Development of Radar Pulse Compression Techniques Using Computational Intelligence Tools Ajit Kumar Sahoo

Development of Radar Pulse Compression Techniques Using Computational Intelligence Tools Ajit Kumar Sahoo
Development of Radar Pulse
Compression Techniques Using
Computational Intelligence Tools
Ajit Kumar Sahoo
Roll - 507EC005
Department of Electronics and Communication Engineering
National Institute of Technology Rourkela
Rourkela – 769 008, India
Development of Radar Pulse
Compression Techniques Using
Computational Intelligence Tools
Thesis submitted in partial fulfillment of the requirements for the degree of
Doctor of Philosophy
in
Electronics and Communication Engineering
by
Ajit Kumar Sahoo
(Roll - 507EC005)
under the guidance of
Prof. Ganapati Panda
Department of Electronics and Communication Engineering
National Institute of Technology Rourkela
Rourkela, Orissa, 769008, India
Electronics and Communication Engineering
National Institute of Technology, Rourkela
Rourkela-769 008, Orissa, India.
Dr. Ganapati Panda FNAE, FNASc.
Professor
March 17, 2012
Certificate
This is to certify that the thesis entitled “Development of Radar Pulse
Compression Techniques Using Computational Intelligence Tools”
by Ajit Kumar Sahoo, submitted to the National Institute of Technology,
Rourkela for the degree of Doctor of Philosophy, is a record of an original research
work carried out by him in the department of Electronics and Communication
Engineering under my supervision. I believe that the thesis fulfills part of the
requirements for the award of degree of Doctor of Philosophy. Neither this thesis
nor any part of it has been submitted for any degree or academic award elsewhere.
Ganapati Panda
Acknowledgement
I take the opportunity to express my reverence to my supervisor Prof. G. Panda
for his guidance, inspiration and innovative technical discussions during the course
of this work. He encouraged, supported and motivated me throughout the work. I
always had the freedom to follow my own ideas for which I am very grateful.
I am thankful to Prof. S. Meher, Prof. K. K. Mahapatra, Prof. S. K. Patra
of Electronics and Communication Engg. department and Prof. K. B. Mohanty
of Electrical Engg. department for extending their valuable suggestions and help
whenever I approached.
My special thanks to Dr. D. P. Acharya and Dr. Sitanshu Sahu for their constant
inspiration and encouragement during my research.
My hearty thanks to Jagganath, Trilochan, Upendra, Sudhansu, Prakash, Pyari,
Nithin, Vikas and Yogesh for their help, cooperation and encouragement.
I acknowledge all staff, research scholars and juniors of ECE department, NIT
Rourkela for helping me.
I am also grateful to Prof. S. K. Sarangi, Director NIT Rourkela for providing
me adequate infrastructure and other facilities to carry out the investigations for my
research work.
I take this opportunity to express my regards and obligation to my family
members whose support and encouragement I can never forget in my life.
I am indebted to many people who contributed through their support, knowledge
and friendship to this work and made my stay in Rourkela an unforgettable and
rewarding experience.
Ajit Kumar Sahoo
Abstract
Pulse compression techniques are used in radar systems to avail the benefits of large
range detection capability of long duration pulse and high range resolution capability
of short duration pulse. In these techniques a long duration pulse is used which is
either phase or frequency modulated before transmission and the received signal
is passed through a filter to accumulate the energy into a short pulse. Usually,
a matched filter is used for pulse compression to achieve high signal-to-noise ratio
(SNR). However, the matched filter output i.e. autocorrelation function (ACF)
of a modulated signal is associated with range sidelobes along with the mainlobe.
These sidelobes are unwanted outputs from the pulse compression filter and may
mask a weaker target which is present nearer to a stronger target. Hence, these
sidelobes affect the performance of the radar detection system. In this thesis, few
investigations have been made to reduce the range sidelobes using computational
intelligence techniques so as to improve the performance of radar detection system.
In phase coded signals a long pulse is divided into a number of sub pulses each of
which is assigned with a phase value. The phase assignment should be such that the
ACF of the phase coded signal attain lower sidelobes. A multiobjective evolutionary
approach is proposed to assign the phase values in the biphase code so as to achieve
low sidelobes. Basically, for a particular length of code mismatch filter is preferred
over matched filter to get better peak to sidelobe ratio (PSR). Recurrent neural
network (RNN) and recurrent radial basis function (RRBF) structures are proposed
as mismatch filters to achieve better PSR values under various noise conditions,
Doppler shift and multiple target environment.
Polyphase and linear frequency modulated (LFM) codes yield lower sidelobes
compared to biphase codes. Various weighing functions are used to further suppress
the sidelobes of polyphase and LFM codes. In this thesis, convolutional windows
are used as weighing function to achieve high PSR magnitude at different Doppler
shift conditions.
In high range resolution radar wide bandwidth signals are used for transmission.
The conventional narrowband hardware may not support the instantaneous wide
bandwidth. Therefore, the wide bandwidth signal is split into several narrowband
signals which are transmitted and recombined coherently at the receiver to get the
effect of the wideband signal. However, the ACF of such narrow band pulse train
suffers from grating lobes and hence reduce the range resolution capability of the
pulse train. In this work, evolutionary computation algorithms are proposed to
optimally choose the parameters of stepped frequency LFM pulse train to achieve
reduced grating lobes, low peak sidelobe and narrow mainlobe width.
Keywords:
Pulse Compression,
Matched filter,
Sidelobes,
ACF,
Multiobjective, RNN, RRBF, LFM, Polyphase Codes, Convolutional
Windows, Grating Lobes.
vi
Contents
Certificate
iii
Acknowledgement
iv
Abstract
v
List of Figures
x
List of Tables
xiv
List of Acronyms
xv
1 Introduction
1
1.1
Pulse compression . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
1.2
Matched filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4
1.2.1
Matched filter for a narrow bandpass signal . . . . . . . . . .
7
1.2.2
Matched filter response to Doppler shifted signal . . . . . . . .
8
1.2.3
Properties of ambiguity function . . . . . . . . . . . . . . . . .
9
1.2.4
Cuts through ambiguity function . . . . . . . . . . . . . . . . 10
1.3
Radar signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.1
Frequency modulated signal . . . . . . . . . . . . . . . . . . . 11
1.3.2
Phase coded signal . . . . . . . . . . . . . . . . . . . . . . . . 12
1.4
Background and scope of the thesis . . . . . . . . . . . . . . . . . . . 15
1.5
Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.6
Objective of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.6.1
1.7
Structure and chapter wise contribution of the thesis . . . . . 18
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
vii
2 Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
22
2.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2
Merit measures and problem formulation . . . . . . . . . . . . . . . . 24
2.3
Techniques used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4
2.3.1
Genetic algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3.2
Multi objective GA . . . . . . . . . . . . . . . . . . . . . . . . 29
Generation of pulse compression codes . . . . . . . . . . . . . . . . . 35
2.4.1
Using genetic algorithm . . . . . . . . . . . . . . . . . . . . . 35
2.4.2
Using NSGA-II . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.5
Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.6
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3 Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
43
3.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.2
Problem formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.3
Techniques used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.4
3.5
3.3.1
Adaptive linear combiner . . . . . . . . . . . . . . . . . . . . . 47
3.3.2
Artificial neural network . . . . . . . . . . . . . . . . . . . . . 51
Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.4.1
Sidelobe suppression using adaptive linear combiner . . . . . . 64
3.4.2
Sidelobe suppression using MLP, RNN, RBF, RRBF . . . . . 64
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4 Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
76
4.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.2
LFM and polyphase codes . . . . . . . . . . . . . . . . . . . . . . . . 79
4.3
4.2.1
LFM signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.2.2
Polyphase codes . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Problem formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
viii
4.3.1
For LFM signal . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.3.2
For polyphase codes . . . . . . . . . . . . . . . . . . . . . . . 89
4.4
Windows used for sidelobe suppression . . . . . . . . . . . . . . . . . 90
4.5
Simulation results
4.6
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.5.1
Analysis for LFM signals . . . . . . . . . . . . . . . . . . . . . 93
4.5.2
Analysis for polyphase codes . . . . . . . . . . . . . . . . . . . 98
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5 Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
104
5.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.2
LFM pulse train . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.3
Problem formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.4
5.5
5.3.1
Problem formulation -1 . . . . . . . . . . . . . . . . . . . . . . 110
5.3.2
Problem formulation -2 . . . . . . . . . . . . . . . . . . . . . . 110
5.3.3
Problem formulation -3 . . . . . . . . . . . . . . . . . . . . . . 111
Techniques used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.4.1
Particle swarm optimization . . . . . . . . . . . . . . . . . . . 112
5.4.2
NSGA-II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
Determination of parameters of LFM pulse train . . . . . . . . . . . . 115
5.5.1
Using PSO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.5.2
Using NSGA-II . . . . . . . . . . . . . . . . . . . . . . . . . . 116
5.6
Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.7
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
6 Conclusion and Future Work
127
6.1
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.2
Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Bibliography
130
Dissemination of Work
141
ix
List of Figures
1.1
Pulsed radar waveform . . . . . . . . . . . . . . . . . . . . . . . . . .
2
1.2
Transmitter and receiver ultimate signals . . . . . . . . . . . . . . . .
3
1.3
Block diagram of a pulse compression radar system . . . . . . . . . .
4
1.4
Block diagram of matched filter . . . . . . . . . . . . . . . . . . . . .
5
1.5
The instantaneous frequency of the LFM waveform over time . . . . . 11
1.6
Phase modulated waveform . . . . . . . . . . . . . . . . . . . . . . . 12
1.7
Matched filter output of different signals . . . . . . . . . . . . . . . . 13
2.1
Crossover . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.2
Mutation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.3
Flow chart for GA . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.4
NSGA-II procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.5
Flow chart for NSGA-II . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.6
Crossover using binary bits 1 and -1 . . . . . . . . . . . . . . . . . . . 38
2.7
Mutation using binary bits 1 and -1 . . . . . . . . . . . . . . . . . . . 38
3.1
Adaptive linear combiner . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2
Single neuron structure . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.3
Mlutilayer perceptron network . . . . . . . . . . . . . . . . . . . . . . 55
3.4
Block diagram of recurrent neural nerwork . . . . . . . . . . . . . . . 58
3.5
Architecture of radial basis function network . . . . . . . . . . . . . . 59
3.6
Architecture of recurrent radial basis function network . . . . . . . . 61
3.7
26 different possible input sequences for 13-bit Barker codes . . . . . 63
x
3.8
Filter response in dB for 13-bit Barker code obtained using (a)ACF
(b)LMS (c)RLS (d)Modified RLS algorithms
3.9
. . . . . . . . . . . . . 65
Filter response in dB for 35-bit Barker code obtained using (a)ACF
(b)LMS (c)RLS (d)Modified RLS . . . . . . . . . . . . . . . . . . . . 66
3.10 Convergence graphs of different structures for (a)13-bit (b)35-bits
Barker codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.11 Compressed waveforms for 13 bit Barker code using (a)MLP (b)RNN
(c)RBF (d)RRBF structures . . . . . . . . . . . . . . . . . . . . . . . 69
3.12 Input waveform on addition of two 5-DA 13-bit Barker sequence
having same magnitude (a)Left shift (b)Right shift (c)Added
waveform (d)Waveform after flip about the vertical axis . . . . . . . . 72
3.13 Compressed waveforms for 13-bit Barker code having same IMR and
5 DA for (a)MLP (b)RNN (c) RBF (d)RRBF structures . . . . . . . 73
4.1
Real and imaginary part of the chirp signal for T B = 50 . . . . . . . 80
4.2
Amplitude spectrum of chirp signal for T B = 50
4.3
Compressed envelope . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.4
Matched filter output and phase values of 100 element Frank code . . 83
4.5
Matched filter output and phase values of 100 element P1 code . . . . 84
4.6
Matched filter output and phase values of 100 element P2 code . . . . 85
4.7
Matched filter output and phase values of 100 element P3 code . . . . 87
4.8
Matched filter output and phase values of 100 element P4 code . . . . 87
4.9
Frequency response curve . . . . . . . . . . . . . . . . . . . . . . . . . 92
. . . . . . . . . . . 81
4.10 Matched filter output with Hamming weighing at the receiver . . . . 93
4.11 Effect on sidelobes due to Doppler shift
. . . . . . . . . . . . . . . . 94
4.12 Compressed waveforms for T B = 50 for amplitude tapering (α = 0.1)
97
4.13 Compressed waveforms for T B = 50 for cubic phase distortion (∆B =
0.75B and ∆T =
1
)
B
. . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.14 Matched filter output for 100 element Frank code using Hamming
window
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
xi
4.15 Matched filter output of P3 code under different Doppler shift . . . . 100
4.16 Matched filter output of P4 code under different Doppler shift . . . . 100
4.17 Effect on sidelobes due to Doppler shift
. . . . . . . . . . . . . . . . 101
5.1
Stepped frequency LFM pulse train . . . . . . . . . . . . . . . . . . . 105
5.2
Stepped frequency LFM pulse for Tp ∆f = 3, Tp B = 4.5 and N = 8.
Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom: ACF (in dB) . . . . 109
5.3
Stepped frequency LFM pulse for Tp ∆f = 3, Tp B = 0 and N = 8.
Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom: ACF (in dB) . . . . 109
5.4
Stepped frequency LFM pulse for Tp ∆f = 2.5, Tp B = 12.5 and N = 8.
Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom: ACF (in dB) . . . . 118
5.5
Stepped frequency LFM pulse for Tp ∆f = 4, Tp B = 16 and N = 8.
Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom: ACF (in dB) . . . . 118
5.6
Pareto front obtained using NSGA-II for Tp ∆f ∈ [2, 10], c ∈ [2, 10]
and N = 8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
5.7
Stepped frequency LFM pulse for Tp ∆f = 2, c = 5, Tp B = 12 and
N = 8. Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom: ACF (in dB) 121
5.8
Stepped frequency LFM pulse for Tp ∆f = 2, c = 5.1412, Tp B =
12.2824 and N = 8. Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom:
ACF (in dB) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
5.9
Stepped frequency LFM pulse for Tp ∆f = 2.8721, c = 5.0978, Tp B =
17.5135 and N = 8. Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom:
ACF (in dB) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
5.10 Pareto front obtained using NSGA-II for Tp ∆f ∈ [2, 10] c ∈ [2, 10],
ǫ = 0.01 and N = 8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
5.11 Pareto front obtained using NSGA-II for Tp ∆f ∈ [5, 30], c ∈ [2, 10],
ǫ = 0.01 and N = 8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.12 Pareto front obtained using NSGA-II for Tp ∆f ∈ [2, 10], c ∈ [2, 5],
ǫ = 0.01 and N = 8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
xii
5.13 Stepped frequency LFM pulse for Tp ∆f = 9.0188, c = 3.5502, Tp B =
41.0373 and N = 8. Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom:
ACF (in dB) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
5.14 Stepped frequency LFM pulse for Tp ∆f = 4.9667, c = 4.0720, Tp B =
25.1911 and N = 8. Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom:
ACF (in dB) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
5.15 Stepped frequency LFM pulse for Tp ∆f = 3.6048, c = 4.6129, Tp B =
20.2334 and N = 8. Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom:
ACF (in dB) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.16 Stepped frequency LFM pulse for Tp ∆f = 3, c = 5, Tp B = 18 and
N = 8. Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom: ACF (in dB) 125
xiii
List of Tables
1.1
Barker codes
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1
Sequences obtained using GA . . . . . . . . . . . . . . . . . . . . . . 40
2.2
Sequences obtained using NSGA-II . . . . . . . . . . . . . . . . . . . 41
3.1
PSRs obtained using various learning algorithms. . . . . . . . . . . . 64
3.2
PSRs obtained by various structures . . . . . . . . . . . . . . . . . . 68
3.3
Comparison of PSRs in dB at different SNRs for 13-bit Barker code . 70
3.4
Comparison of PSRs in dB at different SNRs for 35-bit Barker code . 70
3.5
Comparison of range resolution ability for 13-bit Barker code of two
targets having same IMR and DA. . . . . . . . . . . . . . . . . . . . . 71
3.6
Comparison of range resolution ability for 35-bit Barker code of two
targets having same IMR and DA. . . . . . . . . . . . . . . . . . . . . 71
3.7
Comparison of range resolution ability for 13-bit Barker code of two
targets having different IMR and DA. . . . . . . . . . . . . . . . . . . 74
3.8
Comparison of 35-bit Barker code for range resolution ability of two
targets having same IMR and DA . . . . . . . . . . . . . . . . . . . . 74
3.9
Doppler shift performance . . . . . . . . . . . . . . . . . . . . . . . . 74
4.1
Comparison of PSR for different Doppler shift for T B = 50
4.2
PSR using amplitude tapering . . . . . . . . . . . . . . . . . . . . . . 96
4.3
PSR using cubic phase distortion . . . . . . . . . . . . . . . . . . . . 98
4.4
Comparison of PSR for different Doppler shift . . . . . . . . . . . . . 102
5.1
Values of Tp ∆f , Tp B obtained for N = 8 and f1 = 0 . . . . . . . . . . 119
xiv
. . . . . 95
List of Acronyms
Radar RAdio Detection And Ranging
CW Continuous Waveform
TB Time-Bandwidth
PCR Pulse Compression Ratio
TR Transreceiver
AWGN Additive White Gaussian Noise
SNR Signal-to-Noise Ratio
PSD Power Spectral Density
ACF Autocorrelation Function
AF Ambiguity Function
LFM Linear Frequency Modulated
PSL Peak Sidelobe Level
PSR Peak to Sidelobe Ratio
ISR Integrated Sidelobe Ratio
MPS Minimum Peak Sidelobe
CI Computational Intelligence
xv
EA Evolutionary Algorithm
MF Merit Factor
GA Genetic Algorithm
NSGA Nondominated Sorting Genetic Algorithm
NSGA-II Nondominated Sorting Genetic Algorithm -II
VEGA Vector Evaluated Genetic Algorithm
MOGA Multiobjective Genetic Algorithm
VLSI Very Large Scale Integrated
ISL Integrated Sidelobe Level
LS Least Square
LP Linear Programming
ALC Adaptive Linear Combiner
FIR Finite Impulse Response
MSE Mean Square Error
LMS Least Mean Square
RLS Recursive Least Square
MLP Multilayer Perceptron
BP Back Propagation
RNN Recurrent Neural Network
RBF Radial Basis Function
xvi
RRBF Recurrent Radial Basis Function
ANN Artificial Neural Network
NN Neural Network
DA Delay Apart
IMR Input Magnitude Ratio
NLFM Nonlinear Frequency Modulated
DFT Discrete Fourier Transform
PSO Particle Swarm Optimization
pdf probability distribution function
xvii
Chapter 1
Introduction
Radar an acronym for RAdio Detection And Ranging. It is an electromagnetic
system used to detect and locate the object by transmitting the electromagnetic
signals and receiving the echoes from the objects within its coverage [1]. The echoes
are used to extract the information about the target such as range, angular position,
velocity and other identifying characteristics. A continuous waveform (CW) is the
simplest radar waveform which is transmitted continuously while receiving target
echoes on a separate antenna. The advantage of CW is the unambiguous Doppler
measurement. However, due to continuous nature of the waveform the target range
measurement is entirely ambiguous.
Most of the modern radar systems employ a pulsed waveform which provides
range information accurately. The primary advantage of pulsed radar is that the
transmitter and receiver can share the same antenna due to pulsating nature of
the waveform. A pulsed waveform is shown in Figure 1.1, where Tp is the pulse
duration and Tr is the pulse repetition time. The unambiguous range Ru that can
be measured by this waveform as described in [2] is
Ru =
cTr
2
(1.1)
where c is the speed of light.
Two important factors to be considered for radar waveform design are range
resolution and maximum range detection. Range resolution is the ability of the
1
Introduction
Chapter 1
Figure 1.1: Pulsed radar waveform
radar to separate closely spaced targets and it is related to the pulse width of the
waveform. The narrower the pulse width the better is the range resolution. But,
if the pulse width is decreased, the amount of energy in the pulse is decreased and
hence maximum range detection gets reduced. To overcome this problem pulse
compression techniques are used in the radar systems.
1.1
Pulse compression
The maximum detection range depends upon the strength of the received echo. To
get high strength reflected echo the transmitted pulse should have more energy for
long distance transmission since it gets attenuated during the course of transmission.
The energy content in the pulse is proportional to the duration as well as the peak
power of the pulse. The product of peak power and duration of the pulse gives an
estimate of the energy of the signal. A low peak power pulse with long duration
provides the same energy as achieved in case of high peak power and short duration
pulse. Shorter duration pulses achieve better range resolution. The range resolution
rres is expressed [2] as
rres =
c
2B
(1.2)
where B is the bandwidth of the pulse.
For unmodulated pulse the time duration is inversely proportional to the bandwidth.
If the bandwidth is high, then the duration of the pulse is short and hence this
offers a superior range resolution. Practically, the pulse duration cannot be reduced
indefinitely. According to Fourier theory a signal with bandwidth B cannot have
duration shorter than 1/B i.e. its time-bandwidth (T B) product cannot be less than
2
Introduction
Chapter 1
unity. A very short pulse requires high peak power to get adequate energy for large
distance transmission. However, to handle high peak power the radar equipment
become heavier, bigger and hence cost of this system increases. Therefore peak
power of the pulse is always limited by the transmitter. A pulse having low peak
power and longer duration is required at the transmitter for long range detection. At
the output of the receiver, the pulse should have short width and high peak power
to get better range resolution. Figure 1.2 illustrates two pulses having same energy
with different pulse width and peak power. To get the advantages of larger range
detection ability of long pulse and better range resolution ability of short pulse, pulse
compression [3] techniques are used in radar systems.
The range resolution depends on the bandwidth of a pulse but not necessarily on the
Figure 1.2: Transmitter and receiver ultimate signals
duration of the pulse [4]. Some modulation techniques such as frequency and phase
modulation are used to increase the bandwidth of a long duration pulse to get high
range resolution having limited peak power. In pulse compression technique a pulse
having long duration and low peak power is modulated either in frequency or phase
before transmission and the received signal is passed through a filter to accumulate
the energy in a short pulse. The pulse compression ratio (P CR) is defined as
P CR =
width of the pulse before compression
width of the pulse after compression
3
(1.3)
Introduction
Chapter 1
The block diagram of a pulse compression radar system is shown in Figure
1.3. The transmitted pulse is either frequency or phase modulated to increase the
bandwidth. Transreceiver (TR) is a switching unit helps to use the same antenna
as transmitter and receiver. The pulse compression filter is usually a matched filter
whose frequency response matches with the spectrum of the transmitted waveform.
The filter performs a correlation between the transmitted and the received pulses.
The received pulses with similar characteristics to the transmitted pulses are picked
up by the matched filter whereas other received signals are comparatively ignored
by the receiver.
Figure 1.3: Block diagram of a pulse compression radar system
1.2
Matched filter
In radar applications the reflected signal is used to determine the existence of the
target. The reflected signal is corrupted by additive white Gaussian noise (AWGN).
The probability of detection is related to signal-to-noise ratio (SNR) rather than
exact shape of the signal received. Hence it is required to maximize the SNR rather
4
Introduction
Chapter 1
than preserving the shape of the signal. A filter which maximizes the output SNR
is called matched filter [5]. A matched filter is a linear filter whose impulse response
is determined for a signal in such way that the output of the filter yields maximum
SNR when the signal along with AWGN is passed through the filter.
An input signal s(t) along with AWGN is given as input to the matched filter
as shown in Figure 1.4. Let N0 /2 be the two sided power spectral density (PSD) of
AWGN. It is required to find out the impulse response h(t) or the frequency response
H(f ) (Fourier transform of h(t)) that yields maximum SNR at a predetermined delay
t0 . In other words, h(t) or H(f ) is determined to maximize the output SNR which
is given by
Figure 1.4: Block diagram of matched filter
SP
NP
=
out
|s0 (t0 )|2
n20 (t)
(1.4)
where SP is the signal power, N P is the output noise power, s0 (t0 ) is the value of
the output signal s0 (t) at t = t0 and n20 (t) is the mean square value of the noise.
If S(f ) is the Fourier transform of s(t), then s0 (t) is obtained as
Z ∞
s0 (t) =
H(f )S(f )ej2πf t df
(1.5)
−∞
The value of s0 (t) at t = t0 is
s0 (t0 ) =
Z
∞
H(f )S(f )ej2πf t0 df
(1.6)
−∞
The mean square value n20 (t) of the noise is evaluated as
Z
N0 ∞
2
|H(f )|2 df
n0 (t) =
2 −∞
5
(1.7)
Introduction
Chapter 1
Substituting (1.6) and (1.7) in (1.4) yields
2
R
∞
j2πf t0
df −∞ H(f )S(f )e
SP
R
=
N0 ∞
N P out
|H(f )|2 df
2
−∞
Using Schwarz inequality the numerator of (1.8) can be written as
Z ∞
2 Z ∞
Z ∞
2
j2πf
t
0
|H(f )| df
|S(f )ej2πf t0 |2 df
df ≤
H(f )S(f )e
−∞
−∞
(1.8)
(1.9)
−∞
In (1.9) the equality holds good if
H(f ) = K1 [S(f )ej2πf t0 ]∗ = K1 S ∗ (f )e−j2πf t0
(1.10)
where K1 is an arbitrary constant and ∗ stands for complex conjugate. Using the
equality sign of (1.9), which corresponds to maximum output SNR, in (1.8)
R∞
|S(f )|2 df
2E
SP
(1.11)
= −∞ N0
=
N P out
N
0
2
where E is the energy of the finite time signal and defined as
Z ∞
Z ∞
2
E=
|s(t)| dt =
|S(f )|2 df
−∞
(1.12)
−∞
From (1.11) it is obvious that the maximum SNR is a function of the energy of the
signal but not the shape. Taking inverse Fourier transform of (1.10) the impulse
response of matched filter is obtained as
h(t) = K1 s∗ (t0 − t)
(1.13)
From (1.13) it is clear that the impulse response of matched filter is a delayed mirror
image of the conjugate of the input signal. From (1.6) and (1.10) the output at t = t0
is given as
s0 (t0 ) = K1
=
R∞
−∞
R∞
K1 −∞
S(f )S ∗ (f )e−j2πf t0 ej2πf t0 df
|S(f )|2 df
= K1 E
6
(1.14)
Introduction
Chapter 1
Equation (1.14) states that regardless of the type of waveform, at the predefined
delay t = t0 the output is the energy of the waveform for K1 = 1. The output of the
matched filter is evaluated as
s0 (t) = s(t) ⊗ h(t)
R∞
= −∞ s(τ )h(t − τ )dτ
R∞
= −∞ s(τ )K1 s∗ (τ − t + t0 )dτ
R∞
= K1 =1,t0 =0 −∞ s(τ )s∗ (τ − t)dτ
(1.15)
where ⊗ denotes the linear convolution operation. The right hand side of (1.15) is
known as autocorrelation function (ACF) of the input signal s(t).
1.2.1
Matched filter for a narrow bandpass signal
Most of the radar signals are narrow bandpass signals. A narrowband signal s(t) [5]
can be represented as
1
1
s(t) = u(t)ej2πf0 t + u∗ (t)e−j2πf0 t
2
2
(1.16)
where u(t) is the complex envelope of s(t) and f0 is the carrier frequency.
From (1.15) and (1.16)
R∞
s0 (t) = K41 −∞ [u(τ )ej2πf0 τ + u∗ (τ )e−j2πf0 τ ]
∗
u (τ − t + t0 )e−j2πf0 (τ −t+t0 ) + u(τ − t + t0 )ej2πf0 (τ −t+t0 ) dτ
(1.17)
Evaluating the products, (1.17) is represented as
R∞
s0 (t) = K41 ej2πf0 (t−t0 ) −∞ u(τ )u∗ (τ − t + t0 )dτ
R∞
+ K41 e−j2πf0 (t−t0 ) −∞ u∗ (τ )u(τ − t + t0 )dτ
R∞
+ K41 ej2πf0 (t−t0 ) −∞ u∗ (τ )u∗ (τ − t + t0 )e−j4πf0 τ dτ
R∞
+ K41 e−j2πf0 (t−t0 ) −∞ u(τ )u(τ − t + t0 )ej4πf0 τ dτ
(1.18)
of first and third terms respectively. So it can be written as
o
n
R
K1
∗
j2πf0 (t−t0 ) ∞
s0 (t) = 2 Re e
u(τ )u (τ − t + t0 )dτ
−∞
o
n
R
∞
+ K21 Re ej2πf0 (t−t0 ) −∞ u∗ (τ )u∗ (τ − t + t0 )e−j4πf0 τ dτ
(1.19)
In (1.18) the second and fourth terms of right hand side are the complex conjugate
7
Introduction
Chapter 1
The second term on the right hand side of (1.19) is the Fourier transform of
u∗ (τ )u∗ (τ − t + t0 ) evaluated at f = 2f0 , which is at much higher frequency than the
spectrum of the complex envelope u(t). So neglecting the second term the expression
in (1.19) becomes
o
R∞
ej2πf0 (t−t0 ) −∞ u(τ )u∗ (τ − t + t0 )dτ
o
i
nh
R∞
= Re K21 e−j2πf0 t0 −∞ u(τ )u∗ (τ − t + t0 )dτ ej2πf0 t
s0 (t) =
K1
Re
2
n
(1.20)
The expression inside the square bracket of (1.20) is defined as new complex envelope
u0 (t) which is expressed as
u0 (t) = K2
Z
∞
−∞
where K2 =
u(τ )u∗ (τ − t + t0 )dτ
(1.21)
K1 −j2πf0 t0
e
.
2
The output of the matched filter is
s0 (t) = Re u0 (t)ej2πf0 t
(1.22)
From (1.21) and (1.22) it is observed that the matched filter output of narrow
bandpass signal has a complex envelope u0 (t) which is obtained by passing the
complex envelope u(t) through its own matched filter.
1.2.2
Matched filter response to Doppler shifted signal
Most of the targets in the environment are non stationary. So the frequency of
the reflected signal from a target experiences Doppler shift. The Doppler shifted
complex envelope is represented as
uD (t) = u(t)ej2πfd t
(1.23)
where fd is the Doppler shift.
Substituting uD (t) for first u(t) in (1.21) and choosing t0 = 0 and K2 = 1
Z ∞
u(τ )ej2πfd τ u∗ (τ − t)dτ
u0 (t, fd ) =
−∞
8
(1.24)
Introduction
Chapter 1
Reversing the operations of τ and t a modified expression obtained as
Z ∞
u(t)u∗ (t − τ )ej2πfd t dt
χ(τ, fd ) =
(1.25)
−∞
Equation (1.25) is one of the versions of the ambiguity function (AF). The AF
describes the output of the matched filter if the input signal is delayed by τ and
Doppler shifted by fd relative to the values for which the matched filter is designed.
The AF was introduced by Woodward [6] which is an important tool for radar
signal analysis. But the AF expressions given in [2,4–6] differ in the sign of τ and fd .
τ gives the information whether the target is farther from or nearer to the reference
and fd gives the information whether the target is moving towards or moving away
from the radar. A standard form of AF which is used in most of the radar systems
is
Z
|χ(τ, fd )| = ∞
∗
j2πfd t
u(t)u (t + τ )e
−∞
dt
(1.26)
where a positive τ corresponds to the target being farther from the radar and a
positive fd corresponds to the target moving towards the radar.
1.2.3
Properties of ambiguity function
Some of the important properties of AF [5] are explained below where energy of u(t)
normalized to unity .
1. It has maximum value at origin (0,0) i.e.
|χ(τ, fd )| ≤ |χ(0, 0)| = 1
(1.27)
2. The total volume under AF is unity and independent of signal waveform.
Z ∞Z ∞
|χ(τ, fd )|2 dτ dfd = 1
(1.28)
−∞
−∞
3. AF is symmetrical with respect to origin
|χ(τ, fd )| = |χ(−τ, −fd )|
9
(1.29)
Introduction
Chapter 1
4. If a complex envelope u(t) has AF |χ(τ, fd )| then addition of linear frequency
modulation, which is equivalent to a quadratic phase modulation, makes the
AF as
2
u(t)ejπkt ⇔ |χ(τ, fd − kτ )|
1.2.4
(1.30)
Cuts through ambiguity function
1. Cuts along the delay axis
The cut along the delay axis is obtained by setting fd = 0 in (1.26) i.e.
Z ∞
∗
|χ(τ, 0)| = u(t)u (t + τ )dt = |R(τ )|
(1.31)
−∞
where R(τ ) is the autocorrelation function of u(t).
2. Cuts along the Doppler axis
Setting τ = 0 in (1.26) yields
Z
|χ(0, fd )| = ∞
−∞
|u(t)|2 ej2πfd t dt
(1.32)
Equation (1.32) states that the cut along the Doppler axis yields the Fourier
transform of the magnitude of the square of the complex envelope u(t).
1.3
Radar signals
In radar system a particular waveform is first determined for a given application and
it is used to design the optimum detection system. The waveform should provide
least amount of uncertainty or ambiguity when the reflected signal is used to extract
the information about the range, the velocity and the number of true targets present
in the environment. The different types of signals those are mostly used in radar
systems are discussed in sequel.
10
Introduction
Chapter 1
1.3.1
Frequency modulated signal
Linear frequency modulated (LFM) signals are used in most of the radar systems to
achieve wide operating bandwidth. In this case the frequency increases (up chirp)
or decreases (down chirp) linearly across the pulse. The instantaneous phase of the
chirp signal is expressed as
1
φ(t) = 2π(f0 t + kt2 )
2
(1.33)
where f0 is the carrier frequency and k is the frequency sweep rate related to pulse
duration Tp and bandwidth B as
k=
B
Tp
(1.34)
The instantaneous frequency is given by
f (t) =
d
1
(f0 t + kt2 ) = f0 + kt
dt
2
(1.35)
Equation (1.35) states that the instantaneous frequency is a linear function of
Figure 1.5: The instantaneous frequency of the LFM waveform over time
time, and hence is called as linear frequency modulation. Figure 1.5 illustrates
the instantaneous frequency of LFM waveform that sweeps from f0 to f1 . The
matched filter responses of an unmodulated pulse (duration 10µs) and an LFM
pulse (duration 10µs and bandwidth 3MHz) are depicted in Figures 1.7(a) and
1.7(b) respectively. From these figures it is evident that the matched filter output
11
Introduction
Chapter 1
of LFM signal has narrow mainlobe width and hence has better range resolution
capability. However it is associated with sidelobes which are unwanted in output
from the filter. The compressed pulse width of LFM signal is 1/B and the PCR is
obtained as
P CR = BTp
1.3.2
(1.36)
Phase coded signal
The increase in bandwidth can also be achieved by phase modulation. In this case a
long pulse width Tp is divided into a number of sub pulses each of width tb as shown
in Figure 1.6. Each sub pulse is assigned with a phase value φi , where i = 1, 2, ...N .
The received echo is passed through a filter to get a single output peak. The most
popular phase coding is biphase or binary coding. A biphase code consists of a
sequence of +1 and -1. The phases of the transmitted waveform is 00 for +1 and
1800 for -1. The coded signal is discontinuous at the point of phase reversal. The
matched filter response of a randomly assigned 10-bit biphase code ([1 -1 1 -1 1 -1
-1 1 1 -1]) is shown in Figure 1.7(c). It is evident from the figure that phase coded
signals are also associated with the sidelobes. The PCR of phase coded pulse is
obtained as
Figure 1.6: Phase modulated waveform
P CR =
Tp
tb
(1.37)
Figure 1.7 shows that the modulated signals provide better range resolution as
compared to unmodulated signals but the matched filter output of the modulated
signals suffer from the sidelobes. These sidelobes may hide the small targets or may
cause false target detection. The sidelobe having largest amplitude is called peak
12
Introduction
Chapter 1
1
0.8
|ACF|
0.6
0.4
0.2
0
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
Delay time
−5
x 10
(a) Matched filter response for unmodulated pulse
1
Mainlobe
|ACF|
0.8
0.6
Sidelobes
0.4
0.2
0
−1
−0.8
−0.6
−0.4
−0.2
0
Delay time
0.2
0.4
0.6
0.8
1
−5
x 10
(b) Matched filter response for frequency modulated pulse (T B = 30)
1
Mainlobe
0.8
|ACF|
0.6
0.4
Sidelobes
0.2
0
0
2
4
6
8
10
12
14
16
18
Delay time
(c) Matched filter response for phase modulated pulse
Figure 1.7: Matched filter output of different signals
13
20
Introduction
Chapter 1
Table 1.1: Barker codes
Code length
Coded signal
PSR in dB
2
1-1,-11
-6
3
11-1
-9.5
4
11-11,111-1
-12
5
111-11
-14
7
111-1-11-1
-16.9
11
111-1-1-11-1-11-1
-20.8
13
11111-1-111-11-11
-22.3
sidelobe. The lower the peak sidelobe level (PSL) the better is the code. To quantify
the the waveform characteristics peak to sidelobe ratio (PSR) and integrated sidelobe
ratio (ISR) are used as measures of performance in radar systems. These are defined
as
peak sidelobe power
mainlobe power
total power in sidelobes
ISR = 10 log10
mainlobe power
P SR = 10 log10
(1.38)
(1.39)
In biphase codes the selection of random phase 0 or π is a difficult task. The phases
are selected so that the matched filter output of the code has lower sidelobes. Barker
codes are the special type of binary codes having sidelobes of unity magnitude.
Exhaustive computer based search reveals that the Barker codes are available for
the length of 2, 3, 4, 5, 7, 11 and 13 only. The Barker codes along with their PSR
values are listed in Table 1.1. The Barker code have maximum compression ratio is
13 and highest PSR magnitude is 22.3 dB.
A longer code is required for many radar application to achieve high pulse
compression ratio. One way to obtain a longer code having lower sidelobe level
is by nesting two Barker codes using Kronecker product. This type of code is called
compound Barker code. If one Barker code has length l1 and that of other is l2 ,
then the compound Barker code is of length l1 l2 and the compression ratio is l1 l2 .
For example a 35-bit compound Barker code is generated by taking the Kronecker
tensor product of 5-bit and 7-bit Barker codes and the resultant code is [1 1 1 -1 -1
14
Introduction
Chapter 1
1 -1 1 1 1 -1 -1 1 -1 1 1 1 -1 -1 1 -1 -1 -1 -1 1 1 -1 1 1 1 1 -1 -1 1 -1]. Although a larger
compression ratio is achieved by compound Barker code, the peak sidelobes are not
proportionally decreased. The codes those yield minimum peak sidelobe level but
do not meet the Barker condition (i.e. maximum PSL is unity) are called minimum
peak sidelobe (MPS) level codes.
If the pulse is allowed to take more than two values, it is known as a polyphase
code. The phases of the polyphase code are chosen in such way that its ACF should
have lower sidelobes. However the polyphase codes are sensitive to Doppler shift.
To overcome this problem the polyphase codes are derived from the phase history
of the frequency modulated pulses. The details of polyphase codes and their pulse
compression methods are discussed in Chapter 4.
1.4
Background and scope of the thesis
A lot of research work has been carried out over past few decades to achieve low
sidelobes and high range resolution in the radar pulse detection system [5]. Biphase
coding techniques are preferred in pulse compression techniques owing to their easy
implementation. The phases of biphase codes are assigned randomly to different
bits of a certain length of code according to different measure of performance. So
efficient techniques are required to assign the phases of biphase codes such that it
would provide better performance indices.
Practically the mismatch filters are used to provide better PSR, with some SNR
loss, than matched filter. Various mismatch filters such as adaptive linear combiner
(ALC) and neural networks are used to suppress the sidelobes [38, 39, 41]. However,
the convergence of the neural network is slow during the training period. Hence
new efficient structures and the corresponding learning algorithms having faster
convergence are required for pulse compression.
Apart from biphase codes, polyphase codes [80, 81] and frequency modulated
codes are also used in radar systems. In the literature different type of windows
are used as weighing function to suppress the sidelobes [89, 90] of polyphase codes
15
Introduction
Chapter 1
and LFM waveforms [71]. Under the Doppler shift conditions the PSR magnitude
provided by the windows are low. Hence efficient amplitude weighing techniques are
needed to achieve lower sidelobes in Doppler shift conditions.
In phased array radar, wide bandwidth waveforms are used to acquire high
range resolution. Generation of such types of waveforms increases the overall cost
and complexity of the system. The conventional hardware designed for narrowband
signals in radar systems may not sustain instantaneous wide bandwidth. To
overcome such limitation the wide bandwidth signal is split into a set of narrowband
signals which are transmitted and received separately.
The effect of wideband
signal is obtained by coherently combining the narrowband signals. Such type
of narrowband signals taken together is called ‘synthetic wideband waveform’ or
‘stepped frequency waveform’ or ‘frequency jumped train’. However, the matched
filter output i.e. ACF of such signals suffers from grating lobes due to constant
frequency step. Therefore there is a need to design a signal having wide bandwidth
but can be processed by the hardware for narrow band signals and its ACF has
suppressed grating lobe, low peak sidelobes and narrow mainlobe width.
1.5
Motivation
Substantial effort has been made to suppress the sidelobes of the different waveforms
using computational intelligence (CI) tools such as evolutionary computing
techniques and neural networks.
• Several existing evolutionary computing techniques have been employed to
assign the phase to different bits of biphase codes using weighted sum of PSL
and merit factor (MF) as cost function [14, 15]. The problem associated with
these methods is to choose the appropriate values of the weights.
• The matched filter does not provide adequate PSR for many radar applications.
Hence to obtain improved PSR, the mismatch filters have been introduced
whose weights are adapted using known input output data. Various mismatch
16
Introduction
Chapter 1
filters using multilayer perceptron (MLP) and radial basis function (RBF)
networks have been reported in the literature. However these filters provide
poor convergence performance and hence the magnitude of PSR is less during
detection. Thus there is need to design improved mismatch filters.
• For polyphase and LFM waveforms the amplitude weighing techniques are
used at the receiver to suppress sidelobes. The targets in the environment
are not always stationary. If the target is in motion, the reflected waveform
is Doppler shifted version of the transmitted waveform. When this Doppler
shifted waveforms are passed through the weighted receiver matched filter the
PSR degrades. Under such situations it is required to improve the PSR.
• The matched filter output i.e. ACF of wide bandwidth stepped frequency LFM
pulse train suffers from grating lobes due to constant frequency step. Several
methods have been implemented to suppress the grating lobes in [113, 114].
These methods generally ignore the PSL and mainlobe width which are
also important measures of the performance for target detection. Therefore,
algorithms need to be developed to choose the parameters of stepped frequency
waveform such that the output of the matched filter provides high range
resolution, lower grating lobes and reduced sidelobes.
Based on the aforementioned motivations, the objectives of the research work of
this thesis is developed. The thesis employs evolutionary, soft computing and signal
processing techniques to solve these problems of pulse compression.
1.6
Objective of the thesis
The main objective of present research work is to propose efficient pulse compression
techniques for different radar signals. The various objectives may be listed as:
• To generate pulse compression biphase codes having lower peak sidelobes and
better MF using multiobjective algorithm.
17
Introduction
Chapter 1
• To develop efficient sidelobe reduction structures using neural networks which
converge faster during the training time as well as provide higher magnitude
of PSR.
• To introduce and assess amplitude weighing technique for LFM waveform and
polyphase codes which is expected to provide better PSR at higher Doppler
shifts.
• To select appropriate parameters of LFM pulse train to achieve reduced grating
lobes, low peak sidelobe level and narrow mainlobe width.
1.6.1
Structure and chapter wise contribution of the thesis
Chapter-1
The concept of pulse compression, matched filter, ambiguity function and
different radar signals are introduced in this chapter. The motivation behind the
application of evolutionary, neural network and signal processing techniques for
pulse compression is outlined. The summary of framework of the research and
contributions are also included.
Chapter-2
In this chapter the biphase codes having lower PSL and better MF in their ACFs
are generated. Genetic algorithm (GA) is used to optimize a cost function consisting
of weighted combination of PSL and MF. However there is difficulty in selection of
proper weight value to optimize the combination. In order to overcome this difficulty
a multiobjective algorithm (based on nondominated sorting genetic algorithm-II
(NSGA-II) ) is proposed which simultaneously optimize the PSL and MF. The
proposed algorithm provides a set of nondominated solutions. Simulations have
been carried out using proposed algorithm to generate pulse compression biphase
codes for length 49 to 59. NSGA-II provides more than one nondominated codes
for each length. A particular code of specified length is chosen in accordance to
the requirement of the environmental condition. If the environmental condition is
18
Introduction
Chapter 1
dominated by distributed clutter then the code having high MF is preferred. On
the other hand if the application requires the detection of target in presence of large
discrete clutter the code having low PSL is chosen.
Chapter-3
Several mismatch filters are investigated in this chapter which provide better PSR
values as compared to the matched filter.
The best binary codes available in
the literature are known as Barker codes having maximum sidelobe level of unity
amplitude. The largest Barker code available is of length 13 having a PSR of
magnitude 22.3 dB which is not adequate for many radar applications. The Barker
codes of larger length are generated by taking Kronecker product of existing Barker
codes. To obtain higher PSR value the mismatch filters such as adaptive linear
combiner, multilayer perceptron (MLP) and radial basis function (RBF) along
with their learning algorithms are investigated. The convergence performance of
MLP and RBF structures is very slow. Therefore recurrent neural network (RNN)
and recurrent RBF (RRBF) structures capable of yielding faster convergence are
proposed for the pulse compression filter. The shifted version of 13-bit and 35-bit
Barker codes are used as input to the different networks. The desired output of
the network is always zero except at one point corresponding to the presence of
target. The convergence rate during training for RNN and RRBF are compared to
that of MLP and RBF networks. After the training is complete the networks are
used for pulse radar detection. The PSR values of RRN and RRBF for different
noise conditions, presence of multiple target and under Doppler shift condition are
evaluated and compared with those of MLP and RBF.
Chapter-4
This chapter presents pulse compression for LFM waveforms and polyphase codes.
The LFM and polyphase codes have lower sidelobes compared to the biphase codes.
LFM waveforms are more Doppler tolerant than phase coded waveforms. Polyphase
codes are derived from the LFM waveform to get the advantage of the Doppler
tolerant property of the LFM waveform. The matched filter output of the LFM
19
Introduction
Chapter 1
waveform yields PSR of -13.2 dB. Different weighing functions are used in the
receiver to achieve high PSR magnitude and the LFM waveform is amplitude tapered
or phase distorted before transmission to get even higher PSR magnitude. The
weighing functions are also used for sidelobe suppression of polyphase codes. If
the target is in motion then the reflected signal is Doppler shifted version of the
transmitted signal. In this chapter convolutional windows are proposed to use as
weighing function for LFM and polyphase codes to achieve better PSR values in
Doppler shift conditions. Simulation study is carried out to assess the performance
of the convolutional windows and is compared to those of conventional windows.
Chapter-5
In this chapter evolutionary computing techniques are proposed to determine
the parameters of stepped frequency LFM pulse train. In case of high range
resolution radar the required bandwidth is very high. The conventional narrowband
hardware may not support the instantaneous wide bandwidth. Therefore, the
wide bandwidth signal is split into narrowband signals which are transmitted and
combined coherently at receiver to get the effect of the wideband signal. But the
ACF of such narrow band pulse train suffers from grating lobes and hence destroys
the range resolution capability of the pulse train. In the proposed work the particle
swarm optimization (PSO) technique is used to determine the parameters of the LFM
pulse train such that all the grating lobes are cancelled. Apart from cancellation
or suppression of grating lobe, minimization of mainlobe width and peak sidelobe
level of ACF are also important for the radar systems. In this chapter NSGA-II
algorithm is proposed to choose the parameters of stepped frequency LFM pulse
train to accomplish reduced grating lobes, low peak sidelobe and narrow mainlobe
width.
Chapter-6
In this chapter the overall contributions of the thesis are reported. This chapter also
contains the details of further research work which can be attempted subsequently.
20
Introduction
Chapter 1
1.7
Conclusion
This chapter provides a brief introduction to radar, pulse compression technique and
different signals used in radar. The merits and demerits of the pulse compression
technique are studied. It also systematically outlines the scope, the motivation
behind this work and the objectives of the thesis. In essence, this chapter provides
an overview of the thesis in a comprehensive manner.
21
Chapter 2
Generation of Pulse Compression
Codes Using Multiobjective
Genetic Algorithm
2.1
Introduction
In a pulse radar system the transmitted pulse width should be as long as possible to
increase the sensitivity of the system and as small as possible at the receiver for better
range resolution. Range resolution is the ability of the radar receiver to discriminate
nearby targets. The performance of range resolution radar would be optimal, if the
coded waveform has impulsive ACF. Biphase coded waveforms support better range
resolution compared to LFM pulses because the windowing functions used with LFM
pulses to lower time sidelobes cause a broadening in the mainlobe. But the ACF
of biphase coded waveforms contain higher range sidelobes, which have a negative
influence on the detection performance of radar systems. A desirable property of
the compressed pulse is that it should have low sidelobes in order to prevent a
weaker target from being masked in the sidelobes of a nearby stronger target. The
lower the sidelobes relative to the mainlobe peak, the better the main peak can be
distinguished and hence the better is the corresponding code. The selection of a pulse
compression code depends on the application and the environmental conditions. If
the application is radar designed for a scenario dominated by distributed clutter,
then integrated sidelobe level (ISL) is very important. On the other hand if the
22
Chapter 2
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
application requires detection of targets in the presence of large discrete clutter, then
the PSL is more important. If the desired ISL or PSL performance is not achieved
with a matched filter, a mismatch filter is used to achieve the desired sidelobe level
with some SNR loss.
Binary pulse compression codes [7] such as the Barker codes [8] or maximal-length
sequences [9] are extensively used in radar systems. The Barker codes which are
known as the best ideal waveform can provide a maximum PCR of 13. Many
practical applications require longer codes to achieve higher PCRs much greater
than 13. Therefore sequences with the lowest possible sidelobes at the longer length
are needed. There is no analytical technique available to construct a sequence for
a given PSL. Time consuming and money consuming exhaustive computer search
program are generally used to generate best possible sequences. By exhaustive
computer search program, Lindner [10] found all binary sequences up to length 40
with minimum PSL. With an improved algorithm Cohen et al. [11] further extended
those results to sequence length 48. For larger sequences some heuristic methods,
such as neural network (NN) and evolutionary algorithms (EAs) are proposed to
search the binary sequences with good aperiodic autocorrelation [12–15]. Using an
NN approach, Hu et al. [12] obtained useful binary sequences for lengths 49 up to 100.
An objective function which consists of weighted sum of PSL and merit factor (MF)
is optimized using genetic algorithm (GA) to generate codes from 49 to 100 [15]. The
demerit of this type of objective function is to choose the accurate weight values. It
is also required to run the program repeatedly for different combinations of weight
values. To overcome this problem, in the proposed work a multiobjective algorithm
is introduced in which PSL and MF are used as two different objective functions to
generate the biphase codes.
23
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
Chapter 2
2.2
Merit measures and problem formulation
Let an L length binary sequence is given by
S = {s1 , s2 , s3 , · · · , sL }
(2.1)
where each element of S has a value either +1 or -1.
The ACF of S for positive delays is given as
Ck (S) =
L−k
X
si si+k
(2.2)
i=1
where k = 0, 1, 2, · · · , L − 1.
Ideally, the range resolution radar signal should have high ACF value for zero shift
and zero value for nonzero shift. A significant problem inherent in biphase pulse
compression is that the ACF does not yield a perfect impulse, that means it does
not produce Ck (S) = 0 for k 6= 0. Any non zero value of Ck (S) for k 6= 0 is referred
to as sidelobe where as the zero-offset correlation value C0 (S) is called the mainlobe.
The difference between a pulse compression waveform and a simple pulse waveform
lies in the existence and magnitude of these sidelobes. These sidelobes limit the
usefulness of a code regardless of the strength of the mainlobe. Codes are chosen for
a given application based on their length and sidelobe levels.
There are two main criteria [16, 17] used to decide the goodness of a pulse
compression code. The first one is the PSL which is the largest sidelobe in the
ACF of the code and defined as
P SL = M ax |(Ck (S))| , k 6= 0
(2.3)
The second one is the merit factor MF which is defined as the ratio of energy in
the main peak of the ACF to the total energy in the sidelobes. As the signal is real
valued the ACF is real and symmetric about the zero delay. The MF is represented
as
MF =
C02 (S)
PL−1 2
2 k=1 Ck (S)
24
(2.4)
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
Chapter 2
The denominator of (2.4) is known as ISL. For a good sequence or code the PSL
should be low and MF should be high. To optimize simultaneously PSL and MF
using GA, the fitness function is defined as
f=
α
+ βM F
P SL
(2.5)
The fitness function f is maximized using suitable values of α and β such that
α + β = 1. However it is a difficult task to choose a proper combination of α and β
to get a optimized code. Hence in the proposed work nondominated sorting genetic
algorithm -II (NSGA-II) is used to optimize multiple objective functions PSL and
MF simultaneously to generate biphase pulse compression codes.
2.3
Techniques used
The techniques which are used to generate pulse compression codes are described in
this section.
2.3.1
Genetic algorithm
The GA is a programming technique that mimics biological evolution process and
uses the genetic operators such as selection, crossover and mutation for problem
solving strategy. GAs are based on Darwin’s theory of evolution i.e. the strong
survivors have better opportunity to transfer their genes to future generations
through reproduction. Species those carry correct combination in their genes are
dominant in the population. Sometimes during the process of evolution, random
changes may occur in genes. If these changes render advantages for the survival,
new species evolve from the old ones. In other words unsuccessful changes are
eliminated by the natural selection.
The GA was originally proposed by J. Holland [18] in 1975 which imitates
nature’s robust way of evolving successful organisms. Afterwards it became popular
due to the publication of D. Goldberg’s book in 1989 [19]. Since then the GAs have
been used in a wide range of applications where optimization is needed. In the GA,
25
Chapter 2
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
a solution is called as an individual or chromosome and each element of chromosome
called as genes. The GA works on a set of chromosomes called as a population. As
the search evolves, the population have fitter and fitter solutions. The various steps
involved in the GA are follows
1. Population initialization
Population size is the number of chromosomes required in one generation. If there
are too few chromosomes, GA have a few possibilities to perform crossover and only
a small part of the search space is explored. On the other hand, if there are too many
chromosomes, GA process will slow down. A chromosome is represented in such a
way that it should contain information about the solution. The chromosomes are
presented in real numbers such as 0.5, -0.3, 1.5 etc or encoded to binary form, using
an encoding process, such as ‘1001010101’, ‘1001001010’ etc. This chapter is dealt
with only binary representation of the chromosomes. M number of chromosomes
are randomly initialized with binary forms.
2. Fitness function evaluation
The initialized population is used to evaluate the objective function which is to be
optimized. This is known as fitness function evaluation since the objective function
value corresponds to the fitness of that chromosome.
3. Selection
Chromosomes from the population are selected by using a mechanism to enter into
a mating pool. Chromosomes from the mating pool are used to produce offspring
which form the basis for the next generation. As the genes of the chromosome
are to be inherited to the next generation, it is desirable the mating pool should
contains good chromosomes. So a selection procedure in GA is used to select
better individuals in the population for the mating pool. The selection pressure
is the degree to which the better chromosomes are favored. The higher the selection
26
Chapter 2
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
pressure, the more the better individuals are favored. Selection pressure helps GA
to improve the population fitness over succeeding generations.
There are many selection methods available in the literature such as roulette
wheel, tournament, rank and steady state selection [20] etc. In this thesis a binary
tournament selection is used to choose a chromosome from the population. In this
selection a tournament size consisting of two chromosomes are randomly chosen
and the winner of the two is the chromosome with the highest fitness value. The
winner is entered into the mating pool. As the mating pool comprised of tournament
winners, it has a higher average fitness than average population fitness. This fitness
difference provides the selection pressure, which helps GA to improve the fitness of
each succeeding generation.
4. Genetic operators
Genetic operators such as crossover and mutation are used to explore and exploit new
and better solutions from the existing solutions in the search space. The operators
are explained below.
a. Crossover
In this operation two chromosomes, called parents, are selected using binary
tournament selection from the existing population and combined together to form
new chromosomes. These newly formed chromosomes are called offspring. It is
always expected that offspring inherits good genes from the parent. A single point
crossover and a two point crossover are shown in Figure 2.1. In case of single point
crossover a point is randomly selected and all the genes after this point are swapped
between the two parent chromosomes to form two offspring. Similarly, in case of two
point crossover two points are randomly selected and the genes in between the two
points are swapped between the two parents to form two offspring. This operation
is carried out with certain probability called as crossover probability which indicates
how often crossover will be performed. If there is a crossover, offspring is made from
parts of parents chromosome. If crossover probability is 100%, then all offspring is
27
Chapter 2
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
(a) Single Point Crossover
(b) Two Point Crossover
Figure 2.1: Crossover
made by crossover. If it is 0%, whole new generation is made from exact copies of
chromosomes from old population. Crossover is made in hope that new chromosomes
will contain good parts of the old chromosomes and therefore the new chromosomes
are better. However it is good to leave some part of the old population survive to
next generation.
b. Mutation
It takes place at the gene level. It introduces random changes into the features of
chromosomes. In GA the probability of mutation is smaller in comparison to the
probability of crossover. If there is no mutation, offspring is taken after crossover
(or copy) without any change. If mutation is performed, part of chromosome is
changed. If mutation probability is 100%, whole chromosome is changed and if
it is 0%, nothing is changed. Mutation reintroduces the genetic diversity back
28
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
Chapter 2
into the population and helps to escape from the local minima. In case of binary
representation of codes a randomly chosen bit is switched from 1 to 0 or 0 to 1 as
shown in Figure 2.2.
Figure 2.2: Mutation
5. Recombination and selection
This process is used to weed out the weaker chromosomes from the population so
that the more productive chromosomes will be used in the next generation. In most
of the cases the fitness function value of a chromosome decides its survival for the
next generation. The current generation population is combined with the offspring
population and the fitness values of each chromosome of the combined population
is evaluated. The best M chromosomes are selected according to the fitness value
to carryout the next generation.
A flow chart for GA operation is depicted in Figure 2.3.
2.3.2
Multi objective GA
In single objective problems one has to find out the best solution which is usually the
global maximum or minimum relying on the problem. In practice, the optimization
problem is associated with multiple, possibly conflicting, objectives and this type
of problem may not have one best solution with respect to all the objectives. A
set of solutions exists in the search space which are superior to rest of the solutions
with respect to all the objectives but are inferior among themselves with respect
to one or more objectives. These solutions are called as nondominated solutions or
Pareto optimal solutions. None of the nondominated solutions is better than the
29
Chapter 2
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
Figure 2.3: Flow chart for GA
30
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
Chapter 2
other or in other words every solution is an acceptable solution. The superiority
of one solution over the other depends upon the knowledge of the problem and its
application. Thus, a solution chosen by a designer may not be accepted by another.
A multiple objectives problem can be solved as a single objective problem by
assigning a weight wi to each objective as follows: minimize
z = w1 z1 (x) + w2 z2 (x) + ..... + wk zk (x)
where z1 (x), z2 (x), ....zk (x) are the objective functions and
(2.6)
Pk
i=1
wi = 1.
In this approach a single set of weight vector produces a single solution.
If
multiple solutions are required the problem has to run repeatedly for different
set of weights. The drawback of this type of approach is judicious selection of a
weight vector for each solution, which is a difficult task. To overcome this difficulty
many multiobjective evolutionary algorithms are found in literature to get a set of
nondominated solution in a single run. In [21–26], the evolutionary algorithms are
amply demonstrated and it is found that these are efficient to find multiple and
diversified nondominated solutions. The difference between single objective GA and
multiobjective GA (MOGA) is the concept of dominance used directly or indirectly
in the selection phase of MOGA. The effective MOGA approximates the true Pareto
front and maintains diversity in the population [21]. Schaffer [22] has proposed the
first practical multiobjective algorithm, called as vector evaluated GA (VEGA). This
algorithm solves each objective separately and then combines sub solution of each
objective. One of the demerits of this algorithm is that it is biased towards some
of the Pareto optimal solutions. In [23], an MOGA is proposed which explores the
solution in all possible directions in the search space. Subsequently many GAs have
been proposed by many researchers to find out improved nondominated solutions in
the objective space. These algorithms are efficient in terms of complexity, rate of
convergence, diversity among the nondominated solution and the interval distance
from the Pareto optimal front.
Deb and Srinivas [25] have proposed a robust
popular nondominated sorting genetic algorithm (NSGA) to solve multiobjective
optimization problems. But this algorithm involves high computational complexity,
31
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
Chapter 2
lacks elitism and difficulty in choosing the optimal value of sharing parameter. An
improved version of NSGA, called NSGA-II, is dealt in [26] which uses the concept
of elitism and does not use the sharing parameter. The various steps of NSGA-II
algorithm are given below.
1. Population initialization:
The population contains a set of M chromosomes. Each chromosome is initialized
randomly with binary bits.
2. Fitness function evaluation:
The fitness functions which are to be optimized are evaluated for each chromosome.
3. Nondominated sort:
The initialized population is sorted according to nondomination.
The sorting
algorithm [26] is given below.
• for each solution x in the main population X do the following
– the domination counter nx , the number of solution that dominate the
solution x, is initialized as zero i.e nx = 0.
– Sx , a set which contains all the solutions to those the solution x dominates,
is initialized as an empty set φ i.e. Sx = φ
– for each solution y in X
∗ if x dominates y
· y is added to the set Sx i.e Sx = Sx ∪ {y}.
∗ else if y dominates x then
· the domination counter of x is incremented i.e. nx = nx + 1.
– if nx = 0 i.e. no solution dominates x then it belongs to the first front.
Assign rank one to the solution i.e. xrank = 1. The first front is updated
by adding x to it i.e. F1 = F1 ∪ {x}.
32
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
Chapter 2
• This process is executed for all solutions in X.
• The front counter i is initialized to one i.e. i = 1.
• The following steps will be executed until ith front is non empty i.e.Fi 6= φ.
– A set Y = φ is defined to store the solutions for next front.
– for each solution x in front Fi
∗ for each solution y in Sx
· ny = ny − 1, the domination count for solution y is decreased.
· if ny = 0, then y belongs to the next front. Hence yrank = i + 1.
The set Y is updated as Y = Y ∪ {y}.
– The front counter incremented by one i.e. i = i + 1.
– Y is set as next front i.e. Fi = Y .
4. Crowding distance:
An efficient multiobjective algorithm not only converges to the true Pareto optimal
set but also requires good spread or diversity among the obtained solutions. The
original NSGA [25] uses a sharing parameter to achieve the diversity among the
solutions. The difficulties of this algorithm are choosing the sharing parameter value
and associated heavy computational complexity. These difficulties are overcome in
NSGA-II by providing better diversity among the solutions using the concept of
crowding distance. It does not require any user defined parameter to maintain the
diversity among the solutions. The crowding distance is calculated front wise as
follows.
• For any front Fi , l is the number of solutions i.e. |Fi | = l.
– The distance of all the solutions are initialized to zero i.e. Fi (Dj ) = 0,
where the index j corresponds to j th solution in front Fi .
– for each objective function m
33
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
Chapter 2
∗ The solutions in front Fi are sorted in ascending order according to
the objective function value i.e. I = sort(Fi , m)
∗ Infinite distance value is assigned to the boundary solutions of front
Fi i.e. Fi (D1 ) = Fi (Dl ) = ∞
∗ for j = 2 to l − 1
Fi (Dj ) = Fi (Dj ) +
I(j+1)·m−I(j−1)·m
max −f min
fm
m
where I(j) · m is the mth objective function value of j th solution in
min
max
I. fm
and fm
are minimum and maximum value of mth objective
function.
• The above procedure is carried out for all the fronts.
The solution that has large crowding distance value means it is far away from others,
hence it is selected first.
5. Selection:
a. Based on nondomination rank: A solution is selected if its nondomination rank
is lower than other.
b. Based on crowding distance: If two solutions belong to the same front, the solution
having higher crowding distance is selected.
The selection procedure is used during binary tournament selection and population
reduction phase.
6. Genetic operators:
These operators are used to produce offspring from the parent. The process of
crossover and mutation are carried out as explained in Section 2.3.1
7. Recombination and selection:
The current generation population is combined with the offspring population and
selection is carried out to choose the best M solutions for next generation. In this
34
Chapter 2
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
algorithm elitism is ensured as all the previous population is added with the current
population for selection process. The algorithm for selection process is given below.
• Let Ag and Bg are current and offspring population respectively of g th
generation. The combined population Cg = Ag + Bg .
• Sort the population Cg according to nondominated sort and determine all the
fronts (F1 , F2 , ....).
• Initialize i = 1
• The following is carried out until (|Cg+1 | + |Fi |) ≤ M
– Assign the crowding distance to the solutions of the front Fi .
– Update Cg+1 by adding all solutions of Fi to it i.e. Cg+1 = Cg+1 ∪ Fi .
– i = i + 1.
• Sort the solutions of front Fi in descending order of the crowding distance.
• Update Cg+1 by adding first (M − |Cg+1 |) solutions of Fi i.e. Cg+1 = Cg+1 ∪
Fi [1 : (M − |Cg+1 |)]. The new population Cg+1 having M chromosomes is used
for next generation.
The NSGA-II procedure and a flow chart is presented in Figures 2.4 and 2.5
respectively.
2.4
Generation of pulse compression codes
2.4.1
Using genetic algorithm
The fitness function defined in (2.5) is maximized to generate the biphase codes.
The various steps are
1. The codes are to be generated are biphase codes i.e. they are consists of 1
or -1. So the chromosomes which are initialized must in the form of 1 or -1.
35
Chapter 2
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
Figure 2.4: NSGA-II procedure
M number of chromosomes are randomly initialized with binary bits (1 or -1)
with each chromosome of length same as the code length to be generated.
2. The ACF, PSL and MF for each chromosome are calculated according to
(2.2), (2.3) and (2.4) respectively. The objective function value which is to
be maximized is evaluated as given in (2.5).
3. The chromosomes are selected according to the binary tournament selection as
described in Section 2.3.1. The selected chromosomes are used for off-spring
generation.
4. The offspring are generated using the genetic operators such as crossover and
mutation as explained in Section 2.3.1. In this case the binary bits with 1 and
-1 are used as shown in Figure 2.6 and 2.7.
5. The current generation population is combined with the parent population and
best M chromosomes are selected according to the fitness value to carryout
the next generation.
Steps from 3 to 5 are carried out until the maximum number of generation is met.
36
Chapter 2
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
Figure 2.5: Flow chart for NSGA-II
37
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
Chapter 2
(a) Single Point Crossover
(b) Two Point Crossover
Figure 2.6: Crossover using binary bits 1 and -1
Figure 2.7: Mutation using binary bits 1 and -1
2.4.2
Using NSGA-II
In the proposed work NSGA-II algorithm is employed to optimize the two objective
functions PSL and MF as explained in Section 2.2. The different steps are
38
Chapter 2
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
1. The population is initialized with random binary bits (1 or -1) as in case of
GA.
2. The two objective functions are evaluated as given in (2.3) and (2.4). PSL is
minimized and MF is maximized simultaneously using the NSGA-II algorithm.
3. The chromosomes are sorted according to nondominated sort and found out
all possible fronts as described in Section 2.3.2.
4. The crowding distance for chromosomes in each front are evaluated according
to the procedure explained in Section 2.3.2.
5. The chromosomes are selected using binary tournament selection according to
Section 2.3.2.
6. The selected chromosomes undergo genetic operations such as crossover and
mutation to produce off-springs as explained in case of GA.
7. The off-spring population is combined with parent population and the best M
chromosome selected for next generation as described in Section 2.3.2
The steps from 5 to 7 are carried out until the maximum number of generations is
met.
2.5
Simulation results
The GA dealt in Section 2.3.1 is used to maximize the fitness function f given in (2.5)
to obtain the desired binary string. The population size M is chosen as 250 for each
code generation. Each chromosome is randomly initialized and the fitness function
f is calculated for all the chromosomes for a given combination of α and β. The
process of selection, crossover and mutation are carried out to produce offspring. The
two point crossover is used to generate the offspring. The probability of crossover
and mutation are chosen to be 0.8 and 0.2 respectively. The current generation
population is mixed with the offspring population and the best M chromosomes are
39
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
Chapter 2
Table 2.1: Sequences obtained using GA
Seq Len
49
50
51
52
53
α
0.9
0.8
0.7
0.6
0.5
0.9
0.8
0.7
0.6
0.5
0.9
0.8
0.7
0.6
0.5
0.9
0.8
0.7
0.6
0.5
0.9
0.8
0.7
0.6
0.5
β
0.1
0.2
0.3
0.4
0.5
0.1
0.2
0.3
0.4
0.5
0.1
0.2
0.3
0.4
0.5
0.1
0.2
0.3
0.4
0.5
0.1
0.2
0.3
0.4
0.5
PSL
5
5
6
6
6
5
5
6
6
7
5
6
6
6
7
5
6
6
6
8
6
6
6
7
7
MF
3.5309
4.0557
4.2271
4.2875
4.6173
3.1807
3.6232
3.7994
3.8941
3.9936
3.6429
3.6841
3.7264
3.7696
3.9529
3.9075
4.0479
4.2516
4.3057
4.7943
3.9075
3.6767
3.7553
4.2561
4.5306
Sequences
1101011100101010110110100001101100111111101100111
1001001110010111100000001010100001000011001011001
1100010111011100101110101001111001101000000100101
0101011000011110011100000010111101110111110110011
1111111100011010001110100010101101001100111100100
11010000100100000010100101110101110000010011011100
11101101100010010000101010001001111001000011001111
10011000011001000110010101001010011111010000010000
01010100010100101101110110110011000110001111100000
10100011001100101001111010001111101111010100000110
100011100101011010001100000001011111000100100100100
101101000101100000100101000011010111110011001100111
001011011111110110100110010101010000110111000011110
010101100011000101010111000000111110111011011011010
011100101110000010011100001001110111101101100101001
1110011110011001000011111000000010010010101010101101
1011111011001011111110000110000101001010110011001100
0011111101100011100010010001111110101101001101110101
1010101001110001110100100110011111100000000100100001
1101011010000000010011110000111011100010111010011001
10011111101100011101101011000111010111111010011010001
00101101110011101110110101111000011000001011111010100
01010010110110110000011110001100010001111110111011101
11000110100101100101100110011000000001111101010111101
01010100011000101111001101101100111101011011110100000
selected to carry out the next generation. The algorithm is run for 100 generations
for each code. For different combinations of α and β the obtained codes from 49
to 53 are tabulated in Table 2.1. 0’s are used in place of -1’s to conserve space.
From Table 2.1 it is observed that by giving different weightage to PSL and MF
different codes are obtained. A particular code is selected according the requirement
of application such as low PSL or high MF. If it requires low PSL, then a high value
of α is required. If it requires high MF, then a high value of β is required. It is
too difficult to choose the appropriate values for α and β to get an optimized code.
The combination α = 0.8 and β = 0.2 produces better code as compared to the
combination α = 0.9 and β = 0.1. Because in both the cases PSL is same but for
the weight combination α = 0.8 and β = 0.2 provides better MF.
40
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
Chapter 2
Table 2.2: Sequences obtained using NSGA-II
Seq Len
49
50
51
52
53
54
55
56
57
58
59
PSL
4
5
6
5
6
7
5
6
5
6
5
6
5
6
7
5
6
5
6
7
5
6
7
5
6
5
6
7
MF
4.2875
5.0869
5.1746
3.7994
3.8941
4.4484
4.3788
5.1403
4.4768
4.7943
3.9675
4.0129
3.8880
3.9728
4.1538
3.6446
3.6800
3.4087
4.0412
4.2609
3.1240
4.1441
4.6151
3.7798
4.8195
3.5020
4.2555
4.4288
Sequences
0011111000111111011000000100101100111001101010101
0011111100111111011000000100101100110100101010101
0011111100111111011010000100111100110100101010101
00010010011100111100100101000110111010101111111010
10010000011100111100100101000110111010101111111010
10010000111100111100100101000110111010101111111010
110000010000000110100110011100010001011110010101101
100010010000000110100110111100010001011110010101001
0000110011001100001111000000001010001011010101101101
0101001110001110100101110110110100000001101100010011
11001001101001000000010001001110101111000010101110001
11111011011101101000011100111001000010111010101100100
011110101101001000100000110000111010100010011101110100
110110101101001000100000110000111010100010011101110000
010110101101001000100000110000111010100010011101110000
1000111100010001011111001000110101011111110101101001001
1001101100010001111110001000011101011111110101111001011
00111000100011011100011010001000000010011011010000111101
00111010100011011100011010001000000010111011010000100101
00111010100011011100011010001000000010011011010000100101
110011010111110101110110101110000100010010000111111011000
110110010111110101110110101110001100010010000111111011000
110110010111010101110100101110001100010010000101111011000
1001010001110111000101011111000011101001110110010000000010
1001010001110011000101011011000011101001110110010000000000
00000100101100001111111100010101100011110100011001001110101
00000101101101001111101100010101110011110100011011001110100
00000100101101001111111100010101110011110100011001001110101
To overcome the difficulty of choosing the values of α and β and to get the
optimized code in a single run, the code is generated by using multiobjective
GA. PSL and MF are the two objective functions used for optimization using
NSGA-II algorithm as explained in Section 2.4.2. The population size M is taken
as 250 and the chromosomes are initialized randomly. The PSL and MF for each
chromosome are found out and the population is sorted based on nondomination.
Each chromosome in the first front have a rank value of 1 and the chromosome
in the second front is assigned a rank value of 2 and so on. Crowding distance is
assigned front wise to each chromosome. Parents are selected from the population
41
Chapter 2
Generation of Pulse Compression
Codes Using Multiobjective Genetic Algorithm
using binary tournament selection based on rank and crowding distance.
The
selected population generate offspring using crossover and mutation operations. The
probability of crossover and mutation are chosen same as earlier. In this case also two
point cross over is used to generate offspring. The offspring population is recombined
with the current population and the best M chromosomes are selected for next
generation. The number of generations is taken as 100. The sequences found by
the proposed method for length 49 to 59 are listed in Table 2.2. In this case all the
possible nondominated solutions can be achieved in a single run. For code length
49 the lowest PSL obtained is 4 and corresponding MF is 4.2875 which is better
the lowest PSL obtained using GA i.e. 5 and corresponding MF is 3.5309. For a
particular length, NSGA-II provides more than one solution. A solution is chosen
according to the requirement of the application such as better PSL or better MF.
2.6
Conclusion
By using NSGA-II algorithm a list of biphase sequences of length 49-59 has been
generated and is listed in Table 2.2 along with their PSL and MF values. The
results reveal that the proposed method performs better than the weighted sum
approach in GA. The search for optimum sequence depends on the selection of the
initial population of parent sequences. As the sequence length increases the search
procedure requires more time for obtaining a good solution. The quality of solution
improves with increase in the number of generations.
42
Chapter 3
Development and Performance
Evaluation of New and Efficient
ANN Mismatch Filters for
Sidelobe Reduction
3.1
Introduction
The objective of pulse compression technique is to achieve appreciable PSR and
acceptable range sidelobes in an economical manner. The types of waveforms used
in this technique decide the cost and complexity of the radar system. The binary
phase codes have better range resolution as compared to frequency coded waveforms.
This advantage is obtained at the cost of high range sidelobes [27,28]. The reduction
or elimination of range sidelobes can be achieved in a pulse compression radar by
the use of a pulse which is coded in both amplitude and phase [29]. However, these
techniques are seldom used because of the expensive amplitude modulation circuitry.
With the increase in compactness and decrease in cost of the digital circuits due to
the revolution of very large scale integrated (VLSI) circuits, it is appropriate to
implement complex techniques offering improved performance. Therefore, efforts
were made to devise alternative methods which could provide acceptable range
sidelobes.
The range sidelobes of binary phase codes are reduced by using different types
43
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
of digital filters whose parameters are determined by use of various algorithms. The
efficiency of different approaches are evaluated by the ISL, PSL, hardware complexity
involved and loss in SNR as compared to matched filter. Two methods are generally
used to suppress the sidelobes. First one employs an additional weighing network
after the matched filter. Rihaczek and Golden [30] have synthesized a filter in
frequency domain for 13-bit Barker code which is a simple network and is able to
suppress the sidelobes to an acceptable level. In this technique the complexity of
the digital processor is reduced due to presence of few tap weights in the tap-delay
filter. The second method is to design a mismatched filter directly [31, 32] instead
of placing a weighing filter after the matched filter. In [31], an optimum mismatch
filter is developed for 13-bit Barker code in least square (LS) sense, which gives
optimal performance in terms of ISL. The weights of this filter is designed in such
a way that the response which approximates (in least square sense) to a sequence,
has all its elements as zeros except for one non zero element present at the central
position. Zoraster [32] has used the linear programming (LP) to determine the filter
weights for reduction of the PSL of 13-bit Barker code. To achieve satisfactory peak
sidelobe and integrated sidelobes, the length of the LP filter should be very large
which effectively increases the hardware cost. Hua and Oksman [33] have combined
the advantages of [30] and [32] to obtain a new algorithm which provides lower peak
sidelobes for 13-bit Barker code. In this method the transfer function of the sidelobe
suppression filter is fitted with a polynomial expansion series in frequency domain,
which consists of some unknown expansion coefficients. By applying inverse Fourier
transform and LP, the coefficients of the transfer function can then be determined.
A mismatch filter in cascade with a finite impulse response (FIR) filter is used in [34]
to suppress the side lobes of Barker codes which requires less multipliers and adders
due to its symmetry.
With the advancement of adaptive signal processing and neural networks,
researchers have put their efforts to design the sidelobe reduction filters using these
techniques. Sidelobe reduction using adaptive filters are discussed in [35, 36] where
44
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
various algorithms like the least mean square (LMS), recursive least square (RLS)
and modified RLS algorithms have been used to reduce the peak sidelobe level.
Among these algorithms the modified RLS yields the highest PSR magnitude of
25.7 dB for 13-bit Barker code which is not suitable for many radar applications.
This necessatiates the use of improved techniques like the neural networks. The
neural network such as MLP has been applied [37–39] for pulse compression which
provides PSR magnitude more than 40 dB for 13-bit Barker code. The weights
of the MLP have been determined by training the network with all possible input
patterns. Kwan and Lee [38] have employed back propagation (BP) algorithm and
achieved acceptably good results. But the convergence speed of the BP algorithm is
inherently slow and the network is sensitive to Doppler shift [39]. To overcome this
drawbacks, the recurrent neural network (RNN) and recurrent radial basis function
(RRBF) networks are proposed for pulse compression.
3.2
Problem formulation
In biphase codes the transmitted pulse of duration Tp is divided into N sub
pulses each of duration tb =
Tp
.
N
The ACF of transmitted code mathematically
is represented [40] as
N −|k|
1 X
xi xi+|k|
yk =
N i=1
k = −N + 1, ....., N − 1
(3.1)
xi = 1 for phase=0 and xi = −1 for phase=π.
In matrix form (3.1) is written as














y−N +1
y−N +2
...
y−1
y0
y1
...
yN −2
yN −1


x1
x2
...








 xN −1



 = 1  xN
 N
 0



 ...



 0

0
0
x1
...
0
0
...
xN −2
xN −1
xN
...
0
0
xN −3
xN −2
xN −1
...
0
0
45
...
...
...
...
...
...
...
...
...
0
0
...
x1
x2
x3
...
xN
0
0
0
...
0
x1
x2
...
xN −1
xN














xN
xN −1
xN −2
...
...
...
...
x2
x1







 (3.2)






Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
The vector in the right hand side of (3.2) is the replica of the transmitted signal. In
other words they are the weighing sequence for the received signal. So (3.2) can be
expressed as
y=
1
Xw1
N
(3.3)
where X is the matrix formed by shifting the input sequence {xi } and w1 is the
weight vector. It is observed from (3.2) that X has 2N − 1 number of patterns.
However, an additional null sequence {0} is considered for no input signal. So there
is 2N number of patterns that are used as the input to the pulse compression filter.
The desired output of the pulse compression filter for an input sequence is modeled
as a all zero vector except at one point at which the desired response is nonzero.
Thus the desired response is represented as
d = [0 0 0 . . . 1 0 0 0 . . .]T
(3.4)
where [.]T denotes the transpose operation. The nonzero component represents the
mainlobe.
The problem is to design a suitable network for the pulse compression using input
output pairs so as to get better performance in terms of PSR for range resolution,
detection in presence of noise and Doppler shift. Different networks such as the
adaptive linear combiner (ALC), MLP, RNN, RBF and RRBF are used as pulse
compression network described in next section. The adaptive network contains
connecting weights which are trained by various learning algorithms such as the LMS,
RLS and BP etc. The weights of the filter, which provide input output relationship,
are determined in an iterative manner.
3.3
Techniques used
In this section various models such as ALC, MLP, RNN, RBF, RRBF and their
learning algorithms are discussed. These models are used as mismatch filter for
radar pulse compression.
46
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
Figure 3.1: Adaptive linear combiner
3.3.1
Adaptive linear combiner
An ALC or nonrecursive adaptive filter is a computational device that attempts
to model the relationship between input and output in an iterative manner. The
general form of an ALC [42, 43] is depicted in Figure 3.1. The input signals to the
ALC are patterns and the nth pattern is represented as
x(n) = [x1 (n), x2 (n), .....xN (n)]T
(3.5)
where N represents the number of elements in each pattern. The ALC contains a
set of adjustable weights given by
w = [w1 , w2 , .....wN ]T
(3.6)
The estimated output of nth pattern is
y(n) = wT x(n)
(3.7)
The training algorithms for ALC are explained below.
(a)Least mean square algorithm
There are many algorithms found in literature to train various adaptive models.
The performance of these algorithms depend on the rate of convergence, training
47
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
time, computational complexity and residual mean square error (MSE). The LMS
algorithm is mostly used to train the weights of adaptive filters because of its low
computational complexity and ease of implementation. From Figure 3.1 the error
signal for nth pattern is obtained as
e(n) = d(n) − y(n)
(3.8)
where d(n) is the desired output for nth pattern.
The cost function to be optimized is
n
1
1X
e2 (n)
ξ=
2 n=1
(3.9)
where n1 is the number of patterns.
The weights associated with the filter are adjusted in such a way that the cost
function is minimized. The proposed study uses an epoch based adaptation for
weight updation. The ALC is trained with all n1 patterns and the change in weight
for each pattern is stored. These change in weights are used for a single update of
the filter weights which in turn constitutes an epoch. The new weights are used to
carry out the next epoch. The LMS algorithm uses the gradient descent technique
to minimize the cost function and the weights are updated [43] as
n1
X
∂ξ
e(n)x(n)
= wk (m) + µ
wk (m + 1) = wk (m) − µ
∂wk (m)
n=1
(3.10)
where k = 1, 2....N and m is the epoch index.
(b)Recursive least square algorithm
The algorithm such as LMS is derived by using some approximation made in the
estimate of the performance function gradient. This type of algorithm have the
disadvantages that they are slow to obtain the optimum weight vector and once close
to it, usually “rattle around” the optimal vector rather than actually converging to
it. To overcome this difficulty, another efficient approach known as RLS algorithm
has been discussed in this section. The advantage gained by the use of the RLS
48
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
algorithm is at the expense of an increase in computational complexity.
Steps involved in RLS algorithm
(a) Accept new pattern x(n) and corresponding desired output d(n).
(b) Compute the estimated output y(n) as expressed in (3.7). Initially assume the
weight vector as zero.
(c) Compute the error e(n) as given in (3.8).
(d) Compute filtered information vector z(n):
z(n) = R−1 (m)x(n)
(3.11)
where R is the autocorrelation matrix of input pattern. The R−1 (m) is assumed
to be exist, where m is epoch index. Initially R−1 (m) is taken as ηI, where I is
an identity matrix of size N × N . The value of η taken as very large i.e. about
104 .
(e) Compute normalized error power q:
q = xT (n)z(n)
(3.12)
(f) Compute gain constant v:
v=
1
1+q
(3.13)
(g) Compute the normalized information vector ẑ(n):
ẑ = vz(n)
(3.14)
(h) Compute the change in weight vector as
∆w(n) = e(n)ẑ
(3.15)
(i) Compute the change in inverse correlation matrix
∆R−1 (n) = −ẑ(n)zT (n)
49
(3.16)
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
The steps from (a) to (i) are carried out for all n1 patterns. The weights are updated
as
w(m + 1) = w(m) +
n1
X
∆w(n)
(3.17)
n=1
and the inverse correlation matrix is updated as
−1
−1
R (m + 1) = R (m) +
n1
X
∆R−1 (n)
(3.18)
n=1
The updated values of the weights and the inverse correlation matrix are used to
carry out next epoch.
(c)Modified RLS algorithm
The modified RLS algorithm is derived by using the condition
|e(n)| ≥ Th
(3.19)
where Th represents a threshold value. The instantaneous error is compared with
the threshold value. If the instantaneous error is greater than the threshold value
then steps from (d) to (i) of the RLS algorithm are evaluated otherwise not. If
|e(n)| < Th , then values of ∆R−1 (n) and ∆w(n) are zero. The weights and the
inverse correlation matrix are updated as given in (3.17) and (3.18) respectively.
Initially the threshold value is chosen as very small and later it is updated for
each epoch based on the maximum error value at that epoch. The updation of
threshold at mth epoch is given by
maxerrm = maximum|em (n)|
(3.20)
Thm = δ ∗ maxerrm
(3.21)
where em (n) is the error vector at mth epoch and maxerrm is the maximum value of
all the errors in error vector. δ, a constant whose value is close or equal to 1, affects
the rate of convergence.
50
Chapter 3
3.3.2
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
Artificial neural network
An artificial neural network (ANN) is an information processing paradigm that
is inspired by the way biological nervous system, such as the brain, process
the information [44, 45].
The first artificial neuron was developed in 1943 by
the neurophysiologist Warren McCulloch and the logician Walter Pits. But the
technology available at that time did not allow them to proceed further.
In
past few decades the ANN has emerged as a powerful learning tool to perform
complex tasks in highly nonlinear dynamic environment. The ANN is capable
of performing nonlinear mapping between the input and output space due to its
large parallel interconnection between different layers and the nonlinear processing
characteristic. Therefore, the ANN is used extensively in the field of communication,
control, instrumentation and forecasting [46–48]. ANN technique is also used for
classification, modeling and optimization problems [49, 50].
An artificial neuron basically consists of a computing element that performs the
weighted sum of the input signal and the connecting weight. The sum is added with
the bias or threshold and the resultant signal is then passed through a nonlinear
function of sigmoid or hyperbolic tangent type. Each neuron is associated with
three parameters whose learning can be adjusted. These are the connecting weights,
the bias and the slope of the nonlinear function. For the structural point of view a
neural network (NN) may be single layer or it may be multilayer. In MLP there is a
number of layers and each layer contains one or many artificial neurons. Each neuron
of the one layer is connected to each and every neuron of the next layer. A trained
neural network can be thought of as an “expert” in the category of information it
has been given to analyze. The advantages of ANN are
(a) Adaptive learning: It is the ability of the network to learn how to do tasks
based on the data given for training or initial experience.
(b) Self-organization: An ANN can create its own organization or representation
of the information as it receives during learning time.
51
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
(c) Real time operation: The ANN computations may be carried out in parallel,
and special hardware devices are being designed and manufactured which take
advantage of this capability.
(d) Fault tolerance via redundant information coding: Partial destruction
of a network leads to the corresponding degradation in performance. However,
some network capabilities may be retained even with major network damage.
Single neuron structure
A neuron is an information processing unit for the operation of a neural network.
The operation in a single neuron involves the computation of the weighted sum
of inputs and threshold. The resultant signal is then passed through a nonlinear
activation function. The basic structure of a single neuron is shown in Figure 3.2.
The output associated with the neuron is computed as
Figure 3.2: Single neuron structure
y=f
"
N
X
wi x i + b
i=1
#
(3.22)
where xi , i = 1, 2...N , are inputs to the neuron, wi is the synaptic weights of the
ith input, b is the bias and f is the nonlinear activation function. The activation
functions generally used in neural computation are discussed below.
Activation functions
The activation or transfer function may be a linear or a nonlinear in nature. A
particular transfer function is chosen to satisfy some specification of the problem that
the neuron is attempting to solve. Some of the activation functions are explained
below.
52
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
1. Log-sigmoid function
This transfer function takes the input and squashes the output into the range
of 0 to 1, according to expression given below
f (x) =
1
1 + e−x
(3.23)
2. Hyperbolic tangent Sigmoid:
This function is expressed as
f (x) = tanh(x) =
ex − e−x
ex + e−x
(3.24)
3. Signum Function:
The expression for this activation function



1
if



f (x) =
0
if




−1 if
is given by
x>0
x=0
(3.25)
x<0
4. Threshold function
This function is given by the expression


 1
if x ≥ 0
f (x) =

 0
if x < 0
(3.26)
This function is represented as



1



f (x) =
x




 0
(3.27)
5. Piecewise linear function
if
x > 0.5
if
− 0.5 ≤ x ≤ 0.5
if
53
x < 0.5
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
The different structures of neural network and their learning algorithms are described
below.
(i) Multilayer perceptron
The MLP is a feed forward network having an input layer, one or more hidden layer
and an output layer. The layer to which the input data is given called as input layer
and the layer from which output is taken called as output layer. All the intermediate
layers are called as hidden layers. The layers are fully interconnected i.e. each neuron
is connected to every neuron in previous and succeeding layers. The input signal
propagates through the network on a layer by layer basis. This network have been
applied successfully to solve many nonlinear and complex problems in several fields.
The structure of a three layer MLP is shown in Figure 3.3 which consists of input
layer, one hidden layer and output layer. i, j and k are the indices used for input,
hidden and output layer respectively. x(n) = [x1 (n), x2 (n), ..., xN (n)]T is the input
to the network for nth pattern and wji is the synaptic weight connecting input xi (n)
to the hidden neuron j. Similarly wkj is the synaptic weight connecting output of j th
hidden neuron output to the k th neuron of output layer. bj and bk are the biases to
the hidden layer and output layer respectively. f represents the nonlinear activation
function for both hidden and output layer. The activation functions can be different
for different layers. The output of j th hidden neuron for nth pattern is
yj (n) = f (aj (n))
where
aj (n) =
N
X
wji xi (n) + bj
i=1
The response of the k
th
(3.28)
!
(3.29)
output node is
yk (n) = f (ak (n))
where
ak (n) =
n2
X
wkj yj (n) + bk
j=1
54
(3.30)
!
(3.31)
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
Figure 3.3: Mlutilayer perceptron network
and n2 is the number of hidden nodes.
Back propagation algorithm
This algorithm is used to train the parameters of the MLP to get optimum cost
function. Basically BP learning comprises of two passes through the different layers
of networks: a forward pass and a backward pass. In forward pass a pattern (input
vector) is applied to the input nodes and its effect propagates through the network
layer by layer. In forward pass the synaptic weights remain constant. On the
other hand during backward pass the synaptic weights and the biases are adjusted
according to the error correction rule. The parameters of the neural network are
updated by BP on epoch basis.
The error of the k th neuron output for nth pattern is
ek (n) = dk (n) − yk (n)
(3.32)
where dk (n) and yk (n) are desired and estimated outputs respectively. The cost
55
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
Chapter 3
function which is to be minimized is defined as
n
n
1
3 X
1X
ek (n)2
ξ=
2 k=1 n=1
(3.33)
where n3 is the number of neurons in the output layer.
According to the gradient descent, each weight change in the network should be
proportional to the negative gradient of the cost function with respect to the specific
weights.
The local gradient for k th output neuron is
δk (n) = −
∂ξ ∂ek (n) ∂yk (n)
′
= (dk (n) − yk (n)) f (ak (n))
∂ek (n) ∂yk (n) ∂ak (n)
(3.34)
′
where f (ak (n)) is the first derivative of f (ak (n)) with respect to ak (n).
The local gradient for j th the hidden nodes is
δj (n) = −
n3
X
k=1
n
3
∂ξ ∂ek (n) ∂yk (n) ∂ak (n) ∂yj (n) X
′
δk (n)wkj (m)f (aj (n))
=
∂ek (n) ∂yk (n) ∂ak (n) ∂yj (n) ∂aj (n) k=1
(3.35)
where m is the epoch index. The local gradients for all patterns are calculated in
each epoch. The change in weights for output layer in mth epoch is
∆wkj (m) = η1
n1
X
δk (n)yj (n)
(3.36)
n=1
where η1 is the learning parameter. The change in biases for output layer is
∆bk (m) = η1
n1
X
δk (n)
(3.37)
n=1
The output layer weights and biases are updated as
wkj (m + 1) = wkj (m) + ∆wkj (m)
(3.38)
bk (m + 1) = bk (m) + ∆bk (m)
(3.39)
The change in weights for hidden layer is
∆wji (m) = η1
n1
X
n=1
56
δj (n)xi (n)
(3.40)
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
The change in biases for hidden layer is
∆bj (m) = η1
n1
X
δj (n)
(3.41)
n=1
The hidden layer weights and biases are updated as
wji (m + 1) = wji (m) + ∆wji (m)
(3.42)
bj (m + 1) = bj (m) + ∆bj (m)
(3.43)
(ii) Recurrent neural network
Recently, significant research work has been carried out to demonstrate the
effectiveness of recurrent neural network in modeling of nonlinear dynamic systems
[51–56]. The RNN has many advantages over static layered networks when used for
system identification and feed back controller [51,52]. Moreover, RNN is capable for
long range prediction in the presence of measurement noise and also able to filter the
noise from the inputs [57]. The RNNs are used to model the plant nonlinearities in
more efficient ways as compared to feed forward network. The RNN has at least one
feed back loop in its architecture which is not present in feed forward network. Thus
in the RNN, there may be one layer with feed back connections as well as there may
be neurons with self feed back where output of the neuron is fed back into itself as
the input. The presence of feed back loop affects heavily on the learning capability of
the network. Contrary to the MLP, the RNN is sensitive and adaptive to past inputs.
Among the several ANN architectures available in the literature, ANN having feed
back and internal dynamics have been considered more suitable for modeling and
control of the nonlinear systems as compared to feed forward network [58].
A block diagram of the RNN is shown in Figure 3.4. The RNN has structure as
that of MLP with feedback or recurrent connections. This network has recurrent
connections from the hidden neurons to a layer of context units consisting of bank
of unit delays [59]. These context units store the outputs of hidden neurons for one
time step and feed them back to the input layer. The inputs to the hidden layers
are combination of the present inputs and the outputs of the hidden layer which
57
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
Chapter 3
Figure 3.4: Block diagram of recurrent neural nerwork
are stored from previous time step in context layer. Hence the outputs of the RNN
are functions of present state, previous state (that is stored in context units) and
present inputs.
In this case the the j th hidden layer outputs are calculated as
yj (n) = f (aj (n))
where
aj (n) = f
N
X
wji xi (n) +
i=1
n2
X
h=1
(3.44)
ujh yj (n − 1) + bj
!
(3.45)
where n2 is the number of hidden nodes and ujh are the recurrent layer weights. The
response of the k th neuron in output layer is
yk (n) = f (ak (n))
where
ak (n) =
n2
X
wkj yj (n) + bk
j=1
(3.46)
!
(3.47)
Weight updation for recurrent neural network
The hidden node output is used to compute the response of output layer of the RNN
as given in (3.46). The local gradient and weight update procedure are same as that
of MLP. The change in recurrent layer weights are obtained as
n1
X
δj (n)yj (n − 1)
(3.48)
uhj (m + 1) = uhj (m) + ∆uhj (m)
(3.49)
∆uhj (m) = η1
n=1
Recurrent layer weights are updated as
58
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
Hence all weights are updated based on the corresponding weight correction
equations.
(iii) Radial basis function network
The radial basis function network can be viewed as a feed forward neural network
with a single hidden layer which computes the distance between input pattern and
the center [60]. It consists of three layers, an input layer, a hidden layer and an
output layer. The input layer connects the network to the environment. The
second layer is the only hidden layer which transfer the input space nonlinearly
using radial basis function. The hidden space is greater than the input space in
most of the applications. The response of the network provided by the output
layer which is linear in nature. The RBF network is suitable for solving function
approximation, system identification and pattern classification because of its simple
topological structure and their ability to learn in an explicit manner [61, 62]. The
Figure 3.5: Architecture of radial basis function network
basic architecture of RBF network is shown in Figure 3.5. Here x(n) is the input to
the network and φ represents the radial basis function that perform the nonlinear
mapping and M represents the total number of hidden units. Each node has a center
vector ck and spread parameter σk , where k = 1, 2, ....M .
Radial basis functions
The radial basis functions are represented by φ(kx, ck), where k.k represents the
Euclidean norm. The radial basis functions which are generally used in various
59
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
applications are
1. Multiquadratics
1
φ(r) = (r2 + c2 ) 2
f or c > 0, r ∈ R
(3.50)
f or c > 0, r ∈ R
(3.51)
f or σ > 0, r ∈ R
(3.52)
wk (m)φ (x(n), ck (m), σk (m)))
(3.53)
2. Inverse multiquadratics
1
φ(r) = (r2 + c2 )− 2
3. Gaussian function
φ(r) = exp(−
r2
)
2σ 2
Learning algorithm for RBF network
The error for the nth pattern is obtained as
e(n) = d(n) −
M
X
k=1
where d(n) is the desired output. If the Gaussian function chosen as the radial basis
function
M
X
kx(n) − ck (m)k2
e(n) = d(n) −
wk (m)exp −
σk2 (m)
k=1
!
(3.54)
The cost function is defined as
n
1
1X
ξ=
e2 (n)
2 n=1
(3.55)
where n1 is the number of training patterns. It is required to adjust the free
parameters such as weight, center and spread so as to minimize ξ. According to
the gradient descent algorithm the free parameters for mth epoch are updated as
∂ξ
∂wk (m)
(3.56)
∂ξ
∂ck (m)
(3.57)
wk (m + 1) = wk (m) − µw
ck (m + 1) = ck (m) − µc
60
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
σk (m + 1) = σk (m) − µσ
∂ξ
∂σk (m)
(3.58)
where µw , µc and µσ are learning parameters and k = 1, 2...M . Finally the updation
equations are defined as
wk (m + 1) = wk (m) +
n1
X
µw e(n)φ (x(n), ck (m), σk (m))
(3.59)
n=1
ck (m + 1) = ck (m) +
n1
X
µc
n=1
σk (m + 1) = σk (m) +
n1
X
n=1
where
e(n)wk (m)
φ (x(n), ck (m), σk (m)) [x(n) − ck (m)] (3.60)
σk2 (n)
µσ
e(n)wk (m)
φ (x(n), ck (m), σk (m)) [kx(n) − ck (m)k2 ]
σk3 (m)
(3.61)
kx(n) − ck (m)k2
φ (x(n), ck (m), σk (m)) = exp −
σk2 (m)
!
(3.62)
(iv) Recurrent radial basis function
The RRBF [63] combines the advantages of RBF and dynamic representation of
time. The RRBF network has been applied for modeling [64], noise cancellation
[65, 66] and time series [67] prediction. This network has faster convergence [68]
while maintaining the modeling capability of neural networks. The architecture of
Figure 3.6: Architecture of recurrent radial basis function network
RRBF model is shown in Figure 3.6. The model of RRBF is similar to RBF with
an input layer, one hidden layer and an output layer. In this network each output of
61
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
Chapter 3
the hidden neurons are fed back to their corresponding input through a delay. The
estimated output of the network for nth pattern is
y(n) =
M
X
wk (m)φ (x(n), ck (m), σk (m))
(3.63)
k=1
where
φ (x(n), ck (m), σk (m)) =
!
kx(n) − ck (m)k2
+ gk (m)φ (x(n − 1), ck (m), σk (m))
(3.64)
exp −
σk2 (m)
Learning algorithm for RRBF
In this case the cost function is same as that of RBF as defined in (3.55). wk , ck and
σk are updated as that of RBF using the currently defined φ (x(n), ck (m), σk (m)).
The recurrent weights are updated as
gk (m + 1) = gk (m) − µg
∂ξ
∂gk (m)
(3.65)
where µg is the learning parameter.
n
1
X
∂ξ
wk (m)φ (x(n), ck (m), σk (m)) φ (x(n − 1), ck (m), σk (m))
=
∂gk (m) n=1
(3.66)
From (3.66) and (3.67)
gk (m + 1) =
gk (m) − µg
n1
X
n=1
wk (m)φ (x(n), ck (m), σk (m)) φ (x(n − 1), ck (m), σk (m)) (3.67)
where k = 1, 2...M .
3.4
Simulation results
This section illustrates the performance of various networks for radar pulse
compression. First, the performance of ALC trained with LMS, RLS and Modified
62
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
RLS algorithm is presented. Subsequently the performances of MLP, RBF along
with the proposed RNN and RRBF are presented. All the networks are trained
with time shifted sequences of the 13-bit and 35-bit Barker codes. The time shifted
sequence for 13-bit Barker code is presented in Figure 3.7. A null sequence {0} is
added to the shifted sequence that represents radar has not received any information.
So there are 26 patterns for 13-bit Barker code. Similarly the number of patterns
in case of 35-bit Barker code is 70. In these training sequences the desired output
of the network is 1 when the proper Barker code present in the input, otherwise the
output is zero.
Figure 3.7: 26 different possible input sequences for 13-bit Barker codes
63
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
Table 3.1: PSRs obtained using various learning algorithms.
Algorithms
3.4.1
13-bit Barker code 35-bit Barker code
(PSR in dB)
(PSR in dB)
ACF
-22.27
-13.97
LMS
-23.86
-16.6
RLS
-24
-16.65
modified RLS
-25.52
-18
Sidelobe suppression using adaptive linear combiner
The adaptive linear combiner is uniquely defined by its weight coefficients. The
length of the weight vector is taken to be same as the input sequence and weights
are initialized to zero. 13-bit and 35-bit Barker codes are used as inputs to the filter.
Several algorithms like LMS, RLS and modified RLS with suitable parameter values
are used for updating the weights of the linear combiner to minimize the error. For
LMS algorithm the convergence parameter µ is chosen as 0.01 and for RLS the value
of η is chosen as 104 . Similarly for modified RLS, Th = 0 and δ = 0.995 are used. The
network is trained for 500 epochs for each algorithm. Once the training is over, the
network can be used as pulse compression filter. Output of pulse compression filter
for different algorithm for 13-bit and 35-bit Barker codes are depicted in Figures 3.8
and 3.9 respectively. The PSRs for ACF, LMS, RLS and modified RLS using 13-bit
and 35-bit Barker codes as input are given in Table 3.1. From the table it is clear
that the modified RLS algorithm gives the highest PSR of magnitude 25.52 dB for
13-bit Barker code and 18 dB for 35-bit Barker code. But these low magnitude of
PSR values are not suitable for many radar applications.
3.4.2
Sidelobe suppression using MLP, RNN, RBF, RRBF
The MLP and RNN consist of input layer one hidden layer and output layer. The
log-sigmoid function is used as the activation function in hidden and output layers.
64
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
Chapter 3
0
Filter response in dB
−5
−10
−15
−20
−25
−30
−35
0
5
10
15
20
25
15
20
25
15
20
25
15
20
25
Time delay
(a)
0
Filter response in dB
−5
−10
−15
−20
−25
−30
−35
0
5
10
Time delay
(b)
0
Filter response in dB
−5
−10
−15
−20
−25
−30
−35
0
5
10
Time delay
(c)
0
Filter response in dB
−5
−10
−15
−20
−25
−30
−35
0
5
10
Time delay
(d)
Figure 3.8: Filter response in dB for 13-bit Barker code obtained using (a)ACF
(b)LMS (c)RLS (d)Modified RLS algorithms
65
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
Chapter 3
0
Filter response in dB
−5
−10
−15
−20
−25
−30
−35
0
10
20
30
40
50
60
70
40
50
60
70
40
50
60
70
50
60
70
Time delay
(a)
0
Filter response in dB
−5
−10
−15
−20
−25
−30
−35
0
10
20
30
Time delay
(b)
0
Filter response in dB
−5
−10
−15
−20
−25
−30
−35
0
10
20
30
Filter response in dB
(c)
0
Filter response in dB
−5
−10
−15
−20
−25
−30
−35
0
10
20
30
40
Time delay
(d)
Figure 3.9: Filter response in dB for 35-bit Barker code obtained using (a)ACF
(b)LMS (c)RLS (d)Modified RLS
66
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
The number of input neurons are same as the length of the input code i.e. 13 for
13-bit Barker code and 35 for 35-bit Barker code. The number of hidden layer and
output layer neurons are chosen as three and one respectively. The weights and the
biases are randomly initialized. The learning parameter η1 is chosen as 0.8. The RBF
and RRBF consist of seven hidden neurons having Gaussian radial basis function
and one output neuron is used. Weight(w), centre(c) and spread (σ) parameters are
randomly initialized. The values of learning parameters µw , µc and µσ for RBF are
chosen as 0.75, 0.8 and 0.75 respectively. Similarly the values of learning parameters
µw , µc , µσ and µg for RRBF are chosen as 0.8, 0.8, 0.75 and 0.8 respectively. All the
four networks are trained for 500 epochs according to their learning algorithm given
in Section 3.3. After completion of the training, the neural network can be used for
pulse radar detection by using various set of input sequences.
Convergence performance
The MSE of all the networks for 13-bit and 35-bit Barker codes are depicted in
Figure 3.10. From the figure it is evident that the RRBF based approach offers
better convergence speed and very low residual error after training for 13-bit and
35-bit Barker codes as compared to all other networks.
PSR performance
After the training is over, different inputs are applied to the networks to examine
PSR performance. The compressed output of different networks for 13-bit Barker
code is shown in Figure 3.11. The PSR values of all the networks for 13-bit and
35-bit Barker codes are listed in Table 3.2. The table shows that the proposed
RRBF network have achieved highest PSR magnitude for both 13-bit and 35-bit
Barker codes compared to all other approaches.
Noise performance
Noise is a random signal which interferes with the target echoes. If the noise is
very high it may mask the target echo. So it is also required to examine the noise
67
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
Chapter 3
−1
10
−2
MLP
RNN
RBF
RRBF
Mean square error(log10)
10
−3
10
−4
10
−5
10
−6
10
0
50
100
150
200
250
300
350
400
450
500
Epochs
(a)
−1
10
−2
MLP
RNN
RBF
RRBF
Mean square error(log10)
10
−3
10
−4
10
−5
10
−6
10
0
50
100
150
200
250
300
350
400
450
500
Epochs
(b)
Figure 3.10: Convergence graphs of different structures for (a)13-bit (b)35-bits
Barker codes
Table 3.2: PSRs obtained by various structures
Structures 13-Bit Barker Code 35-Bit Barker Code
(PSR in dB)
(PSR in dB)
MLP
-42.61
-40.87
RNN
-45.75
-44.93
RBF
-60.43
-56.42
RRBF
-64.31
-62.35
68
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
Chapter 3
Filter response
1.2
1
0.8
0.6
0.4
0.2
0
0
5
10
15
20
25
30
20
25
30
20
25
30
20
25
30
Time delay
(a)
Filter response
1.2
1
0.8
0.6
0.4
0.2
0
0
5
10
15
Time delay
(b)
1.2
Filter response
1
0.8
0.6
0.4
0.2
0
−0.2
0
5
10
15
Time delay
(c)
1.2
Filter response
1
0.8
0.6
0.4
0.2
0
−0.2
0
5
10
15
Time delay
(d)
Figure 3.11: Compressed waveforms for 13 bit Barker code using (a)MLP (b)RNN
(c)RBF (d)RRBF structures
69
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
rejection ability of different networks. The inputs having different SNR ranging
from 0 dB to 20 dB are applied to the networks and the output PSR for 13-bit and
35-bit Barker codes are listed in Tables 3.3 and 3.4 respectively. These tables show
that as the SNR increases the magnitude of PSR also increases. The RRBF provides
highest magnitude of PSR in all SNR values compared to those obtained by all other
approaches.
Table 3.3: Comparison of PSRs in dB at different SNRs for 13-bit Barker code
Structures SNR=0dB SNR=5dB SNR=10dB SNR=15dB SNR=20dB
MLP
-14.23
-28.61
-36.71
-38.53
-39.82
RNN
-17.11
-32.17
-38.35
-40.59
-41.76
RBF
-35.28
-45.23
-50.33
-55.77
-57.62
RRBF
-40.24
-49.27
-57.30
-60.12
-61.24
Table 3.4: Comparison of PSRs in dB at different SNRs for 35-bit Barker code
Structures SNR=0dB SNR=5dB SNR=10dB SNR=15dB SNR=20dB
MLP
-15.18
-29.17
-32.43
-36.95
-38.12
RNN
-19.52
-32.74
-37.83
-40.87
-42.65
RBF
-40.25
-48.25
-52.78
-54.44
-55.17
RRBF
-42.25
-54.69
-57.47
-58.60
-60.57
Range resolution ability
The range resolution is to analyze the ability of a particular network to distinguish
between two targets by measurement of their ranges in the radar system. The two
targets which are to be resolved must be separated by at least the range equivalent
of the width of the processed echo. To compare the range resolution ability two
overlapping codes of same length are considered with n-delay apart (DA) having
70
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
Table 3.5: Comparison of range resolution ability for 13-bit Barker code of two
targets having same IMR and DA.
Structures
MLP
2-DA
3-DA
4-DA
5-DA
(PSR in dB) (PSR in dB) (PSR in dB) (PSR in dB)
-36.53
-38.52
-37.32
-36.16
RNN
-40.24
-41.23
-39.23
-38.78
RBF
-53.32
-55.25
-56.76
-54.23
RRBF
-59.72
-58.28
-60.73
-58.71
Table 3.6: Comparison of range resolution ability for 35-bit Barker code of two
targets having same IMR and DA.
Structures
MLP
2-DA
3-DA
4-DA
5-DA
(PSR in dB) (PSR in dB) (PSR in dB) (PSR in dB)
-34.41
-34.83
-33.75
-32.62
RNN
-38.23
-37.79
-36.87
-35.25
RBF
-48.34
-47.61
-49.82
-47.13
RRBF
-53.72
-55.14
-54.25
-53.23
same or different input magnitude ratio (IMR). The IMR is defined as the magnitude
of first pulse train over that of the delayed pulse train. Figure 3.12 shows the added
input waveform of equal magnitude (IMR=1) with 5 delay apart for 13-bit Barker
code. The compressed output for this input for all the network are shown in Figure
3.13. In this case the PSR is calculated by taking lower value of the two mainlobes.
By varying the DA from 2 to 5 the PSR for 13-bit and 35-bit Barker codes are
obtained and shown in Tables 3.5 and 3.6 respectively. In Tables 3.7 and 3.8 the
PSR for different IMRs and DAs for all the networks are listed. From these tables
it is evident that the PSR values for RRBF are the best among those offered by
all other networks i.e. RRBF based pulse compression technique have best range
resolution ability compared to those of other networks.
71
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
1
0.8
0.6
0.4
Input
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
0
2
4
6
8
10
12
14
16
18
20
12
14
16
18
20
12
14
16
18
20
Time delay
(a)
1
0.8
0.6
0.4
Input
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
0
2
4
6
8
10
Time delay
(b)
2
1.5
Added input
1
0.5
0
−0.5
−1
−1.5
−2
0
2
4
6
8
10
Time delay
(c)
2
1.5
1
Input
0.5
0
−0.5
−1
−1.5
−2
0
2
4
6
8
10
Time delay
12
14
16
18
20
(d)
Figure 3.12: Input waveform on addition of two 5-DA 13-bit Barker sequence having
same magnitude (a)Left shift (b)Right shift (c)Added waveform (d)Waveform after
flip about the vertical axis
72
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
Chapter 3
Filter response
1.2
1
0.8
0.6
0.4
0.2
0
0
5
10
15
20
25
30
20
25
30
20
25
30
20
25
30
Time delay
(a)
Filter response
1.2
1
0.8
0.6
0.4
0.2
0
0
5
10
15
Time delay
(b)
1.2
Filter response
1
0.8
0.6
0.4
0.2
0
−0.2
0
5
10
15
Time delay
(c)
1.2
Filter response
1
0.8
0.6
0.4
0.2
0
−0.2
0
5
10
15
Time delay
(d)
Figure 3.13: Compressed waveforms for 13-bit Barker code having same IMR and 5
DA for (a)MLP (b)RNN (c) RBF (d)RRBF structures
73
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
Chapter 3
Table 3.7: Comparison of range resolution ability for 13-bit Barker code of two
targets having different IMR and DA.
Structures
MLP
2-DA
3-DA
4-DA
5-DA
2-IMR
3-IMR
4-IMR
5-IMR
(PSR in dB) (PSR in dB) (PSR in dB) (PSR in dB)
-38.17
-30.23
-24.16
-12.14
RNN
-40.23
-36.56
-31.14
-23.65
RBF
-51.38
-49.42
-43.24
-33.18
RRBF
-56.36
-55.42
-50.24
-39.37
Table 3.8: Comparison of 35-bit Barker code for range resolution ability of two
targets having same IMR and DA
Algorithms
MLP
2-DA
3-DA
4-DA
5-DA
2-IMR
3-IMR
4-IMR
5-IMR
(PSR in dB) (PSR in dB) (PSR in dB) (PSR in dB)
-34.46
-27.75
-21.78
-14.54
RNN
-39.44
-33.23
-26.74
-20.68
RBF
-47.77
-45.24
-38.21
-25.23
RRBF
-52.64
-48.71
-43.41
-35.42
Table 3.9: Doppler shift performance
Structures 13-bit Barker code 35-bit Barker code
(PSR in dB)
(PSR in dB)
MLP
-14.35
-28.34
RNN
-30.93
-42.36
RBF
-47.45
-46.42
RRBF
-55.23
-56.34
74
Chapter 3
Development and Performance Evaluation of New and Efficient
ANN Mismatch Filters for Sidelobe Reduction
Doppler shift performance
The influence of Doppler shift should be accounted for evaluating the detection
performance for a moving target. The Doppler tolerance measures the Doppler
sensitivity of the pulse compression technique. The Doppler sensitivity is caused by
the shifting in phase of the individual elements of the code by the target Doppler.
In extreme case the phase shift across the code will be 180o , the last subpulse in
the received code is effectively inverted. For 13-bit Barker code at extreme case the
input will change from “1 1 1 1 1 -1 -1 1 1 -1 1 -1 1” to “-1 1 1 1 1 -1 -1 1 1 -1 1 -1
1”. For 13-bit and 35-bit Barker codes the extreme case Doppler shift PSR values
for different types of network are listed in Table 3.9. From this table it is observed
that the MLP has very low Doppler tolerance and RRBF produces the best PSR
value of -55.23 dB for 13-bit Barker code.
3.5
Conclusion
In this chapter recurrent networks such as RNN and RRBF are proposed for radar
pulse compression. The simulation results reveal that the performance of RRBF
based pulse compression is much better than MLP, RNN and RBF based pulse
compression techniques. The convergence rate of RRBF is higher than that of all
other networks and it has low training error. The RRBF approach provides better
PSR values in different adverse conditions such as noise and Doppler shift conditions.
The range resolution ability of RRBF network is much superior than MLP, RNN and
RBF networks. Although the algorithms are applied for 13-bit and 35-bit Barker
codes, they can also be used for any other biphase codes.
75
Chapter 4
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using
Convolutional Windows
4.1
Introduction
The pulse compression techniques dealt in Chapter 3 are meant for biphase codes.
These codes are easily generated and the correlators for these codes are very simple.
But the compressed output of biphase codes are associated with the high time range
sidelobes and these codes are more prone to Doppler shift. The application of a
pulse compression technique depends on how efficiently it reduces the range sidelobes
associated with the compressed waveforms. The number of Barker codes available
are very less. Hence, these codes seriously suffer from security problem. Apart
from biphase codes, polyphase codes and frequency modulated codes are also used
in radar systems. PSL of frequency modulated pulse and polyphase codes are lower
than that of the biphase codes. The frequency modulated and polyphase codes are
more Doppler tolerant and have less range sidelobes compared to biphase codes.
Different windows those are available in the literature are used for reducing the
range sidelobes of the compressed output in case of LFM and polyphase codes.
In most of the practical radar systems LFM waveform is extensively used because
it is more Doppler tolerant than phase coded signals. The matched filter output of
a point target for an arbitrary pulse is the ACF which forms a Fourier transform
76
Chapter 4
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
pair with the energy spectrum of the signal. For rectangular amplitude weighing,
the energy spectrum of an LFM signal can be approximated as sin(x)/x or sinc(x)
shaped ACF. Hence a compressed LFM signal at the receiver will produce a series
of sidelobes surrounding the mainlobe and the first sidelobe occurs at a level of 13
dB below the peak of the mainlobe. Range sidelobes are inherent part of the pulse
compression mechanism and they occur due to abrupt rise in the signal spectrum.
The conventional method used to suppress these ambiguous sidelobes by modifying
the rectangular shape of the chirp spectrum using amplitude weighing. In radar
systems, weighing techniques in time or in the frequency domain are mostly employed
to reduce these range sidelobes with broadening in the mainlobe. Time domain
weighing is preferred to its frequency domain counterpart, as it produces lower peak
sidelobe in compressed output [69–71].
Although weighing when used both on transmitter and receiver provides better
results, weighing only on receiver is preferred. Weighing on transmitter leads to
power loss hence the available transmit power cannot be fully utilized. In low T B
product LFM waveforms the Fresnel ripples, which are responsible for producing
range sidelobes, are reduced by modifying the chirp waveform before transmission.
Amplitude tapering [69] and phase distortion [69, 72] are used to modify the chirp
waveform to suppress the peak sidelobe as well as to increase the fall off rate of far
sidelobes. Amplitude tapering reduces the far sidelobes effectively. But in most high
power radars, the control of the pulse rise time is very difficult. So an appropriate
phase distortion function is used in the LFM pulse for short rise time and high
power radars. Shennawy et. al. [73] have used an external Hamming window as
weighing function in frequency domain to suppress the range sidelobes from a T B
product of 50 to 720. Using the weighing technique the dynamic range of the pulse
compression system is increased. Hamming weighing has been used to suppress the
range sidelobes for rectangular LFM pulses with T B product less than 170 [74] and
it is observed from the results that Hamming weighing in time domain produces
lower peak sidelobe as compared to Hamming weighing in frequency domain. If
77
Chapter 4
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
weighing at the receiver is used, the reflected waveform from the object no longer
matched to the receiver filter which causes mismatch loss. In place of LFM, nonlinear
LFM (NLFM) can be used for transmission which does not need any weighing at
the receiver [75–77] to overcome the mismatch loss. If the objects are in motion
in the environment then the waveforms reflected are Doppler shifted version of the
transmitted signal and the matched filter output of these Doppler shifted signal
produces very low PSR magnitude. If the T B product of the transmitted signal
is low then the PSR magnitude becomes even worse. The NLFM signals are more
affected by Doppler shift and are difficult to design than the LFM signals.
Although LFM signal is popularly used in radar, the group of phase coded
pulses is also an active research area for particular radar applications.
Phase
coded waveforms are more compatible for digital generation and compression [78].
However, these waveforms are affected more in the presence of Doppler shift as
compared to the LFM signal. To get the Doppler shift advantage of LFM signal
various polyphase codes are derived from LFM signal [79–81]. The codes such as
Frank [79], P1 and P2 [80] are derived from step approximation to LFM waveform.
These codes provide lower peak sidelobes than that offered by the best biphase
codes [10] for a particular length. In [81] two more polyphase codes, P3 and P4 are
discussed which are derived from the LFM signals. These codes are more Doppler
tolerant as compared to P1 and P2 codes. Although polyphase codes have lower
sidelobes in their ACF, it is required to further reduce the sidelobes for many
radar applications. Many sidelobe reduction techniques for polyphase codes such as
amplitude weighing [82,83] and the post compression sliding window techniques [84]
are found in the literature . When Doppler shift is zero these techniques substantially
reduce the sidelobes of the compressed pulse. Due to Doppler shift, objects with
larger velocities experience detection range degradation. Grating lobes are appeared
in the ACF of Frank and P1 code with increasing in Doppler shift [85].
Lee
and Griffiths [86, 87] have proposed Woo filter for polyphase codes which provides
optimum uniform sidelobe level and excellent Doppler shift performance. A modified
78
Chapter 4
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
P4 code which have better sidelobe and Doppler shift characteristics is presented
in [88]. But the mainlobe of the compressed output for this code is broadened
by a factor of two. To overcome the problem of mainlobe broadening, Lee [89] has
proposed asymmetric weighing at the receiver. Two variants of Woo filter is proposed
in [90] for P4 codes which have better resistance for Doppler shift as compared to
Woo Filter.
In this chapter convolutional windows are proposed to use as the weighing
functions at the receiver filter to reduce the sidelobes of LFM and polyphase codes.
To assess the performance of convolutional windows under Doppler shift condition,
exhaustive simulation studies are carried out under different Doppler shift conditions
and the results are compared with the conventional windows.
4.2
LFM and polyphase codes
The LFM and polyphase codes that are used in the radar system are explained
below.
4.2.1
LFM signal
An LFM pulse having rectangular envelope mathematically described as
B 2
|t| ≤ Tp /2
t
s (t) = exp j2π f0 t +
2Tp
(4.1)
where f0 = center frequency, B = bandwidth and Tp = pulse duration of s(t).
Applying Fourier transform the spectrum of s(t) is calculated as
Z ∞
S (f ) =
s(t)e−j2πf t dt
(4.2)
−∞
From (4.1) and (4.2) the spectrum is expressed [73] as
r
πTp
Tp
2
[Z(u2 ) − Z(u1 )] e−j B (f −f0 )
S (f ) =
2B
where complex Fresnel integral Z(u) is
Z u
Z u
π π 2
x dx + j
x2 dx
sin
Z(u) =
cos
2
2
0
0
79
(4.3)
(4.4)
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Chapter 4
Amplitude
Real part of chirp signal
0.5
0
−0.5
−5
−4
−3
−2
−1
0
Time in µ sec
1
2
3
4
5
2
3
4
5
Imaginary part of chirp signal
Amplitude
1
0.5
0
−0.5
−5
−4
−3
−2
−1
0
Time in µ sec
1
Figure 4.1: Real and imaginary part of the chirp signal for T B = 50
The arguments u1 and u2 given by
r
u1 = −2(f − f0 )
Tp
−
2B
r
Tp B
2
(4.5)
r
Tp
Tp B
+
(4.6)
u2 = −2(f − f0 )
2B
2
Z(u) is a function of Fresnel integrals cosine C(u) and sine S(u), where
Z u
π C(u) =
cos
x2 dx
(4.7)
2
0
Z u
π (4.8)
x2 dx
sin
S(u) =
2
0
The Fresnel ripple values defined in (4.7) and (4.8) are high at small arguments of u
r
and vice-versa. The real and imaginary part of the envelope of s(t) and corresponding
spectrum for T B = 50 are depicted in Figures 4.1 and 4.2 respectively. From Figure
4.2 it is observed that the spectrum contains the ripples called as Fresnel ripples,
which causes the sidelobes after compression. The output of the receiver matched
filter or compression filter for T B = 50 is a pulse with
Figure 4.3.
80
sin(x)
x
envelope as shown in
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Chapter 4
80
Amplitude
70
60
50
40
30
20
10
−10
−8
−6
−4
−2
0
2
4
6
8
10
Frequency in MHz
Figure 4.2: Amplitude spectrum of chirp signal for T B = 50
Filter response in dB
0
−10
−20
−30
−40
−50
−1
−0.8
−0.6
−0.4
−0.2
0
Time delay × T
0.2
0.4
0.6
0.8
1
p
Figure 4.3: Compressed envelope
4.2.2
Polyphase codes
The complex envelope of phase coded pulse is expressed as [5]
where rect
t
tb
N
t − (i − 1)tb
1 X
ui rect
u(t) = p
tb
Tp i=1
= 1 for |t| ≤
tb
,
2
(4.9)
ui = exp(jφi ), tb is the sub pulse width and N is
the number of phases given as {φ1 , φ2 ..., φN }.
Polyphase codes have harmonically related phases based on a certain fundamental
phase increments. These codes have better Doppler tolerance and sidelobe
performance than biphase codes. These codes are discrete time sequences having
constant magnitude with a variable phase φi . Polyphase codes have more than
two elements or phase values. Increasing the number of elements in the sequence
enables construction of longer sequences having greater range resolution with a
larger compression ratio. But the trade off is that a more complex matched filter as
81
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Chapter 4
compared to the biphase codes. These codes are derived by detecting a frequency
modulated pulse compression waveform with either a local oscillator at the band
edge of the waveform (single sideband detection) or at band center (double sideband
detection) and the resultant in phase I and quadrature phase Q data are sampled
at Nyquist rate.
Polyphase codes are derived from step approximation to linear frequency
modulation, such as Frank, P1 and P2 or from linear frequency modulation like
P3 and P4 codes.
(a)Frank Code
The Frank code is derived from step approximation to a linear frequency modulated
waveform having N frequency steps and N samples per frequency [79]. So Frank
code having N frequency steps have length Nc = N 2 . The first N samples of the
code have zero phase. The second N samples start with zero phase and increase
with a phase value of 2π/N from sample to sample. The phase of ith sample of j th
frequency step is given by
φi,j =
2π
(i − 1)(j − 1)
N
(4.10)
where i = 1, 2, 3...N and j = 1, 2, 3...N .
The Frank code in N × N

0

 0


 0


 ...

0
matrix form is given as
0
0
...
1
2
...
2
4
...
...
...
N −1
2(N − 1)
...
...
0





2(N − 1) 



...

2
(N − 1)
N −1
(4.11)
where the number in matrix represent the multiplying coefficient with the phase
angle 2π/N . The Frank polyphase code is formed by concatenating the rows of the
Frank matrix and multiplying by fundamental phase increment 2π/N .
The 16-element Frank code is given by
82
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Chapter 4
Filter response in dB
0
−10
−20
−30
−40
−50
0
20
40
60
80
100
120
140
160
180
200
70
80
90
100
Time delay
(a) Matched filter response
6
Phase in radian
5
4
3
2
1
0
0
10
20
30
40
50
60
Sample number
(b) phase
Figure 4.4: Matched filter output and phase values of 100 element Frank code
0 0 0 0 0
π
2
π
Taking modulo 2π gives
0 0 0 0 0 π2 π
3π
2
0 π
2π
3π
2
0 π
0 π
3π
0
0
3π
2
3π
2
9π
2
3π
π
π
2
The matched filter response of 100 element Frank code and its phase values are given
in Figure 4.4. From the figure it is clear that the peak sidelobe occurs below 30 dB
of the main lobe.
(b)P1 Code
These codes are also derived from step approximation to linear frequency
modulation. In case of single sideband detection Frank code is generated and in
case of double sideband detection P1 code is generated. P1 code also consists of
N × N elements and the phase of the ith element of j th group is given by
φi,j = −
π
[N − (2j − 1)][(j − 1)N + (i − 1)]
N
(4.12)
where i = 1, 2, 3...N and j = 1, 2, 3...N .
The matched filter response and phase values of 100 element P1 code is depicted in
83
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Chapter 4
Filter response in dB
0
−10
−20
−30
−40
−50
0
20
40
60
80
100
120
140
160
180
200
70
80
90
100
Time delay
(a) Matched filter response
7
Phase in radian
6
5
4
3
2
1
0
0
10
20
30
40
50
60
Sample number
(b) phase
Figure 4.5: Matched filter output and phase values of 100 element P1 code
Figure 4.5. From Figures 4.4(a) and 4.5(a) it is observed that the ACFs of Frank
and P1 codes are same. But the difference between P1 code and the Frank code is
that P1 code has the highest phase increments from sample to sample at the two
ends of the code but the Frank code has the highest phase increments from sample
to sample in the center of the code. So when the codes are passed through band
pass amplifier of a radar receiver, the P1 code is attenuated mostly at the two ends
of the waveform while the Frank code is attenuated most heavily in the center of
the waveform.
(c)P2 Code
In P2 code the starting phases are different from P1 code but the phase increments
within each phase group is same as that of P1 codes. The phases of P2 codes of
length N 2 is given by
φi,j =
π
π
[N − 1] − (i − j)[N + 1 − 2j]
2N
N
84
(4.13)
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Chapter 4
where i = 1, 2, 3...N and j = 1, 2, 3...N .
The matched filter response and phase value of 100 element P2 code is shown in
Figure 4.6. It has same matched filter output as that of Frank and P1 codes. But,
the P1 and P2 codes are more precompression bandwidth limit tolerant than the
Frank code.
Filter response in dB
0
−10
−20
−30
−40
−50
0
20
40
60
80
100
120
140
160
180
200
70
80
90
100
Time delay
(a) Matched filter response
7
Phase in radian
6
5
4
3
2
1
0
0
10
20
30
40
50
60
Sample number
(b) phase
Figure 4.6: Matched filter output and phase values of 100 element P2 code
(d)P3 Codes
P3 codes [81] are derived from the phase samples LFM signal. These codes are
obtained by converting a LFM waveform to base band using a local oscillator on
one end of the frequency sweep (single sideband detection) and sampling the I and
Q video at Nyquist rate.
Let the waveform have pulse duration Tp and the instantaneous frequency is
f (t) = f0 + kt
85
(4.14)
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Chapter 4
where k is a constant and defined as
k=
B
Tp
(4.15)
B is the bandwidth of the signal. So the signal will support a compressed pulse
length of
tc = 1/B
(4.16)
The pulse compression ratio is obtained as
ρ = Tp /tc = BTp
(4.17)
If the first samples of I and Q are taken at the leading edge of the waveform, the
phases of P3 codes are given by
Z (i−1)tc
φi = 2π
[(f0 + kt) − f0 ]dt = πk(i − 1)2 t2c
(4.18)
0
From (4.17) and (4.18)
φi = π(i − 1)2 /BTp = π(i − 1)2 /ρ
(4.19)
The matched filter output and phase values of 100 element P3 code are shown in
Figure 4.7. From Figure 4.7(a) it is observed that the peak side lobe occurs below
27 dB from the main peak lobe which is 3 dB inferior from the Frank code.
(e)P4 Codes
The P4 code is derived from the same waveform as P3 codes but the local oscillator
frequency is set at f0 + kTp /2. So the phase of P4 codes are
φi = 2π
Z
0
(i−1)tc
[(f0 + kt) − (f0 + kTp /2)]dt = [π(i − 1)2 /ρ] − π(i − 1)
(4.20)
The matched filter output and phase values of 100 element P4 code is shown in
Figure 4.8. From Figures 4.7(a) and 4.8(a) it is observed that both P3 and P4 codes
have the same matched filter output. But P4 code is more tolerant to precompression
bandwidth limitation as compared to P3 code.
86
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Chapter 4
Filter response in dB
0
−10
−20
−30
−40
−50
0
20
40
60
80
100
120
140
160
180
70
80
90
Time delay
(a) Matched filter response
7
6
Phase angle
5
4
3
2
1
0
0
10
20
30
40
50
60
100
Sample number
(b) phase
Figure 4.7: Matched filter output and phase values of 100 element P3 code
Filter response in dB
0
−10
−20
−30
−40
−50
0
20
40
60
80
100
120
140
160
180
200
70
80
90
100
Time delay
(a) Matched filter response
7
Phase angle
6
5
4
3
2
1
0
0
10
20
30
40
50
60
Sample Number
(b) phase
Figure 4.8: Matched filter output and phase values of 100 element P4 code
87
Chapter 4
4.3
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Problem formulation
The LFM signal and polyphase codes are used for transmission and the reflected
waveforms are passed through weighted matched filter. If the target is in motion
the reflected waveform is Doppler shifted version of the transmitted waveform. The
Doppler shifted waveform is passed through the weighted receiver matched filter and
PSRs are calculated for different Doppler shift conditions.
4.3.1
For LFM signal
The matched filter is given by
H(f ) = F [s(t)w(t]∗
(4.21)
where w(t) is the window function. The matched filter output is obtained as
g(t) = F −1 [S(f )H(f ]
(4.22)
where S(f ) = F [s(t)].
The Doppler shifted version of the transmitted signal s(t) is represented as
B 2
|t| ≤ Tp /2
(4.23)
t
sd (t) = exp j2π (f0 + fd )t +
2Tp
where fd is the Doppler shift. The Doppler shifted signal is passed through the
weighted matched filter and the PSR values under different Doppler shift are
obtained. To achieve higher magnitude of PSR the transmitted signal is modified
using amplitude tapering and phase predistortion function as explained below
(I) Amplitude tapering: The Fresnel ripples can be reduced by adding cosine
taper of length αTp to the LFM pulse. The amplitude tapered transmitted
signal is represented as
n
s1 (t) = gT (t) exp j2π f0 t +
B 2
t
2Tp
88
o
|t| ≤ (0.5 + α) Tp
(4.24)
Chapter 4
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
where



1


o
n
p /2
gT (t) =
0.5 1 + cos π |t|−T
αTp



 0
|t| ≤ Tp/2
Tp/ ≤ |t| ≤ (0.5 + α) Tp
2
elsewhere
(4.25)
and α is a parameter.
(II) Phase distortion: In high power pulsed radars the pulse rise time cannot be
controlled easily. So an alternate approach which uses an appropriate phase
distortion of the transmitted LFM signal is used for fast rise time transmitters.
LFM waveform having cubic phase distortion is given by
h
n
s2 (t) = exp j2π f0 t +
B 2
t
2Tp
oi
+ φ(t)
|t| ≤ Tp/2 + ∆T
(4.26)
where


∆B
Tp/ )3

2 (−t −

2

 3∆T
∆B
φ (t) =
(t − Tp/2)3
3∆T 2




0
− Tp /2 − ∆T ≤ t < −Tp /2
Tp /2 ≤ t < Tp /2 + ∆T
(4.27)
elsewhere
and ∆B and ∆T are the parameters.
4.3.2
For polyphase codes
These codes have discrete values having different phases. The matched filter is
designed according to the transmitted polyphase code. The filter is multiplied with
the window functions to achieve lower sidelobes. The Doppler shifted signal is
modeled by multiplying ej2πifd /B to the transmitted signal, where i = 1, 2, 3...N
and N is the code length.
The PSR values are calculated using conventional and convolutional windows as
weighing function at the receiver end under various Doppler shift conditions.
89
Chapter 4
4.4
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Windows used for sidelobe suppression
Windows are time domain weighing functions that are used to reduce Gibbs
oscillations caused by the truncation of a Fourier series. They are employed in a
variety of traditional applications including power spectral estimation, beam forming
and digital filter design [91]. Many windows appeared in the literature are not
optimal. So the use of a particular window depends upon the application. A number
of windows and algorithms are formulated to find an optimal window for a given
application [92–96]. The typical windows that are used in the signal processing
techniques are based on cosine series such as Hamming, Hanning and Blackman [92].
Although many windows have already been introduced in the literature, research is
going on to propose new windows or to parameterize the known windows [97–100].
Window functions are generally categorized as fixed or adjustable. Fixed windows
have window length as the parameter which alters the mainlobe width. Adjustable
windows have two parameters, namely the window length and a parameter that
alters the relative sidelobe amplitude. The best known parametric windows in
the literature are Dolph-Chebysev [101] and Kaiser [102] windows. By varying the
two adjusting parameters of the Kaiser window it can control the mainlobe width
and ripple ratio of the spectrum. Polynomial windows having low computational
complexity is presented in [103]. The frequency response of these windows can
easily be changed by modifying their coefficients in the time domain. Avci and
Nacaroglu [104] have proposed a new class of cosine hyperbolic windows having
low computational complexity due to power series expansion in its time domain
representation.
Convolutional windows are derived by convolving the window with itself.
Reljin et. al. [105] have discussed a class of windows that are generated by the
time convolution of classical windows to obtain both flat top and high sidelobe
attenuation. These windows are suitable for harmonic amplitude evaluation in
nonsynchronous sampling case. The convolutional windows from second to eighth
order for rectangular window are derived in [106]. These windows are applied for
90
Chapter 4
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
high accuracy harmonic analysis and parameter estimation of periodic signals. Phase
difference algorithm based on Nuttal self-convolutional window is used to eliminate
the measurement errors of dielectric loss factor [107]. Dielectric loss factor is caused
by non-synchronised sampling and non-integral periodic truncation conditions. A
self convolution Hanning window used for complex signal harmonics parameter
estimation has been presented in [108]. The convolutional window based phase
correction algorithm suppresses the impact of fundamental frequency fluctuation
and white noise on harmonic estimation. In this chapter convolutional windows are
used for weighing purpose to reduce the range sidelobes that are present in output
of pulse compression filter.
The windows employed for the analysis are
(i) Hamming window:
w(n) = 0.54 − 0.46 cos
2πn
N −1
(4.28)
where n = 0, 1, ........(N − 1)
(ii) Hanning window:
w(n) = 0.5 − 0.5 cos
2πn
N −1
(4.29)
(iii) Kaiser window:
w(n) =
q
2n 2
I0 β 1 − ( N −1 )
I0 (β)
(4.30)
where I0 is the zeroth order modified Bessel function of the first kind and β is
a parameter determines the shape of the window.
(iv) Chebysev window: The Dolph-Chebyshev window is in the frequency domain
is represented as
)]]
[β1 cos( πk
N
−1
cosh[N cosh (β1 )]
k cos[N cos
W (k) = (−1)
−1
(4.31)
where β1 = cosh[ N1 cosh−1 (10α1 )] and α1 determines the level of the sidelobe
attenuation. The sidelobe level σ(dB) = 20α1 . The Dolph-Chebyshev window
91
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Chapter 4
0
Hamming
Conv. Hamming
−10
20*log10(amplitude)
−20
−30
−40
−50
−60
−70
−80
50
100
150
200
250
Sample number
(a) Hamming and Convolutional Hamming window
0
Hamming
Conv. Hamming
−10
20*log10(amplitude)
−20
−30
−40
−50
−60
−70
−80
130
140
150
160
170
180
190
200
210
220
230
Sample number
(b) Zoomed version
Figure 4.9: Frequency response curve
is obtained by taking the inverse DFT of W (k) and scaling the result to have
a peak value of 1.
The convolutional windows are obtained by convolving a particular window with
itself. An N point convolutional window is obtained by convolving two N/2 point
windows. After convolution of two N/2 point windows the number of samples is
N − 1. So a zero is padded to the convolution result to make the length of the
window N and the maximum value is normalized to 1. Frequency responses of
Hamming and convolutional Hamming windows are presented in Figure 4.9. From
this figure, it is observed that the sidelobes of convolutional Hamming window are
lower at the cost of wider mainlobe.
4.5
Simulation results
The windows available in the MATLAB library are used for the simulation purpose.
92
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Chapter 4
Filter response in dB
0
−10
−20
−30
−40
−50
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
Time delay × Tp
0.6
0.8
1
Figure 4.10: Matched filter output with Hamming weighing at the receiver
4.5.1
Analysis for LFM signals
LFM signal with low T B products is associated with Fresnel ripples. The compressed
output of an LFM signal having Hamming weighing at the receiver for T B = 50
is shown in Figure 4.10. The peak sidelobe is approximately 37 dB lower than
the main peak. The outputs of matched filter for different Doppler shifts using
Hamming and convolutional Hamming window are depicted in Figure 4.11. It
is observed that at zero Doppler shift Hamming window yields better PSR value
as compared to convolutional window. As the Doppler shift increases sidelobe
level affected very less in case of convolutional Hamming window as compared to
Hamming window. The PSR values under different Doppler shifts using various
windows are presented in Table 4.1. It is observed from the table that at lower
Doppler shift the conventional windows yield better PSR values as compared to
corresponding convolutional windows. On the other hand for higher Doppler shifts
the convolutional windows provide better PSR values than conventional ones. As
an illustration for
fd
B
= 0.01 the PSR for Hamming window is -36.2 dB and that of
convolutional Hamming window is -34.46. But, for
fd
B
= 0.2 the PSR for Hamming
window is -22.2 dB and that of convolutional Hamming window is -29 dB. From
Figure 4.11 it is obvious that the mainlobe width in case of convolutional window
is wider and the sidelobes near |t| =
Tp
2
93
region are not diminished by weighing
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Chapter 4
0
Filter response in dB
Hamming
Conv. Hamming
−10
−20
−30
−40
−50
−1
−0.8
−0.6
−0.4
−0.2
0
Time delay×Tp
(a)
fd
B
0.2
0.4
0.6
0.8
1
=0
0
Filter response in dB
Hamming
Conv. Hamming
−10
−20
−30
−40
−50
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
Time delay×Tp
(b)
fd
B
0.4
0.6
0.8
1
= 0.05
0
Filter response in dB
Hamming
Conv. Hamming
−10
−20
−30
−40
−50
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
Time delay×T
p
(c)
fd
B
= 0.1
Figure 4.11: Effect on sidelobes due to Doppler shift
94
0.8
1
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Chapter 4
Table 4.1: Comparison of PSR for different Doppler shift for T B = 50
Doppler
Shift( fBd )
PSR using Hamming
window in dB
PSR using convolutional
Hamming window in dB
0.01
0.05
0.1
0.15
0.2
Doppler
Shift( fBd )
-36.2
-32.8
-28.6
-25.2
-22.2
PSR using Hanning
window in dB
-34.46
-33.67
-32.5
-31
-29
PSR using convolutional
Hanning window in dB
0.01
0.05
0.1
0.15
0.2
Doppler
Shift ( fBd )
-31.68
-30.32
-27.62
-24.47
-22
PSR using Kaiser
window in dB (β = 6)
-33.67
-33
-31.79
-30.44
-28.7
PSR using convolutional
Kaiser window in dB
0.01
0.05
0.1
0.15
0.2
-36.6
-33
-28.7
-25.2
-22.3
-34.2
-33.4
-32.26
-30.86
-29
Doppler
Shift( fBd )
0.01
0.05
0.1
0.15
0.2
PSR using Chebysev
PSR using convolutional
window in dB (σ = 50) Chebysev window in dB
-36
-34.89
-29.89
-26.86
-23.86
-34.3
-33.5
-32.5
-31
-29
technique. Hence to reduce the sidelobes at
Tp
2
region as well as peak sidelobe,
amplitude tapering and phase predistortion are used .
Amplitude tapering with α = 0.1 is used to modify the transmitter signal. In
simulation study the amplitude tapers are not used in compression filter. The
filter responses for amplitude tapering using Hamming and convolutional Hamming
95
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Chapter 4
windows are depicted in Figure 4.12. From Figure 4.12(a), it is clear that amplitude
tapering reduces drastically the sidelobes around |t| = Tp /2 region but the reduction
of near in sidelobes is very less. From Figure 4.11, it is evident that the compressed
output of convolutional Hamming window produces peak sidelobe around |t| = Tp /2
region. So to get very low overall sidelobe the transmitted signal is amplitude tapered
and the receiver is weighed with convolutional Hamming window. The output
of pulse compression filter using convolutional Hamming window with amplitude
tapered transmitted signal is presented in Figure 4.12(b). The PSR values using
Hamming, Kaiser, convolutional Hamming and convolutional Kaiser windows for
T B = 50 and T B = 100 for different Doppler shifts are listed in Table 4.2. From
the table, it is observed that for higher T B product the PSR value is better and
amplitude tapering with convolutional window provides better PSR value compared
to that of conventional windows.
Table 4.2: PSR using amplitude tapering
Doppler
Shift
fd
B
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
Hamming
window
PSR in dB
TB=50 TB=100
-42.2
-43
-41.2
-42.7
-39
-41.7
-36.6
-39.3
-34.5
-36.8
-32.6
-34.7
-30.8
-32.7
-29.2
-31
-27.8
-29.4
-26.4
-28
-25.2
-26.5
Convolutional Hamming
window
PSR in dB
TB=50
TB=100
-60
-78.7
-58
-75
-55.5
-71.4
-53
-67.3
-50.2
-63.5
-47.7
-60
-45.3
-55.7
-43
-52
-40.5
-48
-38.1
-44.4
-35.7
-41.2
Kaiser
window (β = 6)
Convolutional Kaiser
window
PSR in dB
TB=50 TB=100
-43.4
-44.5
-41.8
-43.8
-40
-42.5
-38.25
-40.8
-36.6
-39
-35
-37.3
-33.3
-35.7
-31.2
-33.4
-29.3
-31.2
-27.6
-29.8
-26
-28
PSR in dB
TB=50
TB=100
-61.2
-79.6
-59.4
-78.2
-56.6
-75.5
-54
-71.38
-51.2
-66.8
-48.6
-62.7
-46.1
-58.5
-43.75
-54.5
-41.3
-50.7
-39
-46.8
-36.5
-43.3
In case of cubic phase distorted transmitted signal the parameter values used are
∆B = 0.75B and ∆T = 1/B. The filter responses for cubic phase distortion using
Hamming and convolutional Hamming windows are depicted in Figure 4.13. From
96
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Chapter 4
Filter response in dB
0
−20
−40
−60
−80
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
Time delay × Tp
0.4
0.6
0.8
1
(a) Compressed waveform for amplitude tapering with Hamming window
Filter response in dB
0
−20
−40
−60
−80
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
Time delay × Tp
0.4
0.6
0.8
1
(b) Compressed waveform for amplitude tapering with convolutional Hamming window
Figure 4.12: Compressed waveforms for T B = 50 for amplitude tapering (α = 0.1)
Filter response in dB
0
−20
−40
−60
−80
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
Time delay× T
p
(a) Compressed waveform for cubic phase distortion with Hamming window
Filter response in dB
0
−20
−40
−60
−80
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
Time delay× Tp
(b) Compressed waveform for cubic phase distortion with convolutional Hamming window
Figure 4.13: Compressed waveforms for T B = 50 for cubic phase distortion (∆B =
0.75B and ∆T = B1 )
97
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Chapter 4
Figure 4.13(a) it is evident that the compressed output of phase distorted signal
also reduces the sidelobe level around |t| = Tp /2 region. But in this case the fall
off rate of far sidelobes is lesser as compared to amplitude tapering. The output
of pulse compression filter using convolutional Hamming window with cubic phase
distorted transmitted signal is presented in Figure 4.13(b). The PSR values using
Hamming, Kaiser, convolutional Hamming and convolutional Kaiser windows for
T B = 50 and T B = 100 for different Doppler shifts are listed in Table 4.3. The
table illustrates that convolutional windows provide better PSR values as compared
to that of conventional windows. For a particular window, the PSR values for
T B = 100 is better than that of T B = 50.
Table 4.3: PSR using cubic phase distortion
Doppler
Shift
fd
B
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
4.5.2
Hamming
window
PSR in dB
TB=50 TB=100
-39.8
-41.8
-37.6
-39.2
-35.5
-36.8
-33.6
-34.7
-31.8
-32.8
-30.2
-31
-28.8
-29.4
-27.4
-27.9
-26.2
-26.6
-25
-25.3
-23.8
-24.4
Convolutional Hamming
window
PSR in dB
TB=50
TB=100
-49.7
-55.3
-48.4
-53.6
-46.7
-52
-45
-50.3
-43.4
-48.8
-41.8
-47.2
-40.2
-45.6
-38.6
-43.9
-36.8
-41.2
-34.7
-38.1
-32.4
-35.3
Kaiser
window (β = 6)
Convolutional Kaiser
window
PSR in dB
TB=50 TB=100
-41.2
-42.2
-39.7
-41.1
-37.8
-39.4
-35.5
-37.3
-33.9
-35.2
-32
-33.1
-30.3
-31.3
-28.7
-29.5
-27.2
-28
-26
-26.4
-24.6
-25
PSR in dB
TB=50
TB=100
-61
-79.5
-59.3
-78
-56.5
-75
-53.8
-71
-51.1
-66.4
-48.6
-62.3
-46.1
-58.1
-43.7
-54.1
-41.3
-50.3
-39
-46.5
-36.4
-43
Analysis for polyphase codes
Polyphase codes derived from the step approximation to LFM signal (Frank, P1 and
P2 ) do not provide satisfactory results using weighing technique. The output of the
Hamming weighted matched filter for 100 element Frank code is depicted in Figure
4.14 which shows that the peak sidelobe level 25 dB below the mainlobe peak. From
Figures 4.4(a) and 4.14 it is clear that with Hamming weighing the peak sidelobe
98
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Chapter 4
level is increased. So windows are not used for these codes for sidelobe suppression.
From Figures 4.7 and 4.8 it is evident that the matched filter output for P3
Filter response in dB
0
−10
−20
−30
−40
−50
0
20
40
60
80
100
120
140
160
180
200
Time delay
Figure 4.14: Matched filter output for 100 element Frank code using Hamming
window
and P4 codes are identical. The matched filter output of P3 and P4 under different
Doppler shift is depicted in Figures 4.15 and 4.16 respectively. These figures show
that under Doppler shift both P3 and P4 codes provide same performance. So P4
code of length 100 is used for further simulation study and the results are valid also
for P3 codes. Figure 4.17 illustrates the output of matched filter under Doppler
shift when the receiver filter weighted with Hamming and convolutional Hamming
windows. It is observed that at higher Doppler shifts the convolutional windows
provide improved results as compared to that of conventional windows. The PSR
values under different Doppler shifts using different windows are presented in Table
4.4. It is observed that for lower values of Doppler shift the PSR values for classical
windows are better than that of convolutional windows. But as the Doppler shift
increases the PSR magnitudes for classical windows drop rapidly as compared to
convolutional windows.
99
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Chapter 4
Filter response in dB
0
−10
−20
−30
−40
−50
0
20
40
60
80
100
120
140
160
180
120
140
160
180
Time delay
(a)
fd
B
= 0.05
Filter response in dB
0
−10
−20
−30
−40
−50
0
20
40
60
80
100
200
Time delay
(b)
fd
B
= 0.1
Figure 4.15: Matched filter output of P3 code under different Doppler shift
Filter response in dB
0
−10
−20
−30
−40
−50
0
20
40
60
80
100
120
140
160
180
120
140
160
180
Time delay
(a)
fd
B
= 0.05
Filter response in dB
0
−10
−20
−30
−40
−50
0
20
40
60
80
100
200
Sample Number
(b)
fd
B
= 0.1
Figure 4.16: Matched filter output of P4 code under different Doppler shift
100
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Chapter 4
Filter response in dB
0
Hamming
Conv. Hamming
−10
−20
−30
−40
−50
0
20
40
60
80
100
120
140
160
180
Time delay
(a)
fd
B
=0
Filter response in dB
0
Hamming
Conv. Hamming
−10
−20
−30
−40
−50
0
20
40
60
80
100
120
140
160
180
Time delay
(b)
fd
B
= 0.05
Filter response in dB
0
Hamming
Conv. Hamming
−10
−20
−30
−40
−50
0
20
40
60
80
100
120
140
160
Time delay
(c)
fd
B
= 0.1
Figure 4.17: Effect on sidelobes due to Doppler shift
101
180
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
Chapter 4
Table 4.4: Comparison of PSR for different Doppler shift
Doppler
Shift( fBd )
PSR using Hamming
window in dB
PSR using convolutional
Hamming window in dB
0.01
0.05
0.1
0.15
0.2
Doppler
Shift( fBd )
-40
-37.3
-32
-27
-22.18
PSR using Hanning
window in dB
-37.17
-37
-36.5
-35.7
-33
PSR using convolutional
Hanning window in dB
0.01
0.05
0.1
0.15
0.2
Doppler
Shift ( fBd )
-40
-39.6
-37.1
-30
-25
PSR using Kaiser
window in dB (β = 6)
-36.3
-36.1
-35.7
-35
-33.6
PSR using convolutional
Kaiser window in dB
0.01
0.05
0.1
0.15
0.2
Doppler
Shift( fBd )
-40.1
-36.8
-37.9
-36.7
-33.8
-36.2
-28.2
-35.4
-23
-34
PSR using Chebysev
PSR using convolutional
window in dB (σ = 50) Chebysev window in dB
0.01
0.05
0.1
0.15
0.2
4.6
-39
-37.9
-33.3
-27.8
-22.7
-37
-36.8
-36.3
-35.5
-33.8
Conclusion
In this chapter the ability of convolutional windows to suppress the sidelobes are
analyzed and the results are compared with that of conventional windows. Although
the magnitude of PSR at lower Doppler shift in Table 4.1 and 4.4 are better in case of
102
Chapter 4
Effective Sidelobe Suppression of
LFM and Polyphase Codes Using Convolutional Windows
conventional windows, the convolutional windows provide better PSR value at higher
Doppler shift. In case of LFM signal, to decrease the sidelobes around |t| =
Tp
2
region
the transmitted signal is modified using amplitude tapering or phase distortion. It is
further demonstrated that the PSR values of amplitude tapered or phase distorted
transmitted signal with convolutional windows are better than that of conventional
windows at all Doppler shift conditions. However, the mainlobe width achieved
using convolutional windows is wider than that of conventional windows.
103
Chapter 5
Efficient Design of Stepped
Frequency Pulse Train Using
Evolutionary Computation
Techniques
5.1
Introduction
In high range resolution radar, signals having wide bandwidth are used to get narrow
mainlobe width. Generation of such type of wideband waveforms increases the
overall cost and complexity of the system. The conventional narrowband hardware
used in the radar system may not sustain instantaneous wide bandwidth.
To
overcome such limitation the wide bandwidth signal is split into a set of narrowband
signals which are transmitted and received separately.
The effect of wideband
signal is obtained by coherently combining the narrowband signals. Such type of
narrowband signals together is called as ‘synthetic wideband waveform’ [109] or
‘stepped frequency waveform’ or ‘frequency jumped train’.
Generally a pulse train consists of N pulses each of duration Tp and pulse
repetition time Tr . Each pulse has a bandwidth B and center frequency step between
the pulses is ∆f . The amplitude and frequency of a stepped frequency LFM pulse
train is shown in Figure 5.1. In the proposed work the values of Tp , B and ∆f
are assumed to remain constant throughout the pulse train and satisfy the condition
104
Chapter 5
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
B > ∆f > 0. One of the advantages of this type of signal is that the interval between
pulses is utilized to adjust the center frequency of other narrowband components of
the radar system. But matched filter output of such signals suffer from grating lobes
for cases when Tp ∆f > 1 due to constant frequency step ∆f . These grating lobes
reduce the range resolution capability of the signal and hence these are undesirable.
Figure 5.1: Stepped frequency LFM pulse train
Different techniques for acceptable suppression or complete rejection of grating
lobes are dealt in [110–115]. In [110, 111] grating lobes are reduced by varying
the pulse width of the pulse train which destroys the periodicity of the waveform.
An approach to generate a nonlinear synthetic wideband waveform by distributing
the energy nonuniformly over the desired bandwidth is described in [112].
It
offers improved performance in terms of lower range sidelobes, higher range
resolution and/or reduced grating lobes. Levanon and Mozeson [113] have proposed
an analytical technique to establish the relation between parameters of stepped
frequency LFM pulse train such that the first two grating lobes are nullified. They
have also shown in some cases that nullifying the first two grating lobes leads to
removal of all other grating lobes. To establish the required relation between the
parameters Tp , B and ∆f using this approach for more than two grating lobes is too
difficult. In this chapter PSO based technique is suggested which aims to eliminate
105
Chapter 5
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
all the grating lobes by judiciously choosing appropriate values of Tp , B and ∆f .
The method presented in [113] does not suppress the range sidelobes that occur
at the output of the matched filter of the receiver. In [114, 115] the LFM pulses
are replaced by nonlinear LFM pulses that suppress the range sidelobes near the
mainlobe along with grating lobes. However, the nonlinear LFM signals are not
Doppler tolerant.
The LFM pulse can be easily generated and more Doppler tolerant than NLFM
pulse. Therefore LFM waveforms are widely used in pulse radar systems. But
the techniques used in [113–115] to suppress the grating lobes of LFM pulse train
ignore the mainlobe width and PSL. The waveform having wide mainlobe width in
its ACF has low range resolution capability and the waveform that yields high peak
sidelobe in its ACF may hide the small targets or cause false target detection. Hence
there is a need to develop an efficient method to determine the parameters of the
stepped frequency LFM pulse train by considering grating lobes, mainlobe width
and PSL. Keeping this fact in view, in this chapter a new optimization is proposed
approach using NSGA-II algorithm to achieve reduced grating lobes, lower sidelobes
and narrow mainlobe width.
5.2
LFM pulse train
The envelope of a constant frequency or unmodulated pulse of duration Tp is given
by
1
u(t) = p rect
Tp
t
Tp
(5.1)
Frequency modulation is applied to the above constant frequency pulse to get an
LFM signal and its complex envelope is represented as
t
1
exp(jπkt2 )
u1 (t) = p rect
Tp
Tp
(5.2)
where k is the frequency slope of the LFM signal and is defined as
k=±
106
B1
Tp
(5.3)
Chapter 5
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
B1 is the bandwidth of the single pulse. “+” and “-” signs stands for positive
frequency slope and negative frequency slope respectively. Instantaneous frequency
of the LFM signal is obtained as
f (t) =
1 d(πkt2 )
= kt
2π dt
(5.4)
A uniform pulse train having N number of LFM pulses separated by Tr ≥ 2Tp is
expressed as
The multiplication factor
N −1
1 X
u1 (t − nTr )
uN (t) = √
N n=0
√1
N
(5.5)
is included in (5.5) to maintain unit energy. Further
a slope ks is added to the entire LFM pulse train and the complex envelope is
represented as
N
−1
X
1
2
u1 (t − nTr )
us (t) = uN (t)exp(jπks t ) = √ exp(jπks t )
N
n=0
2
(5.6)
where
ks = ±
∆f
Tr
∆f > 0
(5.7)
“+” and “-” signs correspond to positive and negative frequency step respectively.
In this work “+” sign of k and ks is used, but the results equally hold good for “-”
sign also.
So the final bandwidth of each pulse in the LFM pulse train is
B = (k + ks )Tp
(5.8)
The total bandwidth of the LFM pulse train is B + (N − 1)∆f .
The ACF of the signal us (t) is obtained [5] as
|τ | sin(N πτ ∆f ) |τ |
sinc Bτ 1 −
|R(τ )| = 1 −
Tp
Tp N sin(πτ ∆f ) (5.9)
In (5.9) the expression of R(τ ) consists of product of two terms out of which the
first term is the ACF of a single LFM pulse and is given by
|τ
|
|τ
|
sinc Bτ 1 −
|R1 (τ )| = 1 −
Tp
Tp 107
(5.10)
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
Chapter 5
and the second term
sin(N πτ ∆f ) |R2 (τ )| = N sin(πτ ∆f ) produces the grating lobes at τg =
g
∆f
(5.11)
where g = 1, 2, 3.... ⌊Tp ∆f ⌋. These grating
lobes appear in the form of high spikes and reduce the range resolution potential of
the waveform. Nullifying or suppressing these grating lobes essentially depends upon
the occurrence of nulls or minima of |R1 (τ )| at τg . In [113] an analysis is provided
that sets up simple relations between the pulse time duration Tp , its bandwidth
B and frequency step ∆f to nullify first two grating lobes of an LFM pulse train.
However, in some cases nullifying two grating lobes also removes all grating lobes.
Equation (5.9) can be written as
τ sin(N π∆f Tp Tτp ) τ τ
τ
= 1 − sinc Tp B
R
1 − ,
Tp N sin(πTp ∆f Tτp ) Tp Tp
Tp
τ ≤1
Tp (5.12)
From (5.9), (5.10), (5.11) and (5.12) it is clear that |R(τ )|, |R1 (τ )| and |R2 (τ )| are
functions of Tp ∆f and Tp B only for a given value of N . Figure 5.2 shows the plots
of |R1 (τ )|, |R2 (τ )| and ACF for Tp ∆f = 3, Tp B = 4.5 and N = 8. It is observed
that all the grating lobes are completely removed. For comparison purpose the ACF
obtained with fixed frequency pulse train is shown in Figure 5.3 in which the grating
lobes are prominent. The nullification of first two grating lobes always does not
guarantee that all other grating lobes will be nullified or suppressed because the
nulls of |R1 (τ )| do not occur periodically while the peaks of |R2 (τ )| occur with a
period
1
.
∆f
The mainlobe width depends on the first overall null of the expressions
|R1 (τ )| and |R2 (τ )|. The first null of |R2 (τ )| occurs at
|R1 (τ )| occurs at
1
Tp B
1
N Tp ∆f
and the first null of
approximately if Tp B >> 1. So the location of first null of
ACF is given by
τ1stnull
= min
Tp
1
1
,
Tp B N Tp ∆f
(5.13)
N ∆f should be always greater than B in order to get a meaningful increase in
bandwidth. So delay resolution is principally determined by |R2 (τ )| which is equal
to
1
.
N Tp ∆f
108
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
Chapter 5
| R1(τ) |, | R2(τ) |
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.6
0.7
0.8
0.9
1
0
ACF in dB
−10
−20
−30
−40
−50
−60
0.5
τ/ T
p
Figure 5.2: Stepped frequency LFM pulse for Tp ∆f = 3, Tp B = 4.5 and N = 8.
Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom: ACF (in dB)
| R1(τ) |,| R2(τ) |
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.5
0.6
0.7
0.8
0.9
1
ACF in dB
0
−20
−40
−60
τ / Tp
Figure 5.3: Stepped frequency LFM pulse for Tp ∆f = 3, Tp B = 0 and N = 8. Top:
|R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom: ACF (in dB)
109
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
Chapter 5
5.3
Problem formulation
The grating lobes reduce range resolution ability of the pulse train. So it is required
to suppress or nullify these grating lobes. In Section 5.3.1 a problem is formulated
in which the PSO is used to determine the parameters of stepped frequency pulse
train to nullify the grating lobes.
The problem of suppression of grating lobes, minimization of mainlobe width
and peak sidelobe level has been formulated in two different ways in a multiobjective
framework which are presented in Sections 5.3.2 and 5.3.3.
5.3.1
Problem formulation -1
The function defined in (5.10) must be minimum or zero at τ = τg , so that the
grating lobes would be suppressed or nullified. The fitness function which is to be
minimized using PSO is defined as
X |τ
|
|τ
|
g
g
1−
sinc Bτ 1 −
f1 =
T
T
p
p
g
(5.14)
subject to N ∆f > B.
By choosing suitable values for Tp B and Tp ∆f the grating lobes as well as sidelobes
in ACF can be suppressed. The value of Tp B is chosen such that Tp B = (c + 1)Tp ∆f
(where c is a positive number) to ensure B > ∆f , so that there will be some
frequency overlap between the pulses in spite of the frequency steps. If f1 = 0 then
each term in the summation is zero which results in complete elimination of grating
lobes otherwise the grating lobes are suppressed to a minimum level. PSO is used
to find out the required values of Tp ∆f and c so that f1 is minimized.
5.3.2
Problem formulation -2
The peak sidelobe should be as low as possible compared to the mainlobe so that
the target will be easily identified.
NSGA-II algorithm is used to choose the
values of Tp ∆f and c to achieve reduced grating lobes and minimum peak sidelobe
110
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
Chapter 5
level simultaneously. The two objective functions which are to be simultaneously
optimized are expressed as follows:
Minimize f2 = max|R1 (τg )| where g = 1, 2, 3..... ⌊Tp ∆f ⌋
Minimize f3 = P SR in dB
subject to N ∆f > B
5.3.3
Problem formulation -3
The range resolution of stepped frequency LFM pulse train depends upon the
mainlobe width of ACF and is given by
1
N Tp ∆f
for N Tp ∆f > Tp B. In the literature
generally weighing technique is used to suppress the sidelobes of an LFM pulse. The
weighing technique adds more emphasis on the center frequencies as compared to the
end frequencies. As a result the sidelobes are suppressed and mainlobe is widened,
which reduces the range resolution capability of the LFM signal. This effect is also
applicable for stepped frequency LFM pulse train as the condition B > ∆f > 0 is
assumed. The values of Tp ∆f and c are chosen by using a multiobjective algorithm
in such a way that the mainlobe width is lowered (for high range resolution) and
the sidelobes are suppressed. The effect of grating lobes is reduced by putting a
constraint so that the grating lobes are below a threshold level i.e. |R1 (τg )| < ǫ.
The fitness functions which are to be optimized simultaneously are defined as
Minimize f3 = P SR in dB
Minimize f4 =
1
N Tp ∆f
subject to N Tp ∆f > Tp B and |R1 (τg )| < ǫ.
5.4
Techniques used
In this chapter single objective evolutionary algorithm, PSO, and multiobjective
algorithm, NSGA-II, are used to determine the parameters of the LFM pulse train.
An overview of each of the algorithms is presented in sequel.
111
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
Chapter 5
5.4.1
Particle swarm optimization
The PSO was introduced by Kennedy and Eberhart [116] in 1995 which is a
population based and self adaptive search optimization technique. This algorithm
was developed based on simulation of animals social behavior such as bird flocking,
fish schooling etc. Like other population based evolutionary computation algorithm
such as GA, the PSO starts with random initialization of population, called as
swarm, in the search space and each individual is called as a particle. Unlike GA,
the PSO have not direct combination of genetic materials between the particles
during the search. The PSO algorithm employs the social behavior of the particle
in the swarm. Hence, it finds the global solution by adjusting the trajectory of each
particle towards its own best solution and towards the best particle of the swarm
in each generation [116–118]. The PSO is very popular because of the simplicity
of implementation of the algorithm and ability to converge quickly to a acceptable
good solution.
In PSO, the trajectory of each particle in search space is altered according to its
own velocity, own flying experience and flying experience of other particles in the
swarm.
The position of ith particle in D dimensional search space is given by
xi = [xi1 , xi2 , ..... xiD ]T
(5.15)
and the velocity of ith particle is expressed as
vi = [vi1 , vi2 , ..... viD ]T
(5.16)
The fitness function value is found out according to the user defined fitness function
which is to be optimized.
Let pbesti be the best position i.e. the best fitness value obtained by the ith
particle at time t. So
pbesti = [pbesti1 , pbesti2 , ..... pbestiD ]T
112
(5.17)
Chapter 5
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
The fittest particle found in the swarm at time t is
gbest = [gbest1 , gbest2 , ...... gbestD ]T
(5.18)
The position and velocity of each particle is updated as
vid (t + 1) = wvid (t) + c1 r1 (pbestid (t) − xid (t)) + c2 r2 (gbestd (t) − xid (t))
xid (t + 1) = xid (t) + vid (t + 1)
(5.19)
(5.20)
where d = 1, 2, ....D and w is a positive constant or positive linear or nonlinear
function of time [119, 120].
w is called inertia weight which plays the role of
balancing the local and global searches. c1 and c2 are two positive constants known
as acceleration coefficients and r1 and r2 are two random numbers in between 0
and 1. The first term in the right hand side of (5.19) corresponds to the previous
velocity which provides the necessary momentum and the second term stands for the
cognitive component which represents the personal thinking of each particle. The
cognitive component promotes the particles to move towards their own best position.
The third term is called as social component which constitutes the cooperative effect
of the particles in finding the global optimal solution. The social component always
drags the particle towards the global particle found so far.
The population is initialized with random positions and random velocities are
assigned to each particle. The fitness function value is evaluated according to defined
objective function i.e. to be optimized. At each generation the velocity and position
of the each particle are updated according to (5.19) and (5.20) respectively. In a
particular generation if a particle finds better position than previously found then
its location is stored in the memory. A maximum velocity i.e. Vmaxd is defined for
each dimension for the velocity vector vid in order to control the excessive roaming
of the particle outside the defined search space. If vid exceeds Vmaxd , then vid is set
to Vmaxd .
113
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
Chapter 5
5.4.2
NSGA-II
The NSGA-II which is described in Chapter 2 dealt with binary coded chromosome.
In this chapter the real coded chromosomes are used and the genetic operators
such as crossover and mutation are different than binary coded GA. For real coded
chromosomes simulated binary crossover and polynomial mutation are used to
generate offsprings. The steps such as population initialization, fitness function
evaluation, crowding distance assignment, selection, recombination process are same
as explained in Section 2.3.2 except the genetic operators.
Genetic operators:
Genetic operators such as crossover and mutation are used to explore and exploit
new and better solution from the existing solutions in the objective space. Real
coded NSGA-II uses simulated binary crossover [121, 122] and polynomial mutation
[122, 123] to produce offspring.
1. Simulated binary crossover: A random number y is generated between 0 and
1. From a defined probability distribution function (pdf) another variable α
is found such that the area under the pdf from 0 to α is equal to y. The pdf
is defined as

 0.5 (η + 1) αηc
c
P (α) =
 0.5 (ηc + 1) η1+2
α
c
if α ≤ 1
if α > 1
(5.21)
where ηc is the distribution index for crossover. This pdf is obtained by using
the transformation
α (y) =

1
 (2y) (ηc +1)
1

1
[2(1−y)] (ηc +1)
if y ≤ 0.5
otherwise
(5.22)
After obtaining α the off-spring children are computed as
1
[(1 − αk ) x1,k + (1 + αk ) x2,k ]
2
1
= [(1 + αk ) x1,k + (1 − αk ) x2,k ]
2
c1,k =
(5.23)
c2,k
(5.24)
114
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
Chapter 5
where αk is the value of α for k th component of the chromosome.
ci,k is the k th component of ith child.
xi,k is the k th component of ith parent which is selected for crossover.
2. Polynomial mutation: Mutation in GA restores lost or unexpected genetic
materials into the solution to avoid convergence of the algorithm into a
sub-optimal solution. The polynomial mutation is defined as
ck = xk + xuk − xlk δk
(5.25)
where ck is the mutated child produced from parent xk . xuk and xlk are the
upper and lower bound of xk . δk is the small variation which is obtained by

 (2r ) ηm1+1 − 1
k
δk =
 1 − [2 (1 − rk )] ηm1+1
if rk ≤ 0.5
otherwise
(5.26)
where ηm is the distribution index for mutation and rk is a random number in
between 0 and 1.
5.5
Determination of parameters of LFM pulse
train
5.5.1
Using PSO
The fitness function defined in (5.14) is minimized to determine the parameters of
LFM pulse train. The various steps are
1. The population of size M is initialized randomly in the given search space and
each particle in the population consists of two dimensions corresponds to Tp ∆f
and c. Random velocities are assigned to each particle .
2. The fitness function for each chromosome is evaluated according to (5.14).
The particle having best fitness value called as gbest. Initially the pbest for
a particle assumed as particle position itself.
115
Chapter 5
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
3. The velocity and position of each particle are updated as given in (5.19) and
(5.20) respectively.
4. The fitness function value is evaluated for the new position for each particle
and compared with the corresponding pbest positions fitness value. If for a
particle the new position fitness is better than that of pbest then the pbest
will be replaced by new particle. The particle having best fitness value among
all the pbests is selected as gbest.
Steps 3 and 4 repeated until the predefined condition is satisfied.
5.5.2
Using NSGA-II
The problems defined in Sections 5.3.2 and 5.3.3 use this algorithm to find the desired
parameter values of LFM pulse train. The various steps involved are
1. A population having M chromosomes is randomly initialized and each
chromosome contains two random values corresponds to Tp ∆f and c.
2. The fitness function values f2 and f3 (f3 and f4 for problem-3) are evaluated
as given in Section 5.3.2. (Section 5.3.3 for problem-3)
3. The chromosomes are sorted using nondominated sort and all possible fronts
are obtained as in Section 2.3.2.
4. The crowding distance for chromosomes in each front are evaluated according
to the procedure explained in Section 2.3.2.
5. The chromosomes are selected using binary tournament selection according to
Section 2.3.2.
6. The selected chromosomes undergo for genetic operations such as crossover
and mutation to produce offspring as explained Section 5.4.2.
7. The off-spring population is combined with parent population and the best M
chromosome selected for next generation as described in Section 2.3.2
116
Chapter 5
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
Steps from 5 to 7 are repeated until the maximum number of generations.
5.6
Simulation results
Simulation studies are carried out according to the problem formulation as presented
in Section 5.3. The techniques explained in Section 5.4 are used for grating lobe
suppression. To carry out PSO based estimation task the swarm size and number of
generations are chosen to be 100 and 50 respectively. This choice is based on trial
and error so as to achieve the best possible performance. The particles are randomly
initialized in the defined search space Tp ∆f ∈ [2, 15] and c ∈ [1, 10] . The values
of c1 and c2 are taken as 2. The velocity and position of the particles are updated
according to (5.19) and (5.20) respectively. At the end of all the generations if f1
attains a zero value then each term of the right hand side of (5.14) becomes zero
which means complete elimination of grating lobes. There are more than one set
of Tp ∆f and c present in the defined search space for which f1 = 0. At the end of
all the generations the best particle is saved and the program executed repeatedly
to get the other distinct best solutions. The values of Tp ∆f , Tp B and B/∆f for
N = 8 are listed in Table 5.1 for f1 = 0. Figures 5.4 and 5.5 show the plots of
|R1 (τ )|, |R2 (τ )| and |R(τ )| for Tp ∆f = 2.5, Tp B = 12.5 and Tp ∆f = 4, Tp B = 16
respectively. From these figures it is observed that the peaks of the |R2 (τ )| (grating
lobes) exactly coincide with the nulls of |R1 (τ )|, as a result there are nulls in |R(τ )|
at those points.
NSGA-II algorithm is employed for optimization of f2 and f3 associated in
problem 2. The population size and the number of generations are taken to be
100 and 50 respectively. The distribution indices for crossover (ηc ) and mutation
(ηm ) are chosen as 20 each. The probabilities of crossover and mutation are set to
be 0.9 and 0.1 respectively. Tp ∆f and c are the two variables judiciously chosen
by the NSGA-II algorithm to get low sidelobe level and reduced grating lobes. The
initialized population is sorted based on nondomination and each solution is assigned
with a crowding distance. The selection, crossover, mutation and recombination are
117
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
Chapter 5
|R (τ)|, |R (τ)|
1
2
0.8
0.6
1
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.5
0.6
0.7
0.8
0.9
1
ACF [dB]
0
−20
−40
−60
τ/T
p
Figure 5.4: Stepped frequency LFM pulse for Tp ∆f = 2.5, Tp B = 12.5 and N = 8.
Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom: ACF (in dB)
|R1(τ)|, |R2(τ)|
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.5
0.6
0.7
0.8
0.9
1
ACF [dB]
0
−20
−40
−60
τ/Tp
Figure 5.5: Stepped frequency LFM pulse for Tp ∆f = 4, Tp B = 16 and N = 8. Top:
|R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom: ACF (in dB)
118
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
Chapter 5
Table 5.1: Values of Tp ∆f , Tp B obtained for N = 8 and f1 = 0
Tp ∆f
2
2.5
3
3.5
4
5
6
7
9
11
13
15
Tp B = (c + 1)Tp ∆f
4
12.5
9
24.5
16
12.5
36
24.5
40.5
60.5
84.5
112.5
B
∆f
=c+1
2
5
3
7
4
2.5
6
3.5
4.5
5.5
6.5
7.5
carried out for each generation according the procedure given in Section 5.4. Figure
5.6 illustrates Pareto front obtained for Tp ∆f ∈ [2, 10], c ∈ [2, 10] and N = 8. This
Pareto front provides the trade-off solutions between grating lobe and the PSR. It
is evident from Figure 5.6 that for all the solutions the sidelobes are below 30 dB
as compared to their respective mainlobes. All the solutions in the Pareto front are
nondominant and a particular solution from the front is chosen according to the
requirements of the application such as low sidelobe level or low grating lobes. For
different values of Tp ∆f and Tp B corresponding values of |R1 (τ )|, |R2 (τ )| and ACF
are shown in Figures. 5.7 to 5.9. In Figure 5.7 the nulls of |R1 (τ )| exactly falls
on the grating lobes of |R2 (τ )| which means that all the grating lobes are canceled
i.e. f2 = 0. In Figure 5.9, the maximum grating lobe amplitude is 0.021 which
is prominently observed around τ /Tp = 0.7 in ACF. But the sidelobes occurring
in Figure 5.7 are below 30.7753 dB from its mainlobe and that of in Figure 5.9 is
32.5 dB below its mainlobe. Therefore the parameter values of stepped frequency
LFM pulse train are chosen according to the requirement of application. If it is
required to suppress all the grating lobes below a certain value ǫ, then a constraint
i.e. f1 < ǫ is associated with the optimization process. Different Pareto fronts can
119
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
Chapter 5
−30.5
PSR in dB (f3)
−31
−31.5
−32
−32.5
−33
0
0.005
0.01
0.015
maximum grating lobe amplitude (f )
0.02
0.025
2
Figure 5.6: Pareto front obtained using NSGA-II for Tp ∆f ∈ [2, 10], c ∈ [2, 10] and
N =8
be obtained by varying the upper and lower limits of Tp ∆f and c depending upon
their ranges for that particular application. For Tp ∆f ∈ [2, 10], c ∈ [2, 10], ǫ = 0.01,
N = 8 and Tp ∆f ∈ [5, 30], c ∈ [2, 10], ǫ = 0.01, N = 8 the Pareto fronts are
depicted in Figures 5.10 and 5.11 respectively. The NSGA-II algorithm facilitates
for choosing the parameters from a set of available optimal parameters according
to the requirements of the system under consideration. Small grating lobes or low
sidelobes can be achieved by choosing an appropriate solution from Pareto front.
Suitable overlap ratio, i.e.
B
∆f
= c + 1, is accomplished by properly defining lower
and higher limit of c during population initialization.
For the third problem the population size and number of generations are chosen
to be 200 and 50 for optimizing the values of f3 and f4 . Same set of previously
chosen parameters of NSGA-II algorithm is used in this case. The population is
randomly initialized for two parameters Tp ∆f and c for the given lower and upper
limit. The initialized population is sorted according to the nondomination sorting
and the process of selection, crossover, mutation and recombination are carried out
for each generation according to the procedure laid down in Section 5.5.2. For
Tp ∆f ∈ [2, 10], c ∈ [2, 5], ǫ = 0.01 and N = 8 the Pareto front is shown in the Figure
5.12. This Pareto front provides a trade-off between the grating lobe and mainlobe
width. A solution from the Pareto front is chosen according to the requirements of
120
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
Chapter 5
0.8
0.6
0.4
1
2
|R (τ)|,|R (τ)|
1
0.2
0
0
0.1
0.2
0.3
0.4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.5
0.6
0.7
0.8
0.9
1
ACF in dB
0
−20
−40
−60
τ/T
p
Figure 5.7: Stepped frequency LFM pulse for Tp ∆f = 2, c = 5, Tp B = 12 and
N = 8. Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom: ACF (in dB)
0.8
0.6
0.4
1
2
| R (τ), R (τ) |
1
0.2
0
0
0.1
0.2
0.3
0.4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.5
0.6
0.7
0.8
0.9
1
ACF in dB
0
−20
−40
−60
τ / Tp
Figure 5.8: Stepped frequency LFM pulse for Tp ∆f = 2, c = 5.1412, Tp B = 12.2824
and N = 8. Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom: ACF (in dB)
121
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
Chapter 5
0.8
2
| R (τ) |, |R (τ) |
1
0.6
1
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.5
0.6
0.7
0.8
0.9
1
ACF in dB
0
−20
−40
−60
τ/T
p
Figure 5.9: Stepped frequency LFM pulse for Tp ∆f = 2.8721, c = 5.0978, Tp B =
17.5135 and N = 8. Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom: ACF (in dB)
−30.6
−31
3
PSR in dB (f )
−30.8
−31.2
−31.4
−31.6
−31.8
0
0.001
0.002
0.003
0.004
0.005
0.006
0.007
0.008
0.009
0.01
Maximum grating lobe amplitude (f )
2
Figure 5.10: Pareto front obtained using NSGA-II for Tp ∆f ∈ [2, 10] c ∈ [2, 10],
ǫ = 0.01 and N = 8
122
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
Chapter 5
−28.4
−28.45
PSR in dB(f3)
−28.5
−28.55
−28.6
−28.65
−28.7
−28.75
−28.8
−28.85
0
0.001
0.002
0.003
0.004
0.005
0.006
0.007
0.008
0.009
0.01
Maximum gratiing lobe amplitude (f )
2
Figure 5.11: Pareto front obtained using NSGA-II for Tp ∆f ∈ [5, 30], c ∈ [2, 10],
ǫ = 0.01 and N = 8
−22
−23
−24
−25
f
3
−26
−27
−28
−29
−30
−31
−32
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
f4
Figure 5.12: Pareto front obtained using NSGA-II for Tp ∆f ∈ [2, 10], c ∈ [2, 5],
ǫ = 0.01 and N = 8
application under consideration. If the application demands high range resolution,
a solution having low value of f4 is chosen and if it requires low sidelobe level, a
solution corresponds to high magnitude of f3 is chosen. Figures 5.13 to 5.16 show
|R1 (τ )|, |R2 (τ )| and ACF for different values of Tp ∆f , c and ǫ = 0.01. It is evident
from the figures that reduction of peak sidelobe is achieved at the cost of increase
in mainlobe width or reduction in range resolution.
123
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
Chapter 5
0.8
2
| R (τ) |,| R (τ) |
1
0.6
1
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.5
0.6
0.7
0.8
0.9
1
ACF in dB
0
−20
−40
−60
τ / Tp
Figure 5.13: Stepped frequency LFM pulse for Tp ∆f = 9.0188, c = 3.5502, Tp B =
41.0373 and N = 8. Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom: ACF (in dB)
0.8
2
|R (τ)|, |R (τ)|
1
0.6
1
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.5
0.6
0.7
0.8
0.9
1
ACF in dB
0
−20
−40
−60
τ / Tp
Figure 5.14: Stepped frequency LFM pulse for Tp ∆f = 4.9667, c = 4.0720, Tp B =
25.1911 and N = 8. Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom: ACF (in dB)
124
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
Chapter 5
0.8
2
| R (τ) |, | R (τ) |
1
0.6
1
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.5
0.6
0.7
0.8
0.9
1
ACF in dB
0
−20
−40
−60
τ / Tp
Figure 5.15: Stepped frequency LFM pulse for Tp ∆f = 3.6048, c = 4.6129, Tp B =
20.2334 and N = 8. Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom: ACF (in dB)
0.8
1
2
| R (τ) |, | R (τ) |
1
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.5
0.6
0.7
0.8
0.9
1
ACF in dB
0
−20
−40
−60
τ/T
p
Figure 5.16: Stepped frequency LFM pulse for Tp ∆f = 3, c = 5, Tp B = 18 and
N = 8. Top: |R1 (τ )| (dash) and |R2 (τ )| (solid). Bottom: ACF (in dB)
125
Chapter 5
5.7
Efficient Design of Stepped Frequency Pulse Train Using
Evolutionary Computation Techniques
Conclusion
Frequency stepping technique is mostly used in radar technology to achieve high
range resolution by combining the effect of narrowband pulses that span in the
desired bandwidth. The drawback of such type of waveform is the presence of the
grating lobes. In this chapter the PSO algorithm is used to determine the parameters
of LFM pulse train for which all the grating lobes are nullified. Using the PSO various
combinations of Tp ∆f and Tp B are found for which all the grating lobes are nullified
and listed in Table 5.1. Changing the limits of Tp ∆f and c more combination of
Tp ∆f and Tp B can be found for which all the grating lobes are nullified.
The multiobjective NSGA-II algorithm has been applied to determine the
parameters of stepped frequency pulse train to get reduced grating lobes, low
sidelobes and narrow mainlobe width at the matched filter output of the pulse train.
The multiobjective optimization algorithm enables to provide trade off solutions
between different objectives through Pareto front that contains nondomination
solutions. In the proposed work the multiobjective problem has been formulated
in two different manners. One formulation provides the trade-off solutions between
grating lobes and peak sidelobe and the other provides the trade-off solutions
between peak sidelobe and mainlobe width.
126
Chapter 6
Conclusion and Future Work
6.1
Conclusion
In this chapter, the conclusion of the whole thesis is presented and future research
problems are outlined for further investigation in the same or related topics. In
this thesis investigation has been made on developing efficient pulse compression
techniques for phase and frequency modulated waveforms. The main contribution
of the thesis is the use of neural network structures and evolutionary computation
techniques for pulse compression.
Biphase codes of longer sequences having low PSL and high MF are important
research area in the field of radar signal processing. There is no available technique
to generate a certain length code for a given PSL and/or MF. In this thesis a
multiobjective algorithm (NSGA-II) is presented to generate the biphase codes of
length 49 to 59 using PSL and MF as two different objective functions. The use
of NSGA-II algorithm has provided more than one nondominated solutions and a
particular code is to be selected depending upon specific situation. This algorithm
in general can be applied to generate codes of any length.
Mismatch filters are used to provide better PSR than that of matched filter of
a given sequence. Several ANN based mismatch filters are used to achieve reduced
sidelobes for 13-bit and 35-bit Barker codes at the output of the filter. In this work,
the RNN and RRBF structures are proposed to use as a pulse compression filter
127
Conclusion and Future Work
Chapter 6
which have been provided better performance in terms of PSR under various adverse
conditions such as noise, multiple target and Doppler shift. The performance of the
proposed methods is compared to that of MLP and RBF based pulse compression
techniques. The comparison study reveals that the RRBF based pulse compression
technique performs the best among others.
Biphase codes can be easily generated and their matched filters implementation
are also simple. These advantages are achieved at the cost of higher sidelobes.
Polyphase codes have lower sidelobes compared to biphase codes for a certain length
code. The LFM signals are more Doppler tolerant than phase coded signals. Hence,
phase codes such as Frank, P1 , P2 , P3 and P4 codes are derived from the LFM signals
to get the advantages of the Doppler shift performance of LFM signal. The PSRs
offered by LFM and polyphase codes are not adequate for many radar applications.
Various windows are used to suppress the sidelobe of LFM and polyphase codes. In
this thesis the use of convolutional windows are proposed as the weighing function
at the receiver to get better PSR values at higher Doppler shifts compared to that
of conventional windows. Amplitude tapering and phase distortion techniques are
employed to modify the transmitted LFM signal and the convolutional window
is used as the weighing function at the receiver offers better PSR values that of
conventional windows.
In high range resolution radar the bandwidth of the signal should be large
to achieve narrow mainlobe width. To overcome the difficulty of generation and
processing, the wideband signal is split into a number of narrowband signals which
together called as stepped frequency waveform. The ACF of stepped frequency
LFM pulse train suffers from grating lobes for Tp ∆f > 1. Hence the range resolution
capability of the waveform is reduced. A PSO algorithm based technique is proposed
to choose the parameters of LFM pulse train to nullify or suppress the grating
lobes. Further, a widely used multiobjective NSGA-II algorithm based approach
is proposed to determine the optimum parameters of LFM pulse train to achieve
reduced grating lobes, low sidelobes and narrow mainlobe width. The multiobjective
128
Conclusion and Future Work
Chapter 6
problem has been formulated in two different ways. One formulation provides the
trade-off solutions between grating lobes and peak sidelobe and the other provides
the trade-off between peak sidelobe and mainlobe width.
6.2
Future work
The research work presented in the thesis can be further extended in following ways.
• Genetic operators such as crossover and mutation are important operations in
GA and NSGA-II. Better biphase codes can be designed by employing different
variant of these operators.
• Better mismatch filters can be developed using polynomial neural network and
support vector machine for sidelobe suppression and the performance can be
compared with existing methods.
• Time-frequency analysis such as short time Fourier transform, wavelet
transform and S-transform can be used for LFM signals to extract Doppler
information.
• The LFM signal can be replaced by Doppler tolerant hyperbolic frequency
modulated pulse in the pulse train and the multiobjective algorithms can be
employed to enhance sidelobe suppression.
129
Bibliography
[1] M. I. Skolnik, Introduction to radar system. New York: McGraw-Hill, 1980.
[2] C. E. Cook and M. Bernfeld, Radar signals: An introduction to theory and
application. Academic Press, New York, 1967.
[3] D.K. Barton, Pulse Compression. Artech House, 1975.
[4] A.W. Rihaczek, Principle of high resolution radar. McGraw Hill, New York,
1969.
[5] N. Levanon and E. Mozeson, Radar Signals. Wiley New York, 2004.
[6] P. M. Woodward, Probability and information theory, with application to radar.
Pergamon Press, Oxford, 1953.
[7] A.M. Boehmer, “Binary pulse compression codes,” IEEE Trans. on Inf. Theory,
vol. 13, no. 2, pp. 156-167, Apr 1967.
[8] R.H. Barker, “Group Synchronization of Binary Digital Systems,” Commun.
Theory, pp. 273-287, 1953.
[9] M. I. Skolnik, Radar Handbook. New York: McGraw-Hill (2nd ed.), 1990.
[10] J. Lindner, “Binary sequences up to length 40 with best possible autocorrelation
function,” Electron. Lett., vol. 11, no. 21, p. 507, Oct. 1975.
[11] M. N. Cohen, M. R. Fox, and J. M. Baden, “Minimum peak sidelobe pulse
compression codes,” in Proc. IEEE Int. Radar Conf., pp. 633-639, 1990.
[12] F. Hu, P. Z. Fan, M. Darnell and F. Jin, “Binary sequences with good aperiodic
autocorrelation functions obtained by neural network search,” Electron. Lett., vol.
33, no. 8, pp. 688-690, Apr. 1997.
130
Bibliography
[13] S. Wang, “Efficient heuristic method of search for binary sequences with good
aperiodic autocorrelations,” Electron. Lett., vol. 44, no. 12, pp. 731-732, June
2008.
[14] B. Militzer, M. Zamparelli and D. Beule, “Evolutionary search for low
autocorrelated binary sequences,” IEEE Trans. on Evol. Comput., vol. 2, no.
1, pp. 34-39, Apr. 1998.
[15] X. Deng and P. Fan, “New binary sequences with good aperiodic
autocorrelations obtained by evolutionary algorithm,” IEEE Comm. Lett., vol.
3, no. 10, pp. 288-290, Oct. 1999.
[16] K. R. Rajeswari and N. Gangatharan, “Algorithm for design of pulse
compression radar codes,” Electron. lett., vol. 39, no. 11, pp. 865-867, May 2003.
[17] M.J.E. Golay, “The merit factor of long low autocorrelation binary sequences,”
IEEE Trans. on Inf. Theory, vol. IT-28, no. 3, pp. 543-549, May 1982.
[18] Holland J.H., Adaptation in Natural and Artificial Systems., University of
Michigan Press, Ann Arbor, 1975.
[19] D.E.Goldberg, Genetic algorithms in search, optimization and machine
learning. Addition-Wesley, 1989.
[20] D. E. Goldberg and K. Deb, “A comparative analysis of selection schemes used
in GAs,” Foundations of genetic Algorithms, San Mateo, CA: Morgan Kaufmann,
pp. 69-93, 1991.
[21] A. Konak, D.W. Coit and A.E. Smith, “Multi-objective optimization using
genetic algorithms: a tutorial,” Reliability Engg. and System Safety, pp. 992-1007,
2006.
[22] J. D. Schaffer, “Multiple objective optimization with vector evaluated genetic
algorithms,” In Proc.of the first Int. Conf. on Genetic Algorithms, Lawrence
Erlbaum, pp. 93-100, 1985.
[23] C. M. Fonseca and P. J. Fleming, “Genetic algorithms for multiobjective
optimization: formulation, discussion and generalization,” In Proc. of the fifth
Int. Conf. on Genetic Algorithms, pp. 416-423, 1993.
131
Bibliography
[24] J. Horn, N. Nafpliotis and D. Goldberg, “A niched Pareto genetic algorithm for
multiobjective optimization,” In Proc. of the first IEEE Conf. on Evol. Comput.,
pp. 82-87, 1994.
[25] N Srinivas and K. Deb, “Muiltiobjective optimization using nondominated
sorting in genetic algorithm,” Evol. Comput., vol. 2, no. 3, pp. 221-48, 1994.
[26] K. Deb, A. Pratap, S. Agarwal and T. Meyarivan, “A fast and elitist
multiobjective genetic algorithm: NSGA-II,” IEEE Trans. Evol. Comput., vol.
6, no. 2, pp. 182-197, 2002.
[27] N. J. Bucci, H. S. Owen, K. A.Woodward and C. M. Hawes, “Validation of
pulse compression techniques for meteorological functions,” IEEE Trans. Geosci.
Remote Sensing, vol. 35, no. 3, pp. 507-523, May 1997.
[28] A. S. Mudukutore, V. Chandrasekar and R. J. Keeler, “Pulse compression for
weather radars,” IEEE Trans. Geosci. Remote Sensing, vol. 36, no. 1, pp. 125-142,
Jan. 1998.
[29] M.H. Ackroyd, “Synthesis of efficient Huffman sequences,” IEEE Trans. Aerosp.
Electron. Syst., vol. AES-8, no. 1, pp. 2-8, Jan. 1972.
[30] A.W. Rihaczek and R.M. Golden, “Range sidelobe suppression for Barker
codes,” IEEE Trans. Aerosp. Electron. Syst., vol. AES-7, no. 6, pp. 1087-1092,
Nov. 1971.
[31] M.H. Ackroyd and F. Ghani, “Optimum mismatched filter for sidelobe
suppression,” IEEE Trans. Aerosp. Electron. Syst., vol. AES-9, no. 2, pp. 214-218,
Mar. 1973.
[32] S. Zoraster, “Minimum peak range sidelobe filters for binary phase coded
waveforms,” IEEE Trans. Aerosp. Electron. Syst., vol. AES-16, no. 1, pp. 112-115,
Jan. 1980.
[33] C.X Hua and J Oksman, “A new algorithm to optimize Barker code sidelobe
suppression filters,” IEEE Trans. Aerosp. Electron. Syst., vol. AES-26, no.4, pp.
673-677, July 1990.
[34] I. Sarkar and A.T. Fam, “Multiplicative mismatched filters for sidelobe
suppression in Barker codes,” IEEE Trans. Aerosp. Electron. Syst., vol. AES-44,
no.1, pp. 349-359, Jan. 2008.
132
Bibliography
[35] B. M. Zrnic, A. J. Zejak, and Petrovic and J. Oksman, “Pulse compression
radar:Self clutter suppression using modified RLS algorithm,” Int. Conf. on
Telecommunications in Modern Satelite, Cable and Broadcasting Services, vol.
2, pp. 363 - 366, 1999.
[36] J.S.Fu and and Xin Wu, “Sidelobe suppression using adaptive filtering
techniques,” Proc. of CIE Int. Conf. on Radar, pp. 788-791, 2001.
[37] H.K. Kwan and C.K. Lee, “Pulse radar detection using a multilayer neural
network,” Int. Joint Conf. on Neural Netw., vol. 2, pp. 75-80, 1989.
[38] H.K. Kwan and C.K. Lee, “A neural network approach to pulse radar
detection,” IEEE Trans. Aerosp. Electron. Syst., vol. AES-29, no. 1, pp. 9-21,
Jan. 1993.
[39] K.D. Rao and G. Sridhar, “Improving performance in pulse radar detection
using neural networks,” IEEE Trans. Aerosp. Electron. Syst., vol. AES-31, no. 3,
pp. 9-21, July 1995.
[40] F. B. Duh, C.F. Juang and C.T. Lin,“ A neural fuzzy network approach to
radar pulse compression,” IEEE Geosci. and remote sensing Lett., vol. 1, no. 1,
pp. 15-20, Jan 2004.
[41] D.G. Khairnar, S.N. Merchant and U.B. Desai, “Radial basis neural network
approach to pulse radar detection,” IET Radar Sonar Navig., vol. 1, no. 1, pp.
8-17, 2007.
[42] B. Widrow and S.D. Sterns, “Adaptive signal processing,” Prentice-Hall, Inc.
Englewood Cliffs, New Jersey, 1985.
[43] S. Haykin, Adaptive Filter Theory. 4th edition, Pearson Education Asia, 2002.
[44] S. Haykin, Neural Networks: A comprehensive foundation. Second Edition,
Pearson Education Asia, 2002.
[45] Y. H. Pao, Adaptive pattern recognition and neural networks. Addison Wesley,
Reading, Massachusetts, 1989.
[46] D. H. Nguyen and B. Widrow, Neural networks for self-learning control system,
Int. J. Contr., vol. 54, no. 6, pp. 1439-1451, 1991.
133
Bibliography
[47] E. J. Dayhoff, Neural network architecture-An introduction. Van Norstand
Reilold, New York, 1990.
[48] P. S. Sastry, G. Santharam and K. P. Unnikrishnan, “Memory neural networks
for identification and control of dynamical systems, IEEE Trans. Neural Netw.,
vol. 5, pp. 306-319, 1994.
[49] N. K. Bose and P. Liang, Neural Network Fundamentals with graphs, algorithms,
applications. TMH Publishing Company Ltd., 1998.
[50] Richard O. Duda, Peter E. Hart and David G. Stork, Pattern classification.
2nd edition, John Wiley & Sons, INC. 2001.
[51] R. K. Seyab, and Y.Cao, “Nonlinear system identification for predictive control
using continuous time recurrent neural networks and automatic differentiation,”
Journal of Process Control, vol. 18, no. 6, pp. 568-581, July 2008.
[52] J. Draye, D. Pavisic, and G. Libert, “Dynamic recurrent neural networks: a
dynamical analysis,” IEEE Trans. Syst., Man and Cybern., part B, vol. 26, no. 5,
pp. 692-706, 1996.
[53] E. Kosmatoupoulos, M. Polycarpou and A.Iannou, “High-order neural network
structures for identification of dynamical systems,” IEEE Trans. Neural Netw.,
vol. 6, no. 2, pp. 422-431, 1995.
[54] S. Kumpati, F. Narenda and K. Parthasarathy, “Identification and control of
dynamical systems using neural networks,” IEEE Trans. of Neural Netw., vol. 1,
no. 1, pp. 4-24, 1990.
[55] A. Parlos, K. Chong and A. Atiya, “Application of the recurrent multilayer
perceptron in modeling complex process dynamics,” IEEE Trans. on Neural
Netw., vol. 5, no. 2, pp. 255-266, 1994.
[56] X. M. Ren, A. B.Rad, P. T. Chan and W. L.Lo “Identification and control of
continuous-time nonlinear systems via dynamic neural networks”, IEEE Trans.
Industrial Electron., vol. 50, no. 3, pp. 478-786, 2003.
[57] H. T. Su and T. J. McAvoy, Artificial neural networks for nonlinear process
identification and control. Nonlinear Process Control, Prentice-Hall, New Jersey,
pp.371-428, 1997.
134
Bibliography
[58] D. Linkens and Y.Nyongesa, “Learning systems in intelligent control: an
appraisal of fuzzy, neural and genetic algorithm control applications,” IEE Proc.
of Contr. Theory and Applications, vol. 134, no. 4, pp. 367-385, 1996.
[59] S. F. Toha and M. O. Tokhi, “MLP and Elman recurrent neural network
modeling for TRMs,” IEEE Int. Conf. on Cybern. Intelligent Syst., pp. 1-6, Sept.
2008.
[60] S. V. T. Elanayar and Y. C. Shin, “Radial basis function neural network for
approximation and estimation of nonlinear stochastic dynamic systems,” IEEE
Trans. Neural Network, vol. 5, pp. 594-603, 1994.
[61] S. R. Samantray, P. K. Dash and G. Panda , “Fault classification and location
using HS-transform and radial basis function neural network,” Electric Power
Syst. Research, vol. 76, pp. 897-905, 2006.
[62] Y. J. Oyang, S. C. Hwang, Yu-Yen Ou, C.Y. Chen and Z.W. Chen, “Data
classification with radial basis function networks based on a novel kernel density
estimation algorithm,” IEEE Trans. Neural Netw., vol. 16, no.1, pp. 225-236,
2005.
[63] A. C. Tsoi and S. Tan, “Recurrent neural networks: A constructive algorithm
and its properties,” Neurocomputing, vol. 15, pp. 309-326, 1997.
[64] G. Hardier, “Recurrent RBF networks for suspension system modeling and wear
diagnosis of a damper,” IEEE World Congress on Comput. Intelligence, vol. 3,
pp. 2441 - 2446, 1998.
[65] S.A. Billings and C.F. Fung, “Recurrent Radial Basis Function Networks for
Adaptive Noise Cancellation,” Neural Netw., vol. 8, no. 2, pp. 273-290, 1995.
[66] R. Bambang, “Active noise cancellation using recurrent radial basis function
neural networks,” Asia-Pacific Conf. on Circuits and Syst., vol 2, pp. 231-236A,
2002.
[67] R. Zemouri, D. Racoceanu and N. Zerhouni, “Recurrent radial basis function
network for time-series prediction,” Engineering Applications of Artificial
Intelligence, vol. 16, no. 5, pp. 453-463, 2003.
135
Bibliography
[68] J Mazurek, A. Krzyzak and A.Cichocki, “Rates of convergence of the recursive
radial basis function networks,” IEEE Int. Conf. on Acoustics, Speech, and Signal
Processing (ICASSP), vol. 4, pp. 3317-3320, 1997.
[69] M. Kowatsch and H.R. Stocker, “Effect of Fresnel ripples on sidelobe
suppression in low time-bandwidth product linear FM pulse compression,” Proc.of
the IEE, vol. 129, no. 1, pp. 41-44, Feb. 1982.
[70] H.D. Griffiths and L. Vinagre, “Design of low-sidelobe pulse compression
waveforms,” Electron. Lett., vol. 30, no. 12, pp. 1004-1005, June 1994.
[71] E. D Witte and H.D. Griffiths, “Improved ultra-low range sidelobe pulse
compression waveform design,” Electron. Lett., vol. 40, no. 22, pp. 1448-1450,
Oct. 2004.
[72] C.E. Cook and J. Paolillo, “A pulse compression predistortion function for
efficient sidelobe reduction in a high power radar,” Proc. of the IEEE, vol. 52, no.
4, pp. 377-389, Apr. 1964.
[73] K.M. El-Sheennawy, O.A. Alim and M.A. Ezz-El-Arab, “Sidelobe suppression
in low and high time-bandwidth products of linear FM pulse compression filters,”
IEEE Trans. on Microwave Theory and Techniques, vol. 35, no. 9, pp. 807-811,
Sept. 1987.
[74] J.J.G. Mccue, “A Note on the Hamming weighting of linear FM pulses,” Proc.
of the IEE, vol. 67, no. 11, pp. 1575-1577, Nov. 1979.
[75] J.A. Johnston and A.C. Fairhead, “Waveform design and Doppler sensitivity
analysis for nonlinear FM chirp pulses,” IEE Proc., vol. 133, no. 2, Apr. 1986.
[76] T. Collins and P. Atkins, “Nonlinear frequency modulation chirps for active
sonar,” IEEE Proc. Radar, sonar navig., vol. 146, no. 6, pp. 312-316, Dec. 1999.
[77] P. Yichun, P. Shirui, Y. Kefeng and D. Wenfeng, “Optimization design of NLFM
signal and its pulse compression simulation,” IEEE Int. Radar Conf., pp. 383-386,
May 2005.
[78] B. Getz and N. Levanon, “Comparison between linear FM and phase-code CW
radars,” IEE Proc. Radar, Sonar Navig., vo1. 141, no. 4, pp. 230-240, Aug. 1994.
136
Bibliography
[79] R. L Frank, “Polyphase codes with good nonperiodic correlation properties,”
IEEE Trans. on Inf. Theory, Vol. IT-9, pp. 43-45, 1963.
[80] B. L. Lewis and F. F. Kretschmer Jr., “A new class of polyphase pulse
compression codes and techniques,” IEEE Trans. Aerosp. Electron. Syst., vol.
AES-17, no. 3, pp. 364-372, May 1981.
[81] B. L. Lewis and F. F. Kretschmer Jr., “Linear frequency modulation derived
polyphase pulse compression codes,” IEEE Trans. Aerosp. Electron. Syst., vol.
AES-18, no. 5, pp. 637-641, Sep. 1982.
[82] F. F. Kretschmer and L.R. Welch, “Sidelobe reduction techniques for polyphase
pulse compression codes,” IEEE Int. Radar Conf., pp. 416-421, 2000.
[83] B. Yang, J.Xiang and S. Liu, “A technique for designing sidelobe reduction
digital filter for pulse compression,” Journal of Electronics, vol.22, no.1, pp.
124-129, 2000.
[84] B. L. Lewis, ”Range-time-sidelobe reduction technique for FM-derived
polyphase PC Codes,” IEEE Trans. Aerosp. Electron. Syst., vol. AES-29, no.
3, pp. 834-840, July 1993.
[85] F. F Kretschmer Jr. and B. L. Lewis, “Doppler properties of polyphase pulse
compression waveforms,” IEEE Trans. Aerosp. Electron. Syst., vol. 19, no. 4, pp.
521-531, July 1983.
[86] W.K.Lee and H.D.Griffiths “Pulse compression filters generating optimal
uniform range sidelobe level,” Electron. Lett., vol. 35, no. 11, pp. 873-875, 1999.
[87] W. K. Lee and H. D. Griffiths, “A new pulse compression technique generating
optimal uniform range sidelobe and reducing integrated sidelobe level,” IEEE Int.
Radar Conf., pp. 441-446, 2000.
[88] W.K. Lee and H.D.Griffiths, “Development of modified polyphase P codes with
optimum sidelobe characteristics,” IEEE Proc. Radar, Sonar and Navig., vol. 151,
no. 4, pp. 210-220, Aug. 2004
[89] Woo-Kyung Lee, “A pair of asymmetrical weighting receivers and polyphase
codes for efficient aperiodic correlations,” IEEE Commun. Lett. vol. 10, no. 5, pp.
387-389, May 2006.
137
Bibliography
[90] U. Kumari, K. Rajearakeswari and M. Krishna, “Low sidelobe Pattern using
Woo filter,” Int. J. Electron. Commun.(AEU), vol. 59, no. 8, pp. 499-501, Dec.
2005.
[91] A. Antoniou, Digital Signal Processing:
McGraw-Hill, 2005.
Signal, Systems, and Filters.
[92] F.J. Harris, “On the use of windows for harmonic analysis with the discrete
Fourier transform,” Proc. of the IEEE, vol. 66, no. 1, pp. 51-83, Jan. 1978.
[93] G.C. Temes, V. Barcilon and F.C.I. Marshall, “The optimization of bandlimited
systems,” Proc. of IEEE, vol. 61, no. 2, pp. 196-234, Feb. 1973.
[94] N.C. Gec-kinli and D. Yavuz, “Some novel windows and a concise tutorial
comparison of window families,” IEEE Trans. Acoust. Speech Signal Process., vol.
26, no. 6, pp. 501-507, Dec. 1978.
[95] A.H. Nuttal, “Some windows with very good sidelobe behavior,” IEEE Trans.
Acoust. Speech Signal Process., vol. 29, no. 1, pp. 84-91, Feb. 1981.
[96] J.W. Adams, “A new optimal window,” IEEE Trans. Signal Process., vol. 39,
no. 8, pp. 1753-1769, Aug. 1991.
[97] S. Yang and Y. Ke, “On the three-coefficient window family,” IEEE Trans.
Signal Process., vol. 40, no. 12, pp. 3085-3088, Dec. 1992.
[98] M. Szyper, “New time domain window,” Electron. Lett., vol. 31, no. 9, pp.
707-708, Apr. 1995.
[99] J.K. Gautam, A. Kumar and R. Saxena, “On the modified Bartlett-Hanning
window (family),” IEEE Trans. Signal Process., vol. 44, no. 8, pp. 2098-2102,
Aug. 1996.
[100] S.H. Saeid and J.K. Gautam, “New hyperbolic function window family,”
Electron. Lett. vol. 33, no. 8 pp. 1531-1532, Aug. 1997.
[101] CL. Dolph, “A current distribution for road side arrays which optimizes the
relationship between beam width and sidelobe level,” Proc. IRE, pp. 335-348,
1946.
138
Bibliography
[102] JF. Kaiser and RW. Schafer., “On the use of the I0 −sinh window for spectrum
analysis,” IEEE Trans. Acoust. Speech Signal Process., vol. 28, no. 1, pp. 105-107,
1980.
[103] K. Okarma, “Polynomial windows with low sidelobes level,” Signal Process.,
vol. 87, no. 4, pp. 782-788, Apr. 2007.
[104] K. Avci and A. Nacaroglu, “Cosh window family and its application to FIR
filter design,” Electron. and Commun., vol. 63, no. 11, pp. 907-916, Nov. 2009.
[105] I. Reljin, B. Reljin, V. Papic and P. Kostic , “New window functions generated
by means of time convolution-spectral leakage error,” Electrotechnical Conf.,
MELECON, vol. 2, pp. 878-881, May 1998.
[106] J. Zhang , L. Changhong and Y. Chen, “A new family windows-convolution
windows and their applications,” Sci China Ser E-Tech Sci, vol. 48, no. 4, pp.
468-480, 2005.
[107] Y. Gao, Z. Teng, J. Wang and H. Wen, “Dielectric loss factor measurement on
Nuttal self- convolution window phase difference correction,” The Ninth Int.Conf.
on Electronic Measurement and Instruments(ICEMI), pp. 714-719, 2009.
[108] W. He , T. ZhaoSheng , G. SiYu , W. JingXun , Y. BuMing and W. Yi and C.
Tao, “Hanning self-convolution window and its application to harmonic analysis,”
Science in China, Series E: Technological Sciences, vol. 52, no. 2, pp. 467-476,
2009.
[109] D. R. Wehner, High resolution radar. Artech House, Boston, 1995.
[110] D. E. Maron, “Non-periodic frequncy-jumped burst waveforms,” In Proc. of
the IEE Int. Radar Conf., London, pp. 484-488, Oct. 1987.
[111] D. E. Maron, “Frequncy-jumped burst waveforms with stretch processing,” In
Proc. of the IEEE Int. Radar Conf., Arlington, pp. 274-279, 1990.
[112] D. J. Rabideau, “Nonlinear synthetic wideband waveforms,” In Proc. of the
IEEE Int. Radar Conf., Los Angeles, pp. 212-219, 2002.
[113] N. Levanon and E. Mozeson, “Nullifying ACF grating lobes in
stepped-frequency train of LFM pulses,” IEEE Trans. Aerosp. Electron. Syst.,
vol. AES-39, no. 2, pp. 694-703, 2003.
139
Bibliography
[114] I. Gladkova and D. Chebanov, “Suppression of grating lobes in stepped
frequency train,” In Proc. of the IEEE Int. Radar Conf., Virginia, USA, pp.
371-376, 2005.
[115] I. Gladkova and D. Chebanov, “Grating lobes suppression in stepped frequency
train,” IEEE Trans. Aerosp. Electron. Syst., vol. AES-44, no. 4, pp. 1265-1275,
Oct. 2008.
[116] J. Kennedy and R. Eberhart, “Particle swarm optimization,” in Proc.of IEEE
Int. Conf. Neural Netw., pp. 1942-1948, 1995.
[117] M. Clerc, “The swarm and the queen: Toward a deterministic and adaptive
particle swarm optimization,” in Proc. IEEE Int. Congr. Evol. Comput., vol. 3,
pp. 1951-1957, 1999.
[118] M. Clerc and J.Kennedy, “The particle swarm-explosion, stability, and
convergence in a multi-dimensional complex space”, IEEE Trans. Evol. Comput.,
vol.6, pp. 58-73, Feb. 2002.
[119] Y. Shi and R. C. Eberhart, “Parameter selection in particle swarm
optimization,” Evol. Programming, vol. VII, Eds. Berlin, Germany:
Springer-Verlag, pp. 591-600, 1998.
[120] Y. Shi and R. C. Eberhart, “A modified particle swarm optimizer,” Proc. IEEE
Conf. Evol. Comput., Anchorage, AK, pp. 69-73, 1998.
[121] H. Beyer and K. Deb, “On self-adaptive features in real-parameter evolutionary
algorithm,” IEEE Tran. Evol. Comput., vol. 5, no. 3, pp. 250-270, 2001.
[122] K. Deb and R. B. Agarwal, “Simulated binary crossover for continuous search
space,” Complex Syst., vol. 9, pp. 115-148, 1995.
[123] M. M. Raghuwanshi and O. G. Kakde, “Survey on multiobjective evolutionary
and real coded genetic algorithms,” In Proc. of the 8th Asia Pacific Symposium
on Intelligent and Evol. Syst., pp. 150-161, 2004.
140
Dissemination of Work
Journals
1. A. K. Sahoo and G. Panda, “A Multiobjective Optimization Approach to
Determine the Parameters of Stepped Frequency Pulse Train,” Aerospace
Science and Technology-Elseveir, DOI 10.1016/j.ast.2011.10.008.(Accepted
17-Oct.-2011)
2. A. K. Sahoo and G. Panda “Sidelobe Suppression Using Convolutional
Windows in Radar,” Int. J. Signal and Imaging Systems Engineering
(IJSISE), Inderscience Publisher.(Accepted 25-Mar-2011)
3. A. K. Sahoo and G. Panda, “Doppler Tolerant Convolutional Windows
for Radar Pulse Compression,” Int. J. of Electronics and Communication
Engineering, International Research Publication, vol. 4, no.1, pp.145-152,
2011.
4. A. K. Sahoo and G. Panda, “Suppression of grating lobes in
stepped frequency LFM pulse train using PSO,” Aerospace Science and
Technology-Elseveir.(Communicated)
Conferences
1. A. K. Sahoo G. Panda and P.M. Pradhan, “Generation of
pulse compression codes using NSGA-II,
Annual IEEE India
Conference (INDICON),Ahamadabad, pp.1-4, Dec.
18-20, 2009.DOI.
10.1109/INDCON.2009.5409443.
2. A. Sailaja, A. K. Sahoo, G. Panda and V. Baghel, “A recurrent
neural network approach to pulse radar detection,” Annual IEEE India
Conf. (INDICON), Ahamadabad, pp. 1-4, Dec. 18-20, 2009. DOI.
10.1109/INDCON.2009.5409446.
141
Dissemination of Work
3. A. K. Sahoo and G. Panda, “Sidelobe Reduction of LFM Signal Using
Convolutional Windows,” Int. Conf. on Electronics Systems (ICES), NIT
Rourkela, pp. 86-89, Jan. 9-11, 2011.
142
Resume
Ajit Kumar Sahoo
Assistant Professor
Department of Electronics & Communication Engineering
National Institute of Technology Rourkela
Rourkela, Orissa – 769 008, India.
Ph: +91-9861370334(M)
E-mail: [email protected]
Qualification
• Ph.D. (Continuing)
National Institute of Technology, Rourkela, Orissa, India
• M. Tech. (Telematics & Signal Processing)
National Institute of Technology Rourkela, Orissa, India.
• B.E. (Electronics and Telecommunication Engineering)
Biju Patnaik University of Technology, Orissa, India [First division]
• +2 (Science)
Council of Higher Secondary Education, Orissa, India [First division]
• 10th
Board of Secondary Education, Orissa, India [First division]
Publications
• 04 Journal Articles
• 05 Conference Papers
Date of Birth
• 8th April, 1982
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement