Sie sind auf Seite 1von 4

Suruchi Sharma, et al International Journal of Computer and Electronics Research [Volume 4, Issue 2, April 2015]

Design and Analysis of FIR Filter using Artificial


Neural Network
Suruchi Sharma Abhishek Lahariya
Dept. Of ECE, M.P.C.T. Gwalior Dept. of ECE, M.P.C.T, Gwalior
E-mail ID: suruchisharma89@yahoo.in E-mail ID: abhishek.laharia@gmail.com

Abstract−For the designing of low pass FIR filter it II. DESIGNING FIR FILTER USING
require filter specification like sampling frequency(fs), cut off WINDOWMETHOD
frequency(fc),pass band frequency, stop band frequency etc. The windowing method requires minimum amount of
By using these specification we can calculate filter computational effort so window method is simple to implement.
coefficients h(n).These filter coefficients decide the structure For the given window ,the maximum amplitude of ripple in the
of filter. It is easy to get these filter cofficients from filter filter response is fixed. Thus the stop band attenuation is fixed in
specification by using simple calculation, but it is difficult to the given window. The frequency response of fir filter shows the
find filter specification from filter coeffients. In this paper we convolution of spectrum of window function & desired
use Blackman window for FIR filter designing& then using frequency response because of this the pass band & stop band
Neural Network tool to estimate the cut off frequency of edge frequency cannot be preciesly specified. In many
given coefficients of FIR filter. Firstly we have designed the application the expression for the desired filter response will be
10th order digital filter then calculate the coefficient of too complicated. Here we used fdatool for filter designing.
designed filter. We have normalized the frequency and
designed filter at frequency ranges 0 to 1 Hz. Then we have III. ARTIFICIAL NEURAL NETWORK
calculated the coefficients of filter at different frequencies. The development of artificial neural network started 50 years
Some data group of coefficients is used to train the neural ago. ANN is a gross simplification of real networks of neurons.
network designed using generalized regression algorithm In common with biological neural networks, ANN can
and rest are used as test input to neural network. Designing accommodate many inputs in parallel and encode the
and analysis of low pass FIR filter using different artificial information in a distributed fashion. The information that is
neural network algorithm and then comparing their result to stored in a neural network is shared by many of its units. This
obtain most effective among the two and accurate neural type of encoding is in sharp contrast to traditional memory
network design algorithm for FIR filter. schemes, where information is stored in only one memory
Keywords: Specification, Coefficients, Neural Network location .The recall process is time taking process and
generalization is usually absent. The distributed storage scheme
I. INTRODUCTION provides many advantages, most important of them being the
Finite impulse response filter are the primary types of filter redundancy in information representation. Thus, an ANN can
used in DSP. For the designing of low pass FIR filter complex undergo partial destruction of its structure and still to be able to
calculation are required. Mathematically, by substitution the function well. Although redundancy can also be built into other
value of passband, trasition width, passband ripple, stop band type of systems, ANN has a natural way of implementing this.
attenuation, sampling frequency, cut-off frequency in any of the The result is a natural fault-tolerant system which is very similar
methods from window method, frequency sampling method, we to biological system.ANN is an information processing system.
can get the value of filter coefficient h(n).In this paper Blackman In this system, the element neurons, process the information. The
window method is used for filter designing and feed forward signals are transmitted by mean of connection links. These links
distributed time delay and radial basis function neural network possess an associated weight, which is multiplied along with the
algorithms are used to train and test the network. incoming signal for any neural network. The output signal is
obtained by applying activation function to the net input. The
artificial neuron is characterized by architecture, training and
©http://ijcer.org e- ISSN: 2278-5795 p- ISSN: 2320-9348 Page 53
Suruchi Sharma, et al International Journal of Computer and Electronics Research [Volume 4, Issue 2, April 2015]

activation function.

(1)  

Figure 2: Architecture of RBF network


Figure 1: ANN Architecture
2. Feed forward distributed time delay
Here we describe two important types of training algorithm. Feed Forward indicates that network has links that extend in
only one direction i.e. neurons are only connected in forward.
1. Radial Basis Network Time delay neural network (TDNN) is an artificial neural
Radial basis networks may require more numbers of neurons network architecture whose primary purpose is to work on
than the feed-forward back propagation network. They can be sequential data. The TDNN units recognize features independent
designed in a fraction of time. They work best when there are of time-shift (i.e. sequence position) and usually form part of a
many training dataset are available. RBFN is less time larger pattern recognition system. Converting continuous audio
consuming for network training. RBFN consists of three layers: into a stream of classified labels for speech recognition. An input
an input layer, a hidden (kernel) layer, and an output layer. Each signal is augmented with delayed copies as other inputs, the
layer is fully connected to the previous layer. The input variables neural network is time-shift invariant since it has no internal
are assigned to the nodes in the input layer and they pass directly state.
to the hidden layer without weights. The transfer function of IV. FORMULATION OF PROBLEM
hidden nodes is RBF. An RBF is symmetrical about a given The aim of this paper is to estimate the cut off frequency of any
mean or a centre point in a multidimensional space. In the RBFN, given coefficients of fir filter and increase the accuracy. Here, two
a number of hidden nodes with RBF activation functions are algorithms are used to train the network and after comparing the
designed a feed forward parallel architecture. The parameters results of these two, we will get whose Accuracy is more from
associated with the RBF network are optimized during the fdatool and nntool. Here we used Radial basis network and Feed
network training. These parameter values are neither necessarily forward distributed time delay algorithms. For training purpose
the same throughout the network nor they directly related to we use input output dataset derived from fdatool, and for testing
constrained of the actual training vectors. When the training purpose we use a test input and simulate it using nntool.
vectors are presumed to be accurate and it is desirable to perform V. EXPERIMENT
a smooth interpolation between them, then linear combinations In this experiment we use Blackman window technique for
of RBFs can be found by which give no error at the training filter designing and feed forward distributed time delay and
vectors. The method of fitting RBF network to data, for function radial basis function neural network algorithms are used to train
approximation, it is closely related to distance weighted and test the network.
regression. The RBF expansion for one hidden layer and a a) Step 1:
Gaussian RBF is represented by radial basis network. Low pass fir filter designed by fdatool. The order of filter is
10.We use cut of frequency(fc) ranges from 0 to 1.Then set the
value of fc = 0.05 and design the filter. Export the value of filter
coefficient h(n) on workspace. Repeat the same process for
 (1)
value of fc from 0.1 to 1.So this gives the 19 dataset of input and
output. Out of these 19 dataset we use 17 for training and 2 for
testing. Here input is h(n) and output is fc.

©http://ijcer.org e- ISSN: 2278-5795 p- ISSN: 2320-9348 Page 54


Suruchi Sharma, et al International Journal of Computer and Electronics Research [Volume 4, Issue 2, April 2015]

Figure 5(b): Network creation by nntool


Figure 3: Filter designing by fdatool
b) Step 2:
Create the MS excel file for training input, targets and testing
input. Then load these files to MATLAB workspace.

Figure 4: MATLAB Workspace Figure 6: Training and simulation with test input

c) Step 3:
In this step we design the neural network model by nntool.
Training is done by Radial basis network and feed forward
distributed time delay algorithms. After training, simulate the
network by testing input. Then compare the result by data from
fdatool.

Figure 7(a): Result of simulation process by Feed forward


distributed time delay network

Figure 5(a): Network creation by nntool


©http://ijcer.org e- ISSN: 2278-5795 p- ISSN: 2320-9348 Page 55
Suruchi Sharma, et al International Journal of Computer and Electronics Research [Volume 4, Issue 2, April 2015]

V. RESULTS AND DISCUSSION


In the above experiment, according to the Table 1: by this
process we can easily estimate the cut off frequency of any
firfilter. Comparison of mean square error using feed forward
distributed time delay network (FFDTDN), radial basis function
(RBF-ANN) algorithm given for analysis of low pass FIR filter.

Figure 7(b): Result of simulation process by RBF network

Table 1: Comparison of mean square error using FFDTDN and RBF-ANN algorithms used for design of low pass FIR

h(n)Filter Blackmann FFDTDN RBF Mean Square Mean Square


coefficient Window error FFTDN error RBF

h(0) 0.35 0.39677 0.35284 0.00218743 0.00000807

h(1) 0.70 0.69912 0.69789 0.00000077 0.00000445

[4] Seema Rani, Amarpreet kaur, J S Ubhi P, “Comparative study of


VI. CONCLUSION FIR and IIR filters for the removal of Baseline noises from ECG
In the presented work different network structures i.e. signal” International Journal of Computer Science and Information
FFDTDN-ANN, and RBF-ANN have been used to achieve the Technologies, Vol. 2, No.3, 2011
optimum model for the defined problem. After comparision [5] Rama Murthy Garimella, “Finite Impulse Response (FIR) Filter
among these two algorithms, it is concluded that both gives Model of Synapses: Associated Neural Networks” Communications
better approximations but overall RBF network is found more and Networking Research Centre. Fourth International Conference
efficient, fast and showing accuracy of 99.99% and above in on Natural Computation, 978-0-7695-3304-9/08 $25.00 © 2008
most of the cases. IEEE.
REFERENCES [6] Khushboo Pachori, Dr. Amit Mishra, “Design of FIR Digital Filters
[1] Awadhesh Gupta, Girish Chanadra Tripathi, “Design and Analysis using ADALINE Neural Network”, 2012 Fourth International
of Low Pass FIR & IIR Filter and Find Optimum Result Using Conference on Computational Intelligence and Communication
Neural Network”, Global journal of Researches in Engineering, Vol. Networks, 978-0-7695-4850-0/12 $26.00 © 2012 IEEE
10 Issue 7 (Verl.0), pp 61-64, December 2010. [7] Harpreet Kaur, Balwinder Dhaliwal, “Design of Low Pass FIR Filter
[2] Seema Verma, P.K. Singhal &Ravindra Pratap Narwaria, Using Artificial Neural Network”, International Journal of
“Reduction of Artifact by FIR low pass filter using Triangular, Information and Electronics Engineering, Vol.3, No.2,pp 204-207
Rectangular and Parzen windows” International Journal of March 2013.
Computer Information systems, Vol. 5, No.5, pp 75-85, 2012
[3] Sheenu Thapar, Parminde Kaur, Neha Aggarwal, “A Low Pass FIR
Filter Design Using Genetic Algorithm Based Artificial Neural
Network” International Journal of Computer Technology and
Electronics Engineering, Vol. 2, Issue 4, August 2012, ISSN
2249-6343.
©http://ijcer.org e- ISSN: 2278-5795 p- ISSN: 2320-9348 Page 56

Das könnte Ihnen auch gefallen