Beruflich Dokumente
Kultur Dokumente
UNDER GUIDANCE OF
SUBMITTED BY
Supreet Kaur
Assistant Professor
2014ECB1422
(June, 2016)
Signature of Supervisor
The M.Tech Viva Voce Examination of Supreet Kaur has been held on
..................................and accepted.
Signature of Supervisor
Signature of H.O.D
Acknowledgement
The following is an attempt to offer my heartfelt gratitude to the various persons
who have been great help in one way or the other for the successful completion of
this dissertation.
I am indebted to my thesis supervisor Dr. Butta Singh, Department of
Electronics and Communication Engineering for her gracious encouragement and
very valued constructive criticism that has driven me to carry out the project
successfully.
I am deeply grateful to Dr. Jyoteesh Malhotra, Head of Department Electronics
Technology, Guru Nanak Dev University, RC Jalandhar for his support and
encouragement. I thank several anonymous publishers of various journals and
research papers for their help.
I wish to express my heart full thanks to the Faculty Members of Department of
Electronics and Communication Engineering Dr. Deepkamal kaur Randhawa,
Er. Manjeet Singh, Er. Nitika Soni, Er Himali Sarangal, Er. Harmander
kaur, Dr. Vinit Grewal for their goodwill and support that helped me a lot in
successful completion of this project and Guru Nanak Dev University, RC
Jalandhar for availing MATLAB.
A Big thank you goes to my Parents for always being there when I needed
them the most. Finally I express my deep sense of gratitude to the almighty GOD
for their blessings for the completion of this dissertation.
Supreet Kaur
2014ECB1422
ii
Abstract
iii
single point so for this reason it is mostly preferred. Genetic algorithms (GAs)
particularly have emerged as a powerful technique for searching in high
dimensional spaces, capable of solving problems despite their lack of knowledge
of the problem being solved.
In this thesis, the complexity of the system is minimized by reducing the number
of 1s between the actual filter and the desired filter using genetic algorithm
optimization technique. Moreover it provides the detail description of power
dissipated in the CMOS circuits and the technologies which are used to
overcome this problem.
The error reduction table for three Low pass filter shows that hamming window
reduces 25% of error in the filter. The toggling of signal after GA id reduced to
50% in case of Hamming window which further reduces the system complexity
and power dissipated. So at the end it concluded that hamming window gives
better results than other windows
iv
Table of Contents
Candidate's declaration form
Acknowledgement
ii
Abstract
iii
Table of Contents
List of Figures
viii
List of Tables
xi
List of Abbreviations
xii
Chapter 1 Introduction
1-17
1.1 Background
1.5
10
11
11
11
12
12
13
14
15
16
1.8 Motivation
16
17
17
18-25
18
26-45
26
3.1.1 Background
26
26
27
27
28
28
3.6.1 Encoding
29
30
30
30
3.6.2 Population
31
31
31
3.7.1 Selection
32
34
32
Mutation
34
35
vi
3.7.3.1
Mutation Rate
37
3.7.4 Termination of GA
37
37
3.8.1
39
3.8.2
41
3.8.2.1
41
Window Method
44
3.9.1
Problem Formulation
44
3.9.2
44
3.9.3
Solution Methodology
45
45
46-66
48
48
54
60
67
5.1
Conclusion
67
5.2
Future Scope
67
References
68-72
vii
List of Figures
Fig 1.1 Basic Filter
11
Fig 1.7 Leakage Current Types: (a) Reverse Biased Diode Current,
(b) Sub-Threshold Leakage Current
Fig 1.8 Clock Gating
12
13
Fig 1.9 (a) Original Signal Flow Graph (b) Unrolled Signal Flow
Graph
14
14
15
15
16
27
29
29
30
30
30
32
33
35
36
36
38
39
40
viii
41
42
43
43
Fig 4.1 Low Pass Filter Ideal Response with specifications fc=350,
fs 1500,N=16 and Rectangular Window
48
49
Fig 4.3 Low Pass Filter Ideal Response with specifications fc=350,
fs= 2000, N=16 and Rectangular Window
50
50
Fig 4.5 Low Pass Filter Ideal Response with specifications fc=500,
fs=2000, N=16 and Rectangular Window
51
52
Fig 4.7 Low Pass Filter Ideal Response with specifications fc=500,
fs= 3000, N=16 and Rectangular Window
53
53
Fig 4.9 Low Pass Filter Ideal Response with specifications fc=350,
fs= 1500, N=16 and Hamming Window
54
55
Fig 4.11 Low Pass Filter Ideal Response with specifications fc=350,
fs= 2000, N=16 and Hamming Window
56
56
Fig 4.13 Low Pass Filter Ideal Response with specifications fc=500,
fs= 2000, N=16 and Hamming Window
57
ix
58
Fig 4.15 Low Pass Filter Ideal Response with specifications fc=500,
fs= 3000, N=16 and Hamming Window
59
59
Fig 4.17 Low Pass Filter Ideal Response with specifications fc=350,
fs= 1500, N=16 and Hanning Window
60
61
Fig 4.19 Low Pass Filter Ideal Response with specifications fc=350,
fs= 2000, N=16 and Hanning Window
62
62
Fig 4.21 Low Pass Filter Ideal Response with specifications fc=500,
fs= 2000,N=16 and Hanning Window
63
64
Fig 4.23 Low Pass Filter Ideal Response with specifications fc=500,
fs= 3000, N=16 and Hanning Window
65
65
List of Tables
Table 1.1: Comparison of Digital and Analog signal processing
42
66
xi
List of Abbreviations
DSP
MAC
ADC
ASIC
FPGA
DAC
FIR
IIR
CMOS
PMOS
fc
fs
Sampling Frequency
SUMBE
CSD
VLSI
DLMS
NMOS
SM
Signed Magnitude
ARM
TSP
HD
Hamming Distance
GA
Genetic Algorithm
DNA
Deoxyiribonucleic Acid
Pc
Crossover Probability
Pm
Mutation Probability
xii
xiv
CHAPTER 1
INTRODUCTION
1.1 BACKGROUND
The evolution of human society has been always closely tied with
the effective communication and exchange of information so that it can enable to
pass human knowledge and skills from one generation to another generation. In the
twentieth century the last three decades, in particular, are often termed as
"information age". The way in which the information is being transmitted, stored and
processed has entirely changed the availability of powerful and fast computers and
the rapid technology advancement in telecommunications fuelled by the growth of
Internet and multimedia. One of the most important key enabling technologies in the
development and enhancement of communication infrastructure was Signal
Processing. The field of signal processing contains the algorithms and hardware that
allow processing of signals which are produced by natural or artificial means. These
signals include speech, audio, video, images, satellite and weather data, etc.
Processing of these signals can involve data acquisition, data conversion, data
coding, data compression, transmission, display, etc. When these signals are
represented in the discrete form and are being processed by computers or special
purpose digital hardware, they can be identified as an exciting and rapidly expanding
field of Digital Signal Processing (DSP) [Prokias et al (2004)].
In DSP, A filter is a device or process that filters out or removes some unwanted
component or feature from a signal as shown in Fig 1.1. Filtering is a class of signal
processing, which defines the feature of filter being the complete or partial
suppression of some aspect of the signal which are not required in designing. The
filter basically functioned as to remove unwanted parts of the signal, such as random
noise, or to extract the important information from the noisy signal, such as the
frequency components lying within a certain frequency range [Mitra (2005)]. An
electrical network that alters the amplitude and/or phase characteristics of a signal
with respect to frequency is also done by filters. Ideally, a filter will not add any new
frequencies to the input signal, nor it will make any change to the component
frequencies of that signal, but it will change the relative amplitudes and/or their
1
Introduction
phase relationships of various frequency components. Filters are often used in
electronic systems to emphasize signals in required frequency ranges and reject
signals that are in other frequency ranges.
FILTER
Raw signal (unfiltered)
filtered signal
Fig 1.1 Basic Filter
The primary functions of filters are one of the followings [Smith (1997)]:
To limit a signal into a fixed frequency band as in low-pass, high-pass, and bandpass filters.
To divide a signal into two or more sub-bands as in filter-banks, sub-band coders,
frequency multiplexers, and graphic equalizers.
To change some parts of the frequency spectrum of a signal as in telephone
channel equalization and audio or video graphic equalizers.
To graphics the input-output relationship of a system such as telecommunication
channels, human vocal tract, and music synthesizers.
Depending on the basis of the filter equation and structural implementation, filters
may be broadly classified into the following categories:
Linear filters versus nonlinear filters.
Time-invariant filters versus time-varying filters.
Adaptive filters versus non-adaptive filters.
Recursive versus non-recursive filters.
Direct form, cascade form, parallel form and lattice structures.
Introduction
analog circuits made up from electronic components such as resistors, capacitors and
op amps to obtain the required filtering effect. Such analog filter circuits are widely
used in the applications such as noise reduction, video signal enhancement, graphic
equalizers in hi-fi systems, and many other areas [Mitra (2006)]. At all stages of
designing, the signal which is being filtered is an electrical voltage or current which
is the direct analogue of the physical quantity (e.g. a sound or video signal or
transducer output) involved.
Advantages:
Simple and unite methodologies of plan
Realization is fast and simple
Disadvantages:
Little stable and responsive to temperature variations
Expensive to realize designing in large amount
1.2.2 Digital Signal Processing
In digital signal processing, Digital filters consist of digital processor that performs
numerical calculations on sampled values of the signal. The processor may be a
general purpose computer such as a PC, or a specialised digital signal processor chip.
Digital filters are used widely in signal processing applications, such as digital image
processing, spectrum analysis, and pattern recognition. Digital filters eliminate
number of problems that are associated with their classical analog counterparts and
thus are preferably replace with the digital filters. The analog input signal is sampled
firstly and then digitized using an ADC (analog to digital converter). The resulting
values in the form of binary numbers, representing successive sampled values of the
3
Introduction
input signal, are transferred to the processor, where the numerical calculations are
carried out on. Fast DSP processors can handle complex consolidation of filters in
parallel or cascade (series), which makes the hardware requirements relatively simple
and compact in comparison to that of equivalent analog circuitry [Tan et al (2007)].
signal may be processed by digital filter by first being digitized and represented as a
sequence of numbers, then manipulated mathematically, and then reconstructed and
it is manipulated by the circuit [Mitra (2006)]. Digital filters are very important part
of DSP. In fact, their extraordinary performance is one of the key reasons that DSP
has become so popular. The filters have two uses: signal separation and signal
restoration. Signal separation is needed when a signal has been contaminated with
interference, noise, or other signals. For example, imagine a device for measuring the
electrical activity of a baby's heart while still in the womb. The raw signal will likely
be corrupted by the breathing and heartbeat of the mother. A filter might be used to
separate these signals so that they can be individually analyzed. Signal restoration is
used when a signal has been distorted in some way. For example, an audio recording
made with poor equipment may be filtered to better represent the sound as it actually
occurred. Another example is the deblurring of an image acquired with an
improperly focused lens, or a shaky camera [Smith (1997)]. A digital system usually
consists of an analog to digital converter to sample the input signal followed by a
microprocessor and some peripheral components such as memory to store data and
filter coefficients etc as shown in Fig 1.3. Finally a digital to analog converter to
complete the output stage. Program instructions (software) running on the
microprocessors implement the digital filter by performing the necessary
mathematical operations on the numbers receives from the ADC. In some high
performance application, an ASIC or FPGA is used instead of a general purpose
microprocessor, or a specialized DSP with specific parallel architecture for
expediting operations such as filtering [Mitra (2006)].
Introduction
Digital filters may be more expensive than an equivalent analog filter due to their
increased complexity, but they make practical many designs that are impractical or
impossible as analog filters. Since digital filters use a sampling process and discretetime processing, they experience latency (the difference in time between the input
and the response), which is almost irrelevant in analog filters. The cut-off frequency
of the pass-band is a frequency at which the transition of the pass-band to the
transition region occurs. The cut-off frequency of the stop-band is a frequency at
which the transition of the transition region to the stop-band occurs [Brapate (2007)].
Advantages
The following list gives some of the main advantages of digital over analog filters
[Tan et al (2007)].
Digital filters can be designed with an exactly linear phase
They do not suffer from the degradation mechanisms of passive and active
components of analogue filters
Digital filters have better stability, reproducibility and higher orders of
precision
It is possible to realise filters with very low cut-off frequencies
They can be realised as integrated circuits.
Fast DSP processors can handle complex combinations of filters in parallel or
cascade (series), making the hardware requirements relatively simple and
compact in comparison with the equivalent analog circuitry.
Digital filters are easily designed, tested and implemented on a general
purpose computer or workstation.
Introduction
DIGTAL
ANALOG
y ( n)
b x( n k )
k
k 0
Introduction
The output y(n) of the FIR filter is the function of the input signal x(n) as given in Fig
1.4 . The response of this filter consists of the finite samples of M so it is said to be as
Finite Duration Impulse Response filter.
Introduction
1.5.2 IIR Filter
IIR Filter is known as a recursive filter [Smith (1997)] shown in Fig 1.5 has feedback
from output to input, and in general its output sample is a function of the previous
output samples and the present and past input samples as described by the following
equation,
y ( n)
y y ( n k ) p x( n k )
k
k 1
k 0
IIR filters have one or more than one non-zero feedback coefficients. That is, as a
result of the feedback term, if the filter has one or more than one poles, once the filter
has been excited with an impulse there is always an output.
Introduction
more versatile in designing complex structures as well as they give out the best
results.
FIR
IIR
Introduction
data bus power however can very much be minimized by optimizing filter coefficients
so to reduce the distance between successive coefficient values and also by reducing
the total numbar of signal toggling in the opposite direction between the successive
coefficients. The coefficients and the input data samples from the input to the
multiplier during FIR filtering. The multiplier power dissipation thus depends on the
number of toggles and also on the number of 1s in these inputs. The coefficient
optimization for reducing the coefficient memory, data bus power thus also reduces
the multiplier power. Higher power reduction can be achieved by focussing on lower
significant bits of the coefficients during minimization
1.6.1 Power Dissipation Sources
In CMOS, the total power dissipated out from the circuit is obtained by adding the
three components as; [Yeap (1998)], [Sakurai (2002)].
changed through PMOS transistor as shown in Fig 1.6 in order to make voltage
transition from 0 to 1 high the supply voltage Vdd
Due to this voltage transistor the power is dissipated as it is determined from product
is Vdd I c transient current [Chadrakasan et al (1995)] [Veendrick (1984)].
I c CL
dVout
dt
transient turns on the both CMOS and PMOS transistor which results in the short
circuit between Vdd and ground.
10
Introduction
Static dissipation occurs when the leakage current is drawn out from the voltage supplied
to the circuit.
power.
Leakage Power
Two sources which results in the leakage of power are; firstly when the current flows
through the reversed biased transistors, secondly when the currents flows through the
non-conducting transistors. The leakage current is proportional to the exponential of
11
Introduction
the voltage which is at threshold and the leakage area as shown in Fig 1.7. The power
consume by the leakage current is as high as the power consumption of the switching
activity i.e 0.6m .
Fig 1.7 Leakage Current Types: (a) Reverse Biased Diode Current,
(b) Sub-thresshold Leakage Current
In order to increase the energy efficiency of signal processing circuits and required
computing power. Designer has developed various design methodologies that can be
applied in custom, application-specific integrated circuits. These defined
technologies have been successful in increasing energy efficiency and reducing
computational complexity in designing a FIR filter. These technologies also suffer
from some drawback like- they provide limited function that means they are less
flexible.
12
Introduction
hardware platform selection and partition of hardware and software. The designing
of system has a great impact on the power consumption and hence low power
techniques are applied to reduce the power consumption at this level.
However if, for example, the instruction level power model is available for a
processor, software power optimization can be performed [Tiwari et al (1994)]. It is
observe that the frequent use of the cache and faster code are likely to reduce the
consumption of power in a system.
Two techniques used for low power consumption at the system level are: clock
gating as shown in Fig 1.8 and power down. The hardware units that are not working
are shut down that means they are at the sleep mode so that the power can be saved.
The clock drivers which consumes30-40% of the overall power are gated to reduce
switching activity.
13
Introduction
Fig 1.9 (a) Original Signal Flow Graph. (b) Unrolled Signal Flow Graph. .
Introduction
15
Introduction
automatically limits the use of pin ordering [Yeap (1998)]. Hence by this way, the
consumption of power is reduced without any cost complexity
1
(Vdd Vt ) 2
The main objective is to reduce the power while making the overall throughput of the
system fixed.
1.8 MOTIVATION
The impmentation of Finite impulse response (FIR) is done as a series of multiply
and accumulate operations on a programmable Digital Signal Processor (DSP). The
multiply and accumulator (MAC) unit of a DSP experiences large amount of
switching activity due to signal transition which results in the higher power
consumption and also increase the computational cost of the filter. In order to reduce
the computational cost, lesser number of adder and multipliers should be used and
this can only be achieved by the optimization techniques. So the Genetic Algorithm
16
Introduction
technique is used to overcome this computational problem in designing FIR filter as
this algorithm is quite simple.
17
CHAPTER 2
LITERATURE SURVEY
A general desire in any digital signal processing system design is
that the number of operations (additions and multiplications) needed to compute the
filter response is as low as possible. In certain applications, this desire is a strict
requirement, for example due to limited computational resources, limited power
resources, or limited time. The last limitation is typical in real-time applications.
There are several ways in which a filter can have different computational complexity.
For example, the order of a filter is more or less proportional to the number of
operations. This means that by choosing a low order filter, the computation time can
be reduced.
For discrete filters the computational complexity is more or less proportional to the
number of filter coefficients. If the filter has many coefficients, for example in the
case of multidimensional signals such as tomography data, it may be relevant to
reduce the number of coefficients by removing those which are sufficiently close to
zero. In multirate filters, the number of coefficients by taking advantage of its
bandwidth limits, where the input signal is down sampled (e.g. to its critical
frequency), and up sampled after filtering.
18
Literature Survey
Lin et al [1998] limits the search of the prototype class of the filter which is obtained
by using Kaiser Window. The single parameter that is the cut off frequency (fc) is
optimized in Kaiser Window design. This further reduces the complexity of the
designed filter by obtaining a new parameter new . It is concluded that the design
complexity of other non- linear optimization is much more complex than the Kaiser
Window approach. On the other hand, it is well known that stop-band attenuation is
determined by the Kaiser window so direct control can be given over the stop-band
attenuation.
Rasidi et al [2011] proposed a method which determines the use of low power serial
multiplier and adder combining both the multipliers and shift adder and folding
transformation in linear phase architecture so that they can reduce the overall
dynamic power consumption of the FIR filter. The implementation of the filter was
done using Xilinx ISE. Virtex IV FPGA and Xinlin XPower analyzer was used to
analyze the power consumption, the minimum power achieved was 110mw in 100
MHz to 8 taps and 8 bits coefficients.
Joshi et al [2010] had proposed a brief description of the structure and the
characteristics of
Literature Survey
booth encoder circuit which generates the product in the parallel combination. Since
multiplication of both is done by same multiplier in the modified system hence it
reduces the cost and the power dissipation which results in the low computational
complexity.
Anantha et al [1992] discussed the certain considerations which are taken into
account in low power design filter. In this paper, it includes the technology and the
style of logic and the logic implemented. Power dissipation contributes the factor
like, transition which occur due to critical raise condition and due to leakage current
in the circuit. So in order to get rid from this problem a pass-gate logic family with
modified threshold is used. It is found as the best performer for the designing of low
power FIR digital filter.
Sameuli [1989] had presented an optimization algorithm for improving FIR filter
coefficients in which the additional non-zero elements are added in the CSD set in
order to compensate the non uniform distribution in the Canonic Signed Digit (CSD)
set. This algorithm consist of two stages; stage one defines the search methodology
of the optimum scale factor and the second stage is the bivariate local search
methodology in the neighbourhood set. This result in the increase in computational
complexity from additional CSD digits. However it improves the frequency
response. These techniques can be used with other filters to reduce the complexity.
Dusan et al [1981] discussed the comparison between an optimal and sub-optimal
algorithm which is used for designing of finite word length and shows the simulation
results for FIR filter coefficients whose length varies from 15 to 35. The conclusion
was made that when the computer resources are not available for optimal method it is
worth to apply the local search method to the filter which have rounded coefficients.
Darren et al [1995] presented a low power implementation approaches with less
number of hardware overhead than in traditional FIR filter implementations. Parallel
of block processing having duplication of hardware result in the reduction of power
consumption. Parallel processing by block size L requires the critical path to the
charged L times longer as compared to sequential implementation which leads to low
power consumption.
20
Literature Survey
Chetana et al [1995] illustrates that the gate pipelined MAC3 circuits, can be clocked
at very high speed because of the simplicity of pipelined stage. However this results
in the increasing of high power dissipation in the circuit. In comparison to this half
bit pipelined MAC2 was designed which results with the same performance
characteristics as MAC3, but have lesser area and power dissipation in the circuit.
Sankarayya et al [1996] had discussed a new set of algorithm for the realization of
low power FIR filter that uses various orders of differences between the coefficients
for the computation of the convolution with the input data. The results of this
computation can be stored and reused. The technique used in the computational is
necessary as per convolution coefficients are compared to the coefficients which
were used directly. It was shown that the reduction in the results also reduces the net
energy dissipation.
Erdogan et al [1996] presented a new multiplication technique for an application to
a single multiplier CMOS which is based on DSP Processors. Reduction of switching
activity with the multiplier section results in low power consumption. This scheme
uses the transpose direct form FIR filter structure. This reduction has been
demonstrated with two examples as they were discussed in the paper having different
word length and filter order achieving upto 63% of power reduction in the FIR filter.
Hezar et al [1996] presented an efficient design procedure for the filter whose
coefficients are restricted to the ternary set [-1, 0, +1] which is cascaded by a
multiplication free architecture. This programming algorithm had minimized the
errors which occur during the performance of ternary filter coefficients set. Power
reduction in VLSI was very much better than compared to the other current efforts
which seek to implement these design using VLSI implementation.
Horrock et al [1996] had discussed an overview of the methodologies that had been
used in the past few years for DSP. This includes minimization of power at
architecture level and algorithm level. This paper also lists the power consumption
techniques.
An ultra low power delayed least mean square (DLMS) adaptive filter which
operates in the threshold region for hearing aid application is discussed. The
threshold operation which was discussed in this threshold operation is done by the
21
Literature Survey
parallel architecture with pseudo NMOS style. This architecture operates and the low
clock rate and reduced power while remaining the throughput [IEEE transactions on
VLSI (2003)]. DLMS operates at the frequency of 22 KHz using 400mV supply to
achieve 91% improvement in power as compared to CMOS style.
Low power block based filtering cores had been studied and they are specially
designed for low power consumption filter implementation [Erdogan et al (2003)].
This system uses algorithm flow for designing low power consumption filters and
also uses 2 compliments and SM number (sign magnitude) presentation. Its been
concluded that the 49% of power consumption was reduced as compared to
conventional filtering cores and areas overhead of 5% was increased.
Another scheme was described for the implementation of low power cores for the
hearing and applications [Zwyssig et al (2001)]. In this method, there are two power
saving approaches that has been investigated. First method uses the macrocomponent framework in which the assembly of cores were very easy to identify
hierarchical plug in- basis and second
strategy.
These techniques operate at less frequency and also results in less computational cost
in the FIR filter in DSP processor. The ARM based system on chip platform was
used for designing and testing in which the core was embedded. Using this method,
Power handling/manipulating was possible by power management.
A coefficient algorithm was used for the designing of low power FIR filter. In this
the algorithm decomposes the individuals coefficients into two sub-components
[Erdogen et al (1998)]. The heuristic approach was used for the decomposition which
divides the given coefficients. The part which was produced after division is than
implemented using single shift operation and leaves the other parts and reduces the
word length. This results in the reduction of the switching capacitance and results in
the 63% of the power saving.
Mehendale et al [1996] presented an algorithmic and architectural transform for the
designing of low power filter. In this the implementation was done on both hardware
macros and software on programmable DSP. For the implementation these transform
addresses the reduction in the power in the program memory and data buses and
multiplier. Also it discusses the architectural extension to support some of these
22
Literature Survey
transformations. This transforms reduces the computational cost and achieved the
power reduction.
Young et al [2004] discussed low power canonic signed digit (CSD) filter response
and high speed structure using vertical common sub expression. The horizontal
common sub-expression method was used in conventional linear phase CSD filter
because of its inherent symmetrical filter coefficients. In this the similar values linear
filter having significant bits of adjacent filter coefficients were equal. The method is
more efficient where the bit precision of implementation was lower. The
computational cost also decreases.
Suckely [1991] discussed that for FIR filter designing; a genetic algorithm was used
that produces the filter realization very close to the minimum computational
complexity. This method was automatic and efficient. Also it was said earlier that
computational complexity of the filter had been guaranteed to be produce
automatically but still the optimization method was needed to produce the acceptable
run times. The efficient design procedure was needed for the work on the genetic
algorithm application. The genetic algorithm was commonly used because of its
simplicity.
Merakos et al [1997] presented a noval high level transformation for the
implimentation of low power FIR filter in the paper. The new idea was the recording
of the coefficients in the filter which aims at the minimization of the switching
activity. As a measure of the switching activity the Hamming Distance (HD) of the
filter that is between successive coefficients, stored in a memory, was used. The
transformation that can be incorporated both in application specific architectures as
well as in general purpose programmable architecture. The recording of the N
number of coefficients, for the HD optimization of this sequence can be discussed by
a Travelling Salesman Problem (TSP), which is a well known NP complete problem.
A novel heuristic algorithm for a fast and accurate solution to these problem discuses
above was proposed. The experimental results shows that the proposed technique
leads to significant power saving in terms of switching activity reduction as well as
computation cost.
Soni et al [2011] proposed that the Exponential window provides better side-lobe
roll-off ratio as compared to Kaiser Window which is efficiently used for some
23
Literature Survey
applications such as filter design, beam forming, and speech processing. Besides this
proposed a design of digital non-recursive Finite Impulse Response (FIR) filter by
using Exponential window. The far-end stop-band attenuation was most significant
parameter when the signal which was going to be filtered has great concentration of
spectral energy. In a sub-band coding, the filter is processed to separate out various
frequency bands for independent processing. In case of speech, e.g. the far-end
rejection of the energy in the stop-band should be more in oreder to minimize the
energy leakage from one band to another. Therefore, the desinging of filter should be
done in such a way so that it can provide better far-end stop-band attenuation
(amplitude of last ripple in stop-band). The designing of digital filter by Kaiser
window has a better far-end stop-band attenuation than the designing of filter by the
other previously well known adjustable windows such as Dolph- Chebyshev and
Saramaki, which are special cases of Ultraspherical windows, but obtaining a digital
filter which performs higher far-end stop-band attenuation than Kaiser window will
be useful. In this paper, the designing of non-recursive digital FIR filter had been
proposed by using Exponential window. It provides better far-end stop-band
attenuation as compared to filter designed by well known Kaiser window, which is
the advantage of designing a filter by Exponential window over the designing of
filter by Kaiser window. Hence the computational comlexity becomes lesser.
Multipliers play an important role design of digital FIR filters. A novel design
technique for deriving highly efficient multipliers which operates on a limited range
of multiplier [Turner et al [2004] values was discussed in the literature. Using the
technique,
Xilinx
Virtex
field
programmable
gate
array
(FPGA)
.The
implementations for a poly-phase filter and discrete cosine transform were derived
with area reductions of 31%70% and speed increases of 5%35% when they were
compared to designs using general-purpose multipliers. This design gives better
result as compared to the other fixed coefficient methods.
Algorithmic methods focus on design of filter coefficients with lower complexity
rather than on the optimisation of hardware structures. This in turn leads to lower
implementation complexity of filtering structures. There are two categories of
algorithms to solve an approximation problem for FIR filters with powers-of-two
coefficients (PWR2): exact and approximate. Exact algorithms guarantee the optimal
filter design, i. e. a minimum order of the filter for a given specification. An example
24
Literature Survey
of exact algorithms is an exhaustive search (examines all possibilities) and a branchand-bound algorithm [Lim et al (1983)]. Approximate algorithms do not guarantee
the optimality of the design, although they can deliver near-optimal designs in less
time than exact algorithms. The majority of algorithms for the multiplier-less FIR
filter design belong to the category of approximate algorithms.
Yunlong et al. [2011] they proposed a simple method for design an FIR filter as
compare to other existed methods. This method is the simplest except rectangular
window method. Filter transition bandwidth is smaller than 4.65/N for filter order N.
For the same filter specifications filter order obtained by using the new method is
much smaller than by using Kaiser Window if minimum stop-band attenuation is in
the range of 39.5 dB to 48.5dB and corresponding maximum pass-band ripple is
from 0.35 dB to 0.18 dB.
25
CHAPTER 3
PROPOSED METHOD
3.1 GENETIC ALGORITHM
GA is a search technique used in computing to find true or
approximate solutions to optimization and search problems. GA are a particular class
of evolutionary algorithms that use techniques inspired by evolutionary biology such
as inheritance, mutation, selection, and crossover (also called recombination).GAs
dier from classical optimization and search methods in several respects. Rather than
focusing on a single solution
3.1.1 Background
GA is part of evolutionary computing, which is a fast developing area of artificial
intelligence. It was inspired by Darwins theory of evolution.
GA is a stochastic search method which is used for searching an optimal solution to
the evolution function of an optimization problem [Pittman et al (2000)].GA was
proposed by Holland in early seventies [Holland et al (1975)] as a computer program
which mimic the natural evolutionary process. Then GA was extended to a functional
optimization by De Jong [Jong (1980)] and then Goldbug presented a detailed
mathematical model of a GA [Goldbug (1989)].
26
Proposed Method
27
Proposed Method
For i = 1, 2......N
...3.1
In order to minimize f(x), for f(x)>0 then the objective function is written as
maximize
28
Proposed Method
1
1 f ( x)
...3.2
operators
Evolution environment
Fig 3.2 Genetic Evolution Flow
3.6.1 Encoding
As we know that GA searches directly in the solution space. So they are encoded in
the way which can be manipulated by GA. This representation is known as Genetic
or Chromosomes representation. The genetic representation is known as genotype
and the physical appearance of that which is caused by genes is known as phenotype
as shown in Fig 3.3
29
Proposed Method
3.6.1.1 Binary encoding
This encoding is commonly used because of its simplicity. In binary encoding,
each chromosome is a string of bits 0 and 1as given in Fig 3.4. Binary encoding
provides many possible chromosomes even with the small number of alleles
Coefficients 0.421875
2-1
0
-2-4
-2-6
1 0 0 -1 0
30
Proposed Method
3.6.2 Population
Population is set by randomizing the initial sets of solutions. The size of the
population may vary but usually it is fixed to a certain value. Still it is common that
the population is purely generational. This means that the next generation is
constructed by the offsprings, except individual which are preserved only if elitism
operators are used. The population is purely randomized.
The population p t at the generation t can be denoted as a set of chromosomes as
pt xt (1),.xt (2).......xt ( N )
...3.3
F t f t (1), f t (2)...... f t ( N )
...3.4
31
Proposed Method
Crossover and mutation are probabilistic operations and their occurrence
frequencies are controlled by predefined probabilities Pc and Pm
respectively. High frequency of occurrence is assigned to crossover as they
play the major role in GA typically 80-90% and low frequency of occurrence
is assigned to mutation typically 5-10%.
3.7.1 Selection
Selection is defined as the process of selecting which individual should be allowed to
contribute to the next generation in genetic process. The individuals who are selected
survive for a long to reproduce that means they are equivalent of survival of the
fitness. The selection method used in GA is The Roulette Wheel selection method.
3.7.1.1 Roulette wheel method
Parents are selected according to their fitness value. The better the chromosomes are,
the more the chances to be selected the parents have. The circumference of Roulette
wheel is divided into segments and marked for each string proportionate to the
fitness value as clearly seen in Fig 3.8
The wheel is spun n times; each time a segment is selected by the wheel pointer. The
segment with the highest fitness value will have more probability for selection.
Firstly the ranks are provided to the population and then according to that rank the
fitness value are assigned to them.
32
Proposed Method
For example, population having low fitness value is ranked as number 1 and the
population with the highest fitness value is ranked as number N.
The following algorithm can describe this process;
th
The i string in the population is selected with the probability to f i , where f i is the
fitness value. The population size is usually kept fixed, the sum of probabilities of
th
each string being selected for the tool must be one. The probability of i selected
string is
pi
Fi
n
F
j 1
C k ,1 i k .
th
Step 2: compute the probability of i selected string.
Pi
fi
k
f
i 1
Pcj p i ,
i 1
33
Proposed Method
Step 4: generate a random number r, 0 r 1 . Set i =1, Pc 0 0
Step 5: if Pci r and Pc (i 1) r
Proposed Method
3.7.3 Mutation
After crossover process, the strings are subjected to mutation. Mutation of a bit
involves flipping/altering it, changing 0 to 1 and vice versa with a small amount of
mutation probability Pm as shown in Fig 3.11. The bit-wise mutation is processed
bit-by- bit by flipping a coin with a probability of Pm . Flipping a coin with the
probability of Pm is simulated as follows. A number between 0 and 1 is randomly
chosen. If the random number is smaller than Pm then the outcome of the coin
flipping is true otherwise it is false. If at any bit, the outcome is true then bit is
changed, otherwise the bit is not altered. The bits of the strings are independently
muted, that is, the probability of mutation of other bits is not affected by the mutation
of the bits. The mutation operator introduces new genetic structures in the population
by randomly modifying some of its building blocks. It helps the search algorithm to
escape from local minimas traps since the modification is not related to any previous
genetic structure of the population. It creates different structure representing other
sections of the search space. A simple genetic algorithm treats the mutation only as a
secondary operator with the role of restoring lost genetic materials.
35
Proposed Method
Suppose, for an example, all the strings in a population have conveyed to a zero at a
given position, and then cross over cannot regenerate a one at that position while a
mutation could. The mutation is simply an insurance policy against the irreversible
loss of genetic material. The mutation is also used to maintain diversity in the
population.
36
Proposed Method
3.7.3.1 Mutation rate
Mutation rate is the probability of mutation, which is used in calculating the number
of bits to be muted. The diversity is preserved by the mutation operator among the
population, which is also very important for the search. Mutation probabilities are
smaller in natural populations which lead us to conclude that mutation is
appropriately considered a secondary mechanism of GA adoption. Typically, the
population size of 30 to 200 with the mutation rate varying from 0.001 to 0.5 is used
by the simple GA. The flowchart shown in Fig 3.12 describes the working of
mutation.
3.7.4 Termination of GA
Because the GA is a stochastic search method, it is very difficult to formally specify
convergence criteria. As the fitness value of a population may remain constant for a
number of generations before a superior individual is found, the application of
conventional termination criteria becomes problematic. A common practice is to
terminate the GA after a pre-specified number of generations and then test the quality
of the best members of the population against the problem definition. The GA may
be restarted or a fresh search initiated, if no acceptable solutions are found.
Proposed Method
carrying out the filtering operation using fixed word lengths on the filter
performance is carried out in this step.
38
Proposed Method
39
Proposed Method
In the pass-band, the magnitude response has a peak deviation of p and in the stopband, it was a maximum deviation of s .The difference between p and s gives the
transition width of the filter and transition band determines how sharp the filter is.
The magnitude response decreases monotonically from the pass-band to stop-band in
this region. The following are the key parameters of interest:
(i) p Peak pass-band deviation (or ripples).
(ii) s Stop-band deviation.
(iii) s Stop-band edge frequency.
(iv) p Pass-band edge frequency
(v) Fs Sampling frequency.
Thus the minimum stop-band attenuation, AS And the peak pass-band ripple, AP , in
decibels are given as,
AS (stopband attenuation) = 20 log 10 s
AP (passband ripple) =
20 log 10 (1 P )
Another important parameter is the filter length, n, which defines the number of
filter.
40
Proposed Method
( )e jn d
...3.5
1
2
1 e
jn
...3.6
jn
2 f c Sin (n c )
,
n c
41
n 0,
Proposed Method
n0
2 fc
The ideal infinite impulse response is truncated by using various windows. When this
window is multiplied by the ideal transfer function then all the coefficients within the
window are retained and all that are outside the window are discarded.
Table 3.1:Summary of Important Features of Common Window Function .
Window
Transition Width
Rectangular
4 / n
-21 dB
Hamming
8 / n
-53 dB
Hanning
8 / n
-44 dB
1, 0 n M
wn
0, otherwise
Rectangular window:
42
Proposed Method
Hamming Window:
Hanning Window:
43
Proposed Method
H ( Z . ) a n Z n
n 0
H ( , ) a n e jwT
n 0
Where and T are the angular frequency and sampling period respectively.
Error
H () H (, )
2 1/ 2
44
Proposed Method
3.9.3 Solution Methodology
The intent works is to optimize the coefficients of FIR filter, to minimize the number
of 1s and satisfies the desired characteristics in terms of stop band attenuation and
pass band ripples. This section illustrates the steps of algorithm that shows how the
objective is achieved.
3.9.3.1 Steps of algorithm for the minimization of number of 1s using GA.
Step1: Compute filter coefficients h1 n and freq. response H1 of ideal
FIR filter for 0 n N 1.
Step 2: Calculate the number of 1s between the FIR filter coefficients h1 n with
ternary encoding.
Step 3: Set the Number of chromosomes (k), mutation rate Pm ,Crossover rate
Cm Stopping criteria.
Step 4: Populate k sets of possible designed solutions, to produce symmet
-ric coefficients H D n0 i k 1 and 0 n N 1
Step 5: Compute the frequency response of the coefficients chromosomes
for H D in population.
Step 6: Calculate the number of 1s in each of the coefficient chromosome.
Step 7:Evaluate the fitness of the chromosomes
Step 8: Apply Roulette wheel selection.
Step 9: Apply crossover operators at a desired rate.
Step 10: Mutate at a desired rate
Step 11: Evaluate again the fitness of the chromosomes.
Step12: If the Stopping criterion is met store the chromosomes according to the
fitness, else goto stop 8.for minimization of 1s, using G.A.
45
Proposed Method
46
CHAPTER 4
RESULTS AND DISCUSSIONS
This chapter presents the results of low power filter design and low
computational complexity using GA Algorithm. The GA algorithm is implemented
under the MATLAB environment. These filters vary in terms of cut off frequency
(fc), sampling frequency (fs) and number of coefficients (N). The filters are designed
using three windows;
Rectangular Window
Hamming Window
Hanning Window
Minimization of number of 1s using GA is formulated as multi-objective algorithm
which reduces the mean square error as well as the distance between the ideal filter
and the desired FIR filter. The coefficients value quantized to 32 bit coefficients
represents from the initial sets of coefficients for optimization the results are
presented using GA.
The minimization of computational cost complexity using optimization genetic
technique is formulated as a multi-objective minimization problem which minimizes
the number of 1s and mean square error between the ideal filter and the desired FIR
filter. In order to generate the solution, weighted method is used which convert the
multi-objective problem into a single objective problem by assigning the appropriate
weights to the objectives. The solutions are generated by making the trade off
between the objectives. The procedure continues until the best combinations of
solutions have been achieved. The best solution is then selected on the basis of
fitness, reduction in computational complexity corresponds to minimum square error.
This algorithm is applied on three windows based designed FIR filter.
Below are the response graph of the ideal and the desired filters which shows the
reduction in computational cost by decreasing the number o 1s using GA. The graph
shows the linear response of FIR filter. These filters are designed by evaluating the
window techniques of FIR filters. Among three windows best results are obtained by
46
47
Rectangular
1500
350
16
Original coefficients
Actual number of 1s
38
Optimized coefficients
Reduced number of 1s
30
Magnitude (dB)
50
-50
-100
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
0
-200
-400
-600
-800
48
Magnitude (dB)
50
-50
-100
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-500
-1000
-1500
Case 2: Specifications
Window Type
Rectangular
2000
350
16
Original coefficients
Actual number of 1s
38
Optimized coefficients
Reduced number of 1s
34
49
Magnitude (dB)
50
-50
-100
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-200
-400
-600
Magnitude (dB)
0
-20
-40
-60
-80
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-500
-1000
50
Rectangular
2000
500
16
Original coefficients
Actual number of 1s
40
Optimized coefficients
Reduced number of 1s
28
Magnitude (dB)
50
0
-50
-100
-150
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-500
-1000
51
Magnitude (dB)
20
0
-20
-40
-60
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-500
-1000
-1500
Case 4: Specifications
Window Type
Rectangular
3000
500
16
Original coefficients
Actual number of 1s
40
Optimized coefficients
Reduced number of 1s
36
52
Magnitude (dB)
50
-50
-100
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-200
-400
-600
Magnitude (dB)
20
0
-20
-40
-60
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-500
-1000
-1500
53
Hamming
1500
350
16
Original coefficients
Actual number of 1s
26
Optimized coefficients
Reduced number of 1s
24
54
Case 2: Specifications
Window Type
Hamming
2000
350
16
Original coefficients
Actual number of 1s
32
Optimized coefficients
Reduced number of 1s
26
55
Magnitude (dB)
50
-50
-100
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-500
-1000
Magnitude (dB)
0
-20
-40
-60
-80
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-500
-1000
-1500
56
.5
Hamming
2000
500
16
Original coefficients
Actual number of 1s
28
Optimized coefficients
Reduced number of 1s
16
Magnitude (dB)
50
-50
-100
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-500
-1000
57
Magnitude (dB)
50
-50
-100
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-500
-1000
-1500
Case 4: Specifications
Window Type
Hamming
3000
500
16
Original coefficients
Actual number of 1s
32
Optimized coefficients
Reduced number of 1s
22
58
Magnitude (dB)
50
-50
-100
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-500
-1000
Magnitude (dB)
50
-50
-100
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-500
-1000
-1500
59
Hanning
1500
350
16
Original coefficients
Actual number of 1s
28
Optimized Coefficients
Reduced number of 1s
22
Magnitude (dB)
100
-100
-200
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-500
-1000
60
Magnitude (dB)
50
-50
-100
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-500
-1000
Case 2: Specifications
Window Type
Hanning
2000
350
16
Original coefficients
Actual number of 1s
36
Optimized coefficients
Reduced number of 1s
30
61
Magnitude (dB)
50
0
-50
-100
-150
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-500
-1000
Magnitude (dB)
50
-50
-100
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-500
-1000
62
Hanning
2000
500
16
Original coefficients
Actual number of 1s
40
Optimized coefficients
Reduced number of 1s
30
Magnitude (dB)
50
-50
-100
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-500
-1000
63
Magnitude (dB)
20
0
-20
-40
-60
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-500
-1000
-1500
Case 4: Specifications
Window Type
Hanning
3000
500
16
Original coefficients
Actual number of 1s
32
Optimized coefficients
Reduced number of 1s
30
64
Magnitude (dB)
100
-100
-200
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-500
-1000
Magnitude (dB)
20
0
-20
-40
-60
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Normalized Frequency ( rad/sample)
0.9
Phase (degrees)
-500
-1000
65
% reduction
Response in No. of 1s
fc
Rectangular
Hamming
Hanning
fs
350 1500 16
38
30
0.6518
21.05%
350 2000 16
38
34
0.8396
10.52%
500 2000 16
40
28
0.832937
30%
500 3000 16
40
36
0.6617
10%
350 1500 16
26
24
0.629
7.69%
350 2000 16
32
24
0.5718
25%
500 2000 16
28
16
0.673
42.85%
500 3000 16
32
22
0.5367
31.25%
350 1500 16
28
22
0.5096
21.42%
350 2000 16
36
30
0.3898
16.66%
500 2000 16
40
24
0.3146
40%
500 3000 16
32
30
0.2496
6.25%
66
CHAPTER 5
CONCLUSION AND FUTURE WORK
5.1 CONCLUSION
In this present work error minimization algorithms are presented for the
low power and low complexity realization of FIR filters on the Programmable DSP
analysis is presented to show that the number of multipliers and adders in filters
increases the computational cost and the number of signal toggling between the
values forms the main cause of power dissipation. To improve this problem the
optimization technique is used.
Optimization algorithm is presented in detail so as to minimize the number of signal
toggling or to minimize the number of 1s. One such technique presented is the
genetic algorithm for the minimization of these problems discussed above. GA can
be implemented as a computer simulation in which a population of abstract
representation (the chromosomes, or the genotype, or the phenotype) of the candidate
solutions to be optimized problem evolves the better solution. GAs particularly has
emerged as a powerful technique for searching in high dimensional spaces, capable
of solving problems despite their lack of knowledge of the problem being solved.
The error reduction for three Low pass filter shows that hamming window reduces
25% of error in the filter. The toggling of signal after GA is reduced to 50% in case
of Hamming window which further reduces the system complexity and power
dissipated. So at the end it concluded that hamming window gives better results than
other windows.
67
68
REFERENCES
Anantha, P. et al. (1992). Low power CMOS digital. IEEE Journal of Solid
Stare Circuits, 27(4): 473-484.
Barapate, R. A. (2007). Digital Signal Processing, Pune: Tech-Max Publication.
Chadrakasan, A. and Brodersen, R. W. (1995). Low Power Digital CMOS Design.
Kluwer Publisher, 2-5.
Chadrakasan, A. et al.(1995). Optimizing power using transformations. IEEE
Transaction on Computer Aided Design,. 14(1): 12-31.
Darren, N. et al. (1995). Low power digital filter architectures. IEEE Symposium on
Circuit and Systems , 1: 231-234.
Erdogan, A. T. and Arslan, T. ( 2003). Low power block based FIR filtering cores.
Proceedings of the International Symposium on Circuits and Systems (ISCAS '03), 5:
341-344.
Erdogan, A.T. and Arslan, T. (1998). A coefficient segmentation algorithm for low
power implementation of Fir filters. IET Electronics Letter, 34(19): 1817-1819.
Erdogen, A. T. and Arslan, T. (1996). Low power multiplication scheme for FIR filter
implimentation on single multiplies CMOS DSP processors. IEEE Electronics
Letters, 32(21): 123-125.
Goldberg, D. E. (1989). Genetic algorithms
in search
optimization
and
References
Ifeachor, E. C. and Jervis, B. W. (2002). Finite impulse response (FIR) filter design
in Digital Signal Processing. A Practical Approach, South Asia: Prentice hall, 342440.
Jang, Y. and Yang, S. (2004). Low power CSD linear phase FIR filter structure
using vertical common sub-expression. IEEE Electronics Letters Online, 38(15):
777-779.
Jong, K. D. (1980). Hybrid
methods
using
genetic
References
Merakos, P. K. et al. (1997). A novel transformation for reduction of switching
activity in FIR filters implimentation. IEEE 13th Conference on Digital Signal
Processing, Proceedings: 653-656.
Mitra,S .K . (2005). Digital Signal Processing, New York: Tata McGraw Hill.
Nagendra, C. et al. (1995). Low power consideration in the design of pipelined FIR
filters. IEEE Symposium on Low Power Electronics, 32-33.
Phuong , N.H. (2009). The FIR filter Design . The Window Design Method.
Pittman, J. and Murthy, C.A. (2000). Fitting optimal picewise linear functions
using genetic algorithms. IEEE Tarns. on Pattern Analysis Machine Intelligence,
22: 701718.
Proakis, J. G., and Manolakis, D. G.(2004). Digital Signal Processing. Prentice Hall
India publication.
Rabaey, J. M. and Pedram, M. (1996). Low power design methodologies. Kluwer
Publishers, 2-11.
Rajput, R. P. and Swamy, M. N. S. (2012). High speed modified booth encoder
multiplier for signed and unsigned numbers. International Conference on Modelling
and Simulation, 649-654.
Rashidi, B. and Pourormazd, M. (2011). Design and implementation of low power
digital FIR filter based on low power multipliers and adders on Xilinx FPGA.
International Conference on Electronics Computer Technology, 18-22.
Sakurai, T. (2002). Low power and high speed VLSI design with low supply voltage
through cooperation between levels. International Symposium on Quality Electronic
Design. Proceedings, 18(21): 445-450.
Samueli, H. (1989). An improved search for the design of multiplierless FIR filters
with power-of two coefficients. IEEE Transaction circuits and system, 36(7):10441047.
70
References
Sankaraya, N. et al. (1996). Algorithm for low power realization of FIR filters using
differential coefficients. IEEE 10th international conference on VLSI design
proceedings, 370-375.
Smith, S.W. (1997). The scientist and engineer's guide to Digital Signal Processing.
San Diego: California Technical Publications.
Soni, V. et al. (2011). Application of exponential window to design a digital non
recursive FIR filter. International Conference on Communications and Signal
Processing, 1015-1019.
Suckley, D. (1991). Genetic algorithm in the design of FIR filter. IEEE Proceedings
on Genetic Algorithms, 138: 234-238.
Tan, L. and Jiang, J. (2007). Digital Signal Processing: Fundamentals and
Applications, Amsterdam: Academic Press.
Tiwari, V. et al. (1994). Compilation techniques for low energy an overview. In
IEEE Symposium On Low Power Electronics, San Diego, California, USA.
Proceedings, 3839.
Turner, R. H. and Woods, R .F.(2004). Highly efficient limited range multipliers for
LUT-based FPGA architectures. IEEE Transactions on Very Large Scale Integration
(VLSI) Systems, 12 (10):1113-1118.
Ultra-Low-Power DLMS Adaptive Filter for Hearing Aid Applications.(2003). IEEE
Transactions on Very Large Scale Integration (VLSI) Systems, 11(6): 1058-1067.
Veendrick, H. J. M. (1984). Short-circuit dissipation of static CMOS circuitry and
its impact on the design of buffer circuits. IEEE Journal of Solid-State Circuit, 19:
468473.
Yeap, G K. (1998). Practical low power digital VLSI design. Kluwer Academic
Publisher, Norwell, Mass, 2-8.
Yunlong, W. et al. (2011). An extreme simple method for digital FIR filter design in
proceedings. Third International Conference on Measuring Technology and
Mechatronics Automation (ICMTMA), IEEE, 410-413.
71
References
Zwyssig, E. P. et al. (2001). Low power system on chip implementation scheme of
digital filtering cores. IEEE Seminar on Low Power IC Design,proceedings, 5:1-9.
72