generality, some need long
specifications given in 1. Our choice of coefficient calculation computing time, and so on (Bagachi and Mitra, 1996).
method will be influenced by several factors, the most Kaiser window method has been used because of the presence
important of which are the critical requirement in step 1 of ripple parameter beta.

99
ISSN 2249-6343
International Journal of Computer Technology and Electronics Engineering (IJCTEE)
Volume 2, Issue 4, July 2012

III. KAISER WIDOW Step 2: By using ann.m code the following are data base
Kaiser determined empirically that the value of β need to file in xls format has been prepared.
achieve a specified value of A is given by (a) db new_input.xls – Input data information in matrix
0.1102  A  8.7  for A  50 form
 (b) db new_target.xls –Target information in matrix form
  0.5842  A  21  0.07886  A  21
0.4
for 21 A  50 Step 3: Using ann2.m code the training of neural network
0.0 for A  21 has been done.

a)Preparation of data sheet :-
Recall that the case β = 0 is the rectangular window for
With following that are
which A = 21.
(1) Stop Band attenuation (SBT)
(2) Transition width (TW)
Further more, to achieve prescribed values of A and df, M
(3) Pass Band Ripple (PBR)
must satisfy equations
(4) Sampling Frequency (SF)
(5) Filter Length (N)
 A 8 Filter coefficient are calculated and in this thesis works is
 1 for A  21
M  14.36df carried out using approximately 30 such values of all the
 0.922 / df   1 for A  21
 above parameters to calculate the filter coefficients. The
range of different parameters has been taken which are:
Finite Impulse response filters (Öner and paper., 1999) are (a) SBT: (10.1040– 9.3860)
preferred for their stable and linear phase characteristics. But (b) TW: (50 -100)
due to long impulse response of FIR filters there will be more (c) PBR: (0.1- 0.2)
hardware complexity. (d) N: (3–10)
A. Formulation of Problem (e) PBF : (150- 180)
The design of digital filter means basically finding the Using this data set the Artificial Neural Network has been
values of filter coefficients so that given filter specification trained and can be use to calculate filter coefficients for input
are achieved the window based design method are exclusively parameters in this range.Now, ANN is use to design the low
used for calculating there coefficient. We have used Kaiser pass FIR filter.There is very less difference in the ann results
window for this purpose. The Kaiser window function goes and the calculated results by using Kaiser window method
some how in overcoming the incorporating a ripple control
parameter, ANNs have been used for the design of digital Step 4:Now the ANN has been optimized using GA to obtain
filter with pass band edge frequency, transition width, pass the performance goal in lesser iterations.
band ripple, stopband attenuation, sampling frequency as
input parameters. In this paper a low pass FIR filter has been IV. ARTIFICIAL NEURAL NETWORKS
designed using artificial neural network. The optimization of An Artificial Neural Network is an information processing
the network has been done using genetic algorithms. Here the paradigm inspired by the way the densely inter-connected,
multi layer perception feed forward network has been used for parallel structure of the mammalian brain process
the design because this method is efficient, accurate, less information. ANN have successfully applied to a number of
complex and easily implemented. The network has been problems including the identification and control of
trained in such a manner so that the error comes minimum, dynamical systems, communications networks, coordination
means there may of robotics handeye movements (Hintz and Spoffed, 1998). It
be very less difference in the results comes from actual is also referred to as a neuromorphic system, follows
calculations that has been come from matlab and the output connectionist architecture, and parallel distributed
comes from trained artificial neural network. processing. Artificial neural networks are collections of
B. Objective mathematical models that emulate some of the observed
The objectives of the present work are divided into the properties of biological nervous system and draw on the
following sections: analogies of the observed properties of biological nervous
(1) To prepare the data sheet using different values of filter system and draw on the analogies of adaptive biological
parameter achieve the filter coefficient. learning. The key element of the ANN is the novel structure of
(2) Using ANN a low pass FIR filter has been designed with the information processing systems. It is composed of a large
GA as the optimization tool, such that its coefficient match number of highly inter-connected processing elements that
with coefficients obtained with window method. are analogous to neurons and are tied together with weighted
C. Methodology connections that are analogous to synapses. Neural networks,
The major steps are followed in the execution of this work. with their remarkable ability to derive meaning from
Step 1: Development of codes in Mathab for the following. complicated or imprecise data, can be used to extract patterns
(a) ann.m – For data preparation. and detect trends that are too complex to be noticed by either
(b) ann2.m – For training of the neural network humans or other computer techniques.

100
ISSN 2249-6343
International Journal of Computer Technology and Electronics Engineering (IJCTEE)
Volume 2, Issue 4, August 2012
A trained neural network can be thought of as an “expert” Evolutionary algorithms form a subset of evolutionary
in the category of information it has been given to analyse. computation in that they generally only involve techniques
This expert can then be used to provide projections in given implementing mechanisms inspired by biological evolution
new situations of interest. Traditionally, in ANN models, the such as reproduction, mutation, recombination, natural
input at each synapse is a constant value and the activation selection and survival of the fittest.
function operates on weighted sum of inputs.
Some other advanrtages of ANN are as under: VI. RESULTS
Adaptive learning: An ability to learn how to do tasks based The network has been trained using Multilayer Perceptron in
on the data given for training or initial experience. which Error Back Propagation Algorithm has been
Self-Organisation: An ANN can create its own organization specifically used to design Low Pass FIR filter.Using
or representation of the information it receives during “Levenberg –Marquardt”(trainlm) in the neural network
learning time. feedforward (newff) the goal meet condition has been
Real Time Operation: ANN computations may be carried achieved as shown in fig.3.
out in parallel, and special hardware devices are being
designed and manufactured which take advantage of this
capability.
Fault Tolerance via Redundant Information Coding:
Partial destruction of a network leads to the corresponding
degradation of performance. However, some network
capabilities may be retained even with major network
damage.

V. GENETIC ALGORITHMS
Genetic Algorithms (GAs) are derivative free stochastic
optimization methods based loosely on the concepts of natural
selection and evolutionary processes. GAs are parallel search
procedures that can be implemented on parallel processing
machines for massively speeding up their operation [13].
Genetic algorithms belong to the larger class of evolutionary
algorithms (EA), which generate solutions to optimization Fig.3: Training of Artificial Neural Network
problems using techniques inspired by natural evolution, such
as inheritance, mutation, selection, and crossover. In a genetic Input parameters:-Transition width=50Hz,Sampling
algorithm, a population of strings (chromosomes), which Frequency=1KHz,Pass band Ripple=0.1dB,Filter
encode candidate solutions (called individuals, creatures, or Length=3,Passband=150Hz,Stopband Attenuation =
phenotypes) to an optimization problem, evolves toward 10.1040dB.
better solutions. The solutions are represented in binary as Kaiser Vs ANN
strings of 0s and 1s, but other encodings are also possible. The
evolution usually starts from a population of randomly h(n) Kaiser Artificial
generated individuals and happens in generations. Window Neural Network
Method
In each generation, the fitness of every individual in the h(0) 0.2885 0.2805
population is evaluated, multiple individuals are h(1) 0.4231 0.4206
stochastically selected from the current population (based on h(2) 0.2885 0.2667
their fitness), and modified (recombined and possibly h(3) 0.0000 0.0004
randomly mutated) to form a new population. The new h(4) 0.0000 0.0200
population is then used in the next iteration of the algorithm. h(5) 0.0000 0.0068
Commonly, the algorithm terminates when either a maximum h(6) 0.0000 0.0182
number of generations has been produced, or a satisfactory h(7) 0.0000 0.0191
fitness level has been reached for the population. If the
h(8) 0.0000 0.0031
algorithm has terminated due to a maximum number of
h(9) 0.0000 0.0075
generations, a satisfactory solution may or may not have been
h(10) 0.0000 0.0092
reached. Evolutionary computation uses iterative progress,
h(11) 0.0000 0.0047
such as growth or development in a population. This
h(12) 0.0000 0.0000
population is then selected in a guided random search using
parallel processing to achieve the desired end. Such processes Table 1: Kaiser Vs ANN
are often inspired by biological mechanisms of evolution.

101
ISSN 2249-6343
International Journal of Computer Technology and Electronics Engineering (IJCTEE)
Volume 2, Issue 4, July 2012

Fig.6: Performance Plot

Fig 4. Error Graph

Figure shows that the ann response is better than the actual
response and the error is reduced to a large amount.
The Input Data Set taken is TW =52, SF = 1.3 , PBR = 0.12,
N=4, PBF=151,SBT=9.9413.
The ANN has been optimized using GA. The results obtained
are shown in Fig 4,5 & 6

Fig. 7: Regression Plot

Fig 5 : Neural Network Training


Fig.8: Training State

102
ISSN 2249-6343
International Journal of Computer Technology and Electronics Engineering (IJCTEE)
Volume 2, Issue 4, August 2012
VII. CONCLUSION [14] I feachor Emmonual C. and Jervis Barriel W. (2001) “Digital Signal
Processing,” A Practical Approch,” Person Education (Singapore)
The present thesis has illustrated the need of Artificial Neural Ltd., 2001, Second Edition.
Network for the design of low pass FIR Filters rather than [15] Jou Yue-Dar. and Chen Fu-Kun., (2007) “Least square Design of FIR
Filters based on a compacted Feedback Neural Network”, Proceeding
using the complex calculative Window method..
of the IEEE Transaction on Circuits and systems, vol. 54 issue 5, May,
The following conclusions are drawn from this research 2007, pp. 427–431.
work:
(1) Artificial Neural Network is better and easy method of [16] D. Suckley, (1991) “ Genetic Algorithm in the design of FIR filters” IEE
Proceedings in Circuits, Devices & Systems, Vol. 138, issue 2, pp.
Design of low Pass FIR Filter. 234-238.
(2) Using Fourier series, Frequency sampling or Window
methods the filter can be design but for each unknown AUTHOR’S PROFILE:
parameter the filter coefficients have to calculated. In Sheenu Thapar is currently Assistant Professor in ECE Department,
comparison with ANN, the trained network can calculate the Continental Institute of Engineering & Technology, Fatehgarh Sahib,
140407,India. (e-mail: sheenukuk@gmail.com).
Parminder Kaur is currently Assistant professor in ECE Department, Desh
filter Coefficient for unknown parameter in that specified Bhagat Engineering College, Mandi Gobindgarh, 147301, India. (e-mail:
range. parminderkaurece@gmail.com).
(3) The error graphs shows that results come from Kaiser
Neha Aggarwal is currently Assistant professor in ECE Department, Desh
window method and artificial neural network is almost same.
Bhagat Engineering College, Mandi Gobindgarh, 147301, India. (e-mail:
(4) The ANN optimized with GA is meeting the Er.neha_14@yahoo.com).
performance goal in just 7 iterations

REFERENCES
[1] Aggarwal Varun., Jin O Wesley and May O Una., (2006) “Filter
approximation using Explicit Time and Frequency Remain
specification”. Proceeding, of the Annual Symposium on Artificial
Intelligence, 2006, Seattle, Washington, pp. 174–165.
[2] Aktan Mustaf,Yurdakul Arda and Dundar Gunhan (2008),”An
Algorithm for the design of Low –power hardware –Efficient FIR filter
IEEE transaction on circuit and systems vol 55 No.,2008,pp
1536-1545.
[3] C. Y. Lin, M. Wu, J. A. Bloom, I. J. Cox, and M. Miller, “Rotation,
scale, and translation resilient public watermarking for images,” IEEE
Trans. Image Process., vol. 10, no. 5, pp. 767-782, May 2001.
[4] A. Fizelow., P. Brites., A. Ochoa., H. Mertuns., E. Fernandez, and
Garcia-Martinez R. (2007) “Finding Optimal Neural Network
Architecture using Genetic Algorithms,” Proceeding of the advances in
computer and Engineering, 2007, pp. 15–24.
[5] Bhattacharya D. and Antoniou A., (1996) “Real Time Design of FIR
Filters of Feedback Neural Network”, Vol. 3, 1996, pp 1070.
[6] Durrani A. Yasur and Riesgo Teresa., (2007) “LUT-based Power
Macro modeling Technique for DSP architectures”, Proceeding of the
IEEE, Centro de Electronica Industrial, Spain, August, 2007, pp.
1416–1419.
[7] Englihart K.B., Hudgins B. S., Stevenson M. and Parker P. A., (1994)
“Myoelectric Signal Classification using a Finite Impulse Response
Neural Network”. Technical Report, The university of New
Brunswork, Canada, June, 1994, pp. 803–820.
[8] Fan Z. and Mars P., (1997) “Access flow control scheme for ATM
networks using neural-network-based traffic prediction”. Proceeding
for IEEE, Vol. 144, Dec. 5, October 1997, pp. 708–714.
[9] Garimella Rama Marthy. (2008) “Finite Impulse Response FIR filter
Model of Synapses: Associated Neural Network”, Proceeding of the
Fourth Annual IEEE International Confidences on Natural
Computation, 2008, pp. 3304–3309.
[10] Gua Zichao and Uhrig E. Robert., (1992) “Using Genetic Algorithms
to select input for Neural Network”. Proceeding of the IEEE
International conference, May, 1992, pp. 87–95.
[11] Hintz K. J. and Spofford J. J., (1990) “Evoling Neural Network”
Proceedings of the IEEE Transactions on Communication and
Intelligence, May 1990, pp. 333–338.
[12] Haykins, S. (2003), “Neural Networks – A comprehensive
foundation”, Prentice – Hall of India Private Limited, New Delhi 2003.
[13] Ileană Ioan., Rotar Corina and Incze Arpad., (2004) “The optimization
of feedforward”, proceedings of the international and information –
ICTAMI 2004, Thessalo-niki, Greece.

103

Você também pode gostar