SIGNATURE OF BURIED NETWORKS BY MATRIX

PENCIL METHOD
(Signature des réseaux enterrés par Matrix Pencil
Method)

Thesis by
Rupam Ghosh

In Partial Fulfillment of the Requirements for the

Degree of
Master of Science(M2) in Advanced Electronic Systems Engineering

UNIVERSITÉ DE BOURGOGNE
Dijon, France

2022
Defended 06th September 2022
 ii

© 2022
Rupam Ghosh
ORCID: https://orcid.org/0000-0002-2843-7058
All rights reserved
 iii
Thesis Tittle : SIGNATURE OF BURIED NETWORKS BY MA-
TRIX PENCIL METHOD (Signature des réseaux enterrés par
Matrix Pencil Method)

Thesis Category: Masters 2 Internship

Thesis Speciality: Electromagnetic Compatibility(EMC), Applied

Electronics

Place of Internship: CEM(PHOTON), Institut Pascal, Universite

Clermont Auvergne, Clermont-Ferrand, France

Supervisor: Prof. Dr. Khalil El Khamlichi Drissi, Institut Pascal,

UCA and Prof. Dr. Christophe Pasquier, UCA)

Author: Rupam GHOSH

University: Universite de Bourgogne, Dijon, France

 iv
ACKNOWLEDGEMENTS

The completion of this thesis was made possible by the assistance and
support of numerous people. First and foremost, I’d want to thank my
two supervisors.
Prof Dr Khalil El Khamlichi Drissi of the Institut Pascal, Universitè
Clermont Auvergne, always believed in me, assisted me, and encour-
aged me during some difficult times. I like his commitment and
efficiency.
I am grateful to Prof Dr Christophe Pasquier of the Universitè Cler-
mont Auvergne for his advice, inspiration, and technical aid with my
project. He welcomed me and assisted me in adjusting to new research
topics.
And I’d like to thank Dr Jean-Baptist Thomas of Sciences et Tech-
niques at Universitè de Bourgogne in Dijon, France, for his guidance
and assistance during my master’s journey.
I am also thankful to my defence committee’s rapporteurs for their
valuable involvement.
I am eternally grateful to my mother, Smt Rupali Ghosh, and my
family for their love and unending support. This master’s thesis is
dedicated to my mother.
 v
ABSTRACT

The objective of this thesis is to utilise radar backscattered Stepped

Frequency Ultra-Wideband (UWB) signals from Ground Penetrating
Radar to perform Automatic Target Classification (ATC). First, study-
ing scattering theory helps us to comprehend the physical significance
of the retrieved characteristics and mathematically define them. We
suggest creating a corpus of signatures using the Matrix Pencil Ap-
proach, a high-resolution method (MPM). This signature will be made
up of complex natural resonances (NCR) from the GPR, the ground,
and/or the target network. The poles and residues changed by the
influence of the GPR and/or the ground will be used to identify the
signature. As a result, we will attempt to construct signature clusters
by employing various scenarios. The collected signatures will then
be analysed and compared between two A-scans, maybe followed by a
B-scan. The simulation results show that MPMFD-extracted features
provide a feasible solution to automated target categorization. Finally,
the various algorithms are validated using experimental data and real-
world targets.

Keywords:Ground Penetrating Radar, Ultra-wideband Radar, Feature

extraction, Automatic Target Classification, Matrix Pencil Method.
 vi
Résumé

L’objectif de cette thèse est d’utiliser les signaux radar rétrodiffusés

à fréquence ultra-large (UWB) du radar à pénétration de sol pour
effectuer la classification automatique des cibles (ATC). Première-
ment, l’étude de la théorie de la diffusion nous aide à comprendre la
signification physique des caractéristiques récupérées et à les définir
mathématiquement. Nous suggérons de créer un corpus de signatures
à l’aide de l’approche Matrix Pencil, une méthode à haute résolution
(MPM). Cette signature sera constituée de résonances naturelles com-
plexes (NCR) issues du GPR, du sol et/ou du réseau cible. Les pôles
et résidus modifiés par l’influence du GPR et/ou du sol serviront à
identifier la signature. En conséquence, nous tenterons de construire
des grappes de signatures en utilisant divers scénarios. Les signatures
collectées seront ensuite analysées et comparées entre deux A-scans,
éventuellement suivis d’un B-scan. Les résultats de la simulation
montrent que les fonctionnalités extraites du MPMFD fournissent une
solution réalisable à la catégorisation automatisée des cibles. Enfin,
les différents algorithmes sont validés à l’aide de données expérimen-
tales et de cibles réelles.

Mots-clés: radar à pénétration de sol, radar ultra-large bande, extrac-

tion de caractéristiques, classification automatique des cibles, méthode
du crayon matriciel.
 vii
TABLE OF CONTENTS

Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . iv
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . vii
List of Illustrations . . . . . . . . . . . . . . . . . . . . . . . viii
Chapter I: Introduction . . . . . . . . . . . . . . . . . . . . . 2
1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Contributions and thesis organisations . . . . . . . . . 4
Chapter II: Electromagnetism and Scattering Theory . . . . . 5
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Notions of electromagnetism . . . . . . . . . . . . . . 5
2.3 Scattering Theory . . . . . . . . . . . . . . . . . . . . 7
Chapter III: UWB Radar . . . . . . . . . . . . . . . . . . . . 10
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 10
3.2 Radars . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3 UWB techniques . . . . . . . . . . . . . . . . . . . . . 12
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . 13
Chapter IV: Feature Extraction Method . . . . . . . . . . . . 14
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 14
4.2 Matrix Pencil Method . . . . . . . . . . . . . . . . . . 14
4.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . 15
Chapter V: Simulation Results . . . . . . . . . . . . . . . . . 16
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 16
5.2 Analysis of collected data . . . . . . . . . . . . . . . . 17
5.3 Matrix Pencil Method in Frequency Domain . . . . . . 19
5.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . 27
Chapter VI: Conclusion . . . . . . . . . . . . . . . . . . . . . 28
 viii
LIST OF ILLUSTRATIONS

Number Page
3.1 Radar block diagram . . . . . . . . . . . . . . . . . . . 10
3.2 FMCW Block Diagram . . . . . . . . . . . . . . . . . 13
5.1 Simulation Flowchart . . . . . . . . . . . . . . . . . . 16
5.2 Plot of collected Data in Time Domain and Frequency
Domain(after FFT) . . . . . . . . . . . . . . . . . . . . 18
5.3 Evaluation of Singular Values . . . . . . . . . . . . . . 19
5.4 Evaluation of error rates in percentage . . . . . . . . . 20
5.5 AScan as a function of wavelength . . . . . . . . . . . 21
5.6 AScan reconstructed according to the wavelength . . . 22
5.7 Real part of reduced poles . . . . . . . . . . . . . . . . 23
5.8 Imaginary part of reduced poles . . . . . . . . . . . . . 24
5.9 Real part of residues . . . . . . . . . . . . . . . . . . . 25
5.10 Imaginary part of residues . . . . . . . . . . . . . . . . 26
 1

List of Abbreviations

ADC Analog to Digital Converter

ATC Automatic Target Classification
CNRs Complex natural resonances
FFT Fast Fourier Transform
FMCW Frequency Modulated Continuous Wave
GPR Ground Penetrating Radar
MPM Matrix Pencil Method
MPMFD Matrix Pencil Method in Frequency Domain
SEM Singularity Expansion Method
SNR Signal-to-Noise Ratio
UWB Ultra-Wideband
 2
Chapter 1

INTRODUCTION

1.1 Overview
The research presented in this thesis, titled SIGNATURE OF BURIED NET-
WORKS BY MATRIX PENCIL METHOD (Signature des réseaux enterrés par Ma-
trix Pencil Method), was carried out as part of a Masters-2 internship programme be-
tween the Institut Pascal at the University of Clermont Auvergne (Clermont-Ferrand,
France) and the UFR Sciences et Techniques at the University of Bourgogne (Dijon,
France).
The topic is part of a larger issue that affects many businesses and communities. It
is about understanding and charting the French subterranean network. A significant
portion of the national facilities were constructed between 1945 and 1975, and the
accompanying as-built designs are frequently imprecise and, in some cases, incor-
rect. The age of the network, along with urban densification, causes various leaks
(gas, water, sewer) that might cause major accidents.
Ground Penetrating Radar (GPR) imaging is a non-destructive and promising ap-
proach for studying shallow underground structures in densely populated areas. We
propose to generate a corpus of signatures from the own echo obtained during the
transmission-reception of a multi-antenna GPR device (raw A-scan data in ampli-
tude and phase in the frequency domain) using a high-resolution approach dubbed
the Matrix Pencil Method (MPM). This signature will be made up of complex nat-
ural resonances (NCR) from the GPR, the ground, and/or the target network. The
poles and residues changed by the influence of the GPR and/or the ground will be
used to identify the signature.
As a result, we will attempt to create signature clusters by applying different situa-
tions to each simulated or even actual A-scan:
-Ground and GPR without a target network
-Ground and GPR with target array
The collected signatures will subsequently be analysed and compared between two
A-scans and may be followed during a B-scan.
-Ground and GPR without a target network
-Ground and GPR with target array
The autonomous imaging of buried objects is a complicated topic that has piqued
 3
the interest of researchers and associated industries. Many relevant industrial ap-
plications are interested in systems capable of conducting efficient and resilient
autonomous imaging. The interest in Ultra Wideband (UWB) technologies has de-
veloped considerably during the last fifteen years. One of the UWB’s many uses is
radar automated target categorization in autonomous imaging. UWB radar has very
short-duration pulses, resulting in a very broad frequency spectrum. The UWB
approach has the benefit of being able to be utilised for target identification and
localisation.
In this thesis, we propose using UWB radar to classify an unknown object automat-
ically. The two primary processes of Automatic Target Categorization (ATC) are
feature extraction and classification. The target is identified by comparing the de-
rived target qualities to the various target features that have previously been recorded
in a database. A proper selection of characteristics is essential for distinguishing
between distinct targets. Understanding the physical behaviour of a radar system
in its surroundings is essential for optimum feature extraction handling. Based on
this knowledge, characteristics must then be mathematically characterised in accor-
dance with the specified specifications. When an ultra-wideband signal illuminates
a target, the dispersed transient response in the temporal domain is made up of two
successive parts [1].
First, an impulsive component, corresponding to the early time response, results
from the direct reflection of the incoming wave on the object surface. The oscillat-
ing component then emerges from the resonance phenomena of the target at the late
time. When targets are ideal conductors, resonances arise outside the object and
correspond to surface creeping waves.
Currently, various strategies exist to extract the poles of resonance based on the
study of the late time impulse response: Fourier Transform, MUltiple Signal Clas-
sification (MUSIC), Estimation of Signal Parameters through Rotational Invariance
Techniques (ESPRIT), Prony’s method, and so on. The Matrix Pencil Method [2]
is one of the most efficient and precise methods for extracting resonance poles.
The Matrix Pencil approach is used in this paper to estimate the parameters of ex-
ponentially damped sinusoids in the frequency domain. This paper is interested in
the following applications: non-intrusive identification and load monitoring; UWB
radar for communication; localisation; and/or target identification. The effort in-
tends to achieve the following research objectives:
to comprehend raw time domain signal data obtained from real-time stepped fre-
quency GPR.
 4
Before converting the signal to the frequency domain, evaluate the data to remove
any extraneous noise or echo.
To investigate and improve the efficacy of the Matrix Pencil Method in Frequency
Domain feature extraction by using real-time data.

1.2 Contributions and thesis organisations

The remainder of this dissertation is organized as follows:
Chapter 2 presents Electromagnetism and Scattering Theory with Singularity
Expansion Method(SEM)
Chapter 3 gives notions of Ultra-Wide Band Techniques
Chapter 4 presents feature extraction method as Matrix Pencil Method
Chapter 5 gives the simulation and experimental results
Chapter 6 outlines the conclusion. It summarizes the main achievement of this
dissertation and proposes the future research proprieties.
 5
Chapter 2

ELECTROMAGNETISM AND SCATTERING THEORY

2.1 Introduction
We introduce some notions of electromagnetism in this chapter. The scattering the-
ory is then presented, with the Singularity Expansion Method (SEM) demonstrated
as the scattering model.

2.2 Notions of electromagnetism

Maxwell Equations
A full set of equations governing the behaviour of the electromagnetic phenomenon
was presented by James Clark Maxwell (1831–1879), it involves the behaviour of
two vector fields: the electric field 𝐸® , and the magnetic induction 𝐵.
®

The fundamental Maxwell equations are expressed in time domain as:

®
® = − 𝜕𝐵
∇ × ( 𝐸) (2.1)
𝜕𝑡

∇ · 𝐵® = 0 (2.2)

𝜕 𝐸®
∇ × 𝐵® = 𝜇( 𝐽® + 𝜖 ) (2.3)
𝜕𝑡

𝜌
∇ · 𝐸® = (2.4)
𝜖

𝐽® is the electric current density and 𝜌 is the electric charge density. 𝜖 is the electric
permittivity, and 𝜇 is the magnetic permeability. The electric permittivity is related
to its relative permittivity 𝜖𝑟 and the permittivity of the vacuum by the relation
𝜖 = 𝜖0 𝜖𝑟 . In the same way, the magnetic permeability is related to its relative
permeability and the permeability of the vacuum by the relation: 𝜇 = 𝜇0 𝜇𝑟. Further
discussion about the physical significance of each of these equations may be found
in [8, 9].
 6
The interactions between the fields and the medium of propagation are described
by the constitutive relations. These are of the form for a linear, homogeneous, and
isotropic medium.

𝐷® = 𝜖 𝐸® (2.5)

𝐵® = 𝜇 𝐻® (2.6)

Time Harmonic Fields

The Maxwell’s equations are typically used in their harmonic version, taking into
account that fields and sources have sinusoidal time dependence. The electric field
can therefore be expressed as

𝐸® (®
𝑟 , 𝑡) = 𝐸® (®
𝑟 )𝑒 𝑗𝜔𝑡 (2.7)

and Maxwell’s equations (2.1) and (2.3) become:

∇ × 𝐸® = − 𝑗𝜔 𝐵® (2.8)

∇ × 𝐻® = 𝐽® + 𝑗𝜔 𝐷® (2.9)

Propagation equation in unbound media

All sources are presumably placed in the infinite, so they don’t show up in the
equations. The Maxwell’s equations enable the derivation of the electromagnetic
wave propagation equations in an unbounded, linear, homogeneous, and isotropic
medium in time domain of the electromagnetic fields 𝐸® and 𝐻:
®

1 𝜕 2 𝐸®
∇2 𝐸® − =0 (2.10)
𝑣 2 𝜕𝑡 2

1 𝜕 2 𝐻®
∇2 𝐻® − 2 2 = 0 (2.11)
𝑣 𝜕𝑡

where ∇2 is the Laplacian. The propagation speed of the wave in infinite media is:
1
𝜐=√ (2.12)
𝜖𝜇
 7
We look for sinusoidal form solutions to fields with time variations. The elementary
solution’s general form is
®
𝐸®0 𝑒 𝑗 (𝜔𝑡− 𝑘 𝑟®) (2.13)

𝐸®0 is a constant vector, 𝑘® is the wave vector. Since all wave plan points perpendicular
to the vector of propagation k exhibit the same vibratory behaviour, this solution
resembles a plane wave. The wave number k is defined as follows:
2𝜋
𝑘= (2.14)
𝜆
where 𝜆 is the wavelength.

2.3 Scattering Theory

The definition of scattering is the rerouting of radiation away from its initial direc-
tion of propagation, frequently as a result of interactions between the wave and the
target. When the scattering field is radiated in the backward direction to the incident
wave, it is called back-scattered field.
For scattering, various models have been put forth that either use resonances or scat-
tering centres. Baum [10] introduced the Singularity Expansion Method (SEM), a
time-domain mathematical model based on Complex Natural Resonances (CNRs),
in which late time, a subset of distributed fields in the time domain, is expressed as
a series of damped exponentials.
As an alternative to resonance-based processing [11] The Geometric Theory of
Diffraction (GTD) model is frequently employed in the frequency domain to char-
acterise the properties of the scattering centres. According to GTD, the target
backscatter originates from a number of distinct scattering centres.
First, we present the concept of radar cross section in this part (RCS) and then the
SEM is proffered.

Radar Cross Section

Any electromagnetic pulse that illuminates an item scatters incident energy through-
out the entire scene. The term "scattering" refers to this spatial energy distribution,
and the actual object is referred to as a scatterer. The backscattering of the wave’s
energy, which creates the radar echo of the object, is referred to as backscattering.
The intensity of the echo is measured using the radar cross section.
 8
The RCS of an object,𝜎, is defined as an equivalent area intercepting the amount
of power that, when scattered isotopically, produces at the receiver a power density
that is equal to the power density scattered by the actual target. This is given by:
|𝐸 𝑠 | 2
𝜎 = lim [4𝜋𝑟 2 ] (2.15)
𝑟→∞ |𝐸 𝑖 | 2

where:
𝐸 𝑖 is the electric field strength of the incident wave impinging on the target, (𝑉 ·𝑚 −1 )
𝐸 𝑠 is the electric field strength of the scattered wave at the radar, (𝑉 · 𝑚 −1 )
𝑟 is the distance from the target to the radar.
As the name suggests, the RCS has dimensions of area: metre-squared (𝑚 2 ).

Singularity Expansion Method

Two successive components make up the scattered transient response in the time
domain when wideband signals illuminate a target. First, an impulsive component
that corresponds to the early time response results from the incoming wave’s direct
reflection on the surface of the item. The oscillating component then develops
in the late time as a result of target resonance occurrences. Internal and external
modes of the resonances can be distinguished [13]. While surface creeping waves
are responsible for the outward resonances, it is the interior waves that undergo
repeated internal reflections that cause them. When targets are ideal conductors,
resonances exclusively correspond to external modes and take place external to the
item.
The electromagnetic response has been expressed in an extension of the system’s
complex resonances using Baum’s singularity expansion approach. In general,
the following can be stated as the signal model of the observed late time of an
electromagnetically scattered reaction from an object:
𝑀
∑︁
𝑦(𝑡) = 𝑥(𝑡) + 𝑛(𝑡) ≈ 𝑅𝑚 𝑒 𝑆 𝑚 𝑡 + 𝑛(𝑡) 0≤𝑡 ≤𝑇 (2.16)
𝑚=1

where:
𝑦(𝑡) : denotes the time domain response
𝑛(𝑡) : denotes the noise in the system
𝑥(𝑡) : denotes the noiseless signal
𝑅𝑚 : residues or complex amplituse
𝑆 𝑚 = 𝛼𝑚 + 𝑗𝜔𝑚 ,
 9
𝛼𝑚 : damping factors
𝜔𝑚 : are angular frequencies
𝑀 is the total number of modes of the series
After sampling, the time variable, 𝑡 is replaced by 𝑘𝑇𝑠 , where 𝑇𝑠 is the sampling
period. The original continuous time sequence can be rewritten as:
𝑀
∑︁
𝑦(𝑘𝑇𝑠 ) = 𝑥(𝑘𝑇𝑠 ) + 𝑛(𝑘𝑇𝑠 ) ≈ 𝑅𝑚 (𝑧 𝑘 ) 𝑚 + 𝑛(𝑘𝑇𝑠 ) 𝑓 𝑜𝑟 𝑘 = 0, ..., 𝑁 − 1 (2.17)
𝑚=1

𝑧 𝑚 = 𝑒 𝛼𝑚 + 𝑗𝜔𝑚 )𝑇𝑠 𝑚 = 1, 2, ..., 𝑀 (2.18)

We apply the Laplace transform of 𝑥(𝑡) to obtain the transfer function 𝐻 (𝑠) :
𝑀
∑︁ 𝑅𝑚
𝐻 (𝑠) ≈ (2.19)
𝑚=1
𝑠 − 𝑆𝑚

where 𝑠 = 𝜎 + 𝑗𝜔 is the complex variable in the Laplace plane.

Conclusion
This chapter provides background knowledge on electromagnetic theory. Following
that, the scattering theory was provided, which pointed to a time domain SEM
model. The UWB radar is described in depth in the next chapter.
 10
Chapter 3

UWB RADAR

3.1 Introduction
In this chapter we will get hands on some basic parameters with an overview of
Radars and then an overview of the UWB Radar in Automatic Target Classification
(ATC) will be presented.

3.2 Radars
Radar is an acronym for Radio Detection and Ranging. The characteristics of a
target can be ascertained using radio energy. A radar system, in its most basic
configuration, consists of three subsystems: a transmitter, a receiver, and an antenna
system (Figure 3.1). The antenna system broadcasts an electrical signal that is
produced by the transmitter. The signal will partially reflect back to the radar
system if it is incident on a target, such as an animal, car, or aeroplane. At the

Figure 3.1: Radar block diagram

receiver’s output, a decision is made regarding whether or not a target echo signal
is present following amplification by the receiver and with the help of appropriate
signal processing. At that point, the target’s location and other details are discovered.
 11
Informatic Parameters of Radars
Several types of information can be collected by a radar:
-Detection: The radar is currently used for a variety of purposes, but its initial goal
of detecting things in some volume of space remains a significant aspect of all of its
applications. In this scenario, the user is interested in identifying targets in the lit
volume from the clutter and noise that tend to conceal them.
-Range : The range to a target is calculated by measuring the time it takes for the
radar signal to travel at the speed of light from the radar to the target and back. No
other sensor can accurately calculate the distance to a distant target at long range
like radar. Precision can be a few millimetres at low ranges. The distance 𝐷 is given
by:
𝑐 · Δ𝑇
𝐷= (3.1)
2
where 𝑐 is the speed of light, and Δ𝑇 is the temporal shift between transmission and
reception of the radar signal.

-Radial Velocity : The rate of change of range over time is used to calculate a target’s
radial velocity. It can also be determined by measuring the Doppler frequency shift.
The target’s motion indicates that the time between the transmission and receipt of
the radar echo differs for two consecutively sent impulses. The velocity of the target
is calculated from this difference. The system emits a signal each 𝑇𝑅 and receives
the radiated echo by the target together with the generated temporal shift. If for
the first echo, the shift is 𝛿𝑡 1 = 2·𝐷 1
𝑐 with 𝐷 1 the distance radar/target, for the next
echo and because of the target motion, the temporal shift 𝛿𝑡2 will equal to 𝛿𝑡1 plus a
quantity 𝛿𝑡, positive if the target moves away and negative if it comes closer. If 𝑉𝑅
is the velocity, then:
2 · 𝑉𝑅 · 𝑇𝑅
𝛿𝑡 = (3.2)
𝑐
So, the velocity 𝑉𝑅 of the target is:
𝑐 · 𝛿𝑡
𝑉𝑅 = (3.3)
2 · 𝑇𝑅
 12
Ground Penetrating Radar
Ground-Penetrating Radar is a geophysical technology that images the subsurface
using radar pulses. The reflected signals from subsurface structures are detected
using this technology. GPR may be used to a wide range of media, including rock,
soil, ice, fresh water, pavements, and constructions. It can detect things, material
changes, and cavities and fissures. GPRs are either carrier-free or use a low fre-
quency carrier (500 MHz); the low frequency content of the sent signal helps it
to penetrate materials like concrete and earth more easily than a higher frequency
signal.
Consider a metallic item that has been buried in the earth. The sent signal will be
reflected by the object and will be incident on the reception antenna after some time
delay. The time delay may be used to calculate the distance to the buried object,
implying that GPR utilises time-domain methods to find any discontinuities in the
otherwise opaque target.
Since the 1970s, engineers have frequently employed GPR in both civilian and
military applications. However, the first GPRs capable of high performance were
marketed in the early 1990s. Since then, the number of users of this approach
has grown, and GPRs are now used in a wide range of applications, including the
identification of hidden mines, the discovery of fossils, the evaluation of pavements,
and the research of the moon’s subsurface structure, etc [14,15,16,17,18,19,20].

Signal To Noise Ratio (SNR):

The signal to noise ratio of an ultra-wide band ground penetrating radar is given by:
(𝑃𝜏𝐺 𝑡 𝐺 𝑟 𝜆 1𝜆 2 𝜎)
𝑆𝑁 𝑅 = (3.4)
(4𝜋 3 𝑅 4 𝑘𝑇)10𝛼𝑅
where 𝑃 is the transmitted power, 𝜏 is the integration time, 𝐺 𝑡 and 𝐺 𝑟 are the
transmit and receive antenna gains, 𝜆 1 and 𝜆 2 are the longest and shortest wave
lengths respectively, 𝜎 is the target radar cross section, 𝑅 is the target range, 𝑘
is Boltzmann constant, 𝑇 is the effective receiver noise temperature and 𝐵 is the
receiver bandwidth

3.3 UWB techniques

The technical design of ground penetrating radars can be classified into two groups,
Time Domain operated GPR and Frequency Domain Operated GPRs. In this series
of experiments , The Stepped Frequency Continuous Wave type of GPRs will be
focused into.
 13
Frequency Modulated Continuous Wave Radar
FMCW radar (Frequency Modulated Continuous Wave radar = FMCW radar) is a
special type of radar sensor which radiates continuous transmission power like a
simple continuous wave radar (CW Radar). In contrast to this CW radar FMCW
radar can change its operating frequency during the measurement: that is, the
transmission signal is modulated in frequency (or in phase). Possibilities of Radar
measurements through runtime measurements are only technically possible with
these changes in the frequency (or phase). FMCW Radar is mostly used as Radar
Altimeter in order to measure the exact height while landing the aircraft. The
following figure shows the block diagram of FMCW Radar

Figure 3.2: FMCW Block Diagram

3.4 Conclusion
In this chapter definition, and applications of UWB radar have been provided. the
next chapter we will present the methods used for feature extraction.
 14
Chapter 4

FEATURE EXTRACTION METHOD

4.1 Introduction
The primary goal of this thesis is to study the essential parameters of Matrix Pencil’s
signal processing and feature extraction approach as they apply to backscattered data
from SF GPR radar. The two essential processes in ATC are feature extraction and
categorization. We offer the Matrix Pencil approach for feature extraction as a
high-resolution methodology in this project.

High resolution methods

High resolution approaches are used to estimate and/or identify the desirable and/or
undesired signal components contained in a given collection of data. The phrase
"high resolution" refers to the capacity to distinguish between highly similar signal
components. The parametric high resolution approaches are the outcome of clever
exploitations of well-known data structures. This section introduces the Matrix
Pencil Method (MPM).

4.2 Matrix Pencil Method

The Singularity Expansion Method (SEM) by Baum [4] introduces a practical
approach for expressing the late time response of different scatterers in terms of a
limited sum of exponentially attenuated sinusoids as:
∑︁ 𝑟
ℎ𝑟 (𝑡) = 𝑅𝑟 𝑒 𝑆 𝑡 (4.1)
𝑟

𝑅𝑟 is the residue to each pole

𝑆𝑟 is the 𝑟 𝑡ℎ pole of the target: 𝑆 𝑅 = 𝜎𝑟 + 𝑗𝜔𝑟,
𝜎𝑟 = 2𝜋 𝑗 𝑓𝑟 with 𝑓𝑟 natural frequency.
The MPM is applied to the impulse response of a target illuminated by a broadband
electromagnetic wave at late time instants 𝑌𝐿 (𝑡). Following the sample technique,
it is possible to write:
𝑁
∑︁
𝑦𝑟 (𝑘𝑇𝑠 ) = ℎ𝑟 (𝑘𝑇𝑠 ) + 𝑏𝑟 (𝑘𝑇𝑠 ) ≈ 𝑅𝑛 (𝑧 𝑛 ) 𝑘 + 𝑏𝑟 (𝑘𝑇𝑠 ) (4.2)
𝑛=1

with
 15

𝑅𝑛 = |𝑎 𝑛 |𝑒 𝑗 𝜙𝑛 n=1,2,...,N (4.3)

𝑧 𝑛 = 𝑒 𝑆 𝑇 𝑠 = 𝑒 (𝜎
𝑟 + 𝑗𝜔
𝑟 )𝑇
𝑛 𝑠
𝑛 = 1, 2, ..., 𝑁 (4.4)

The MPM is carried out in two steps. The first stage involves resolving Hankel’s
matrix with Singular Value Decomposition (SVD) to estimate the number of poles
𝑁 and the value of resonance poles 𝑍𝑛 .

 𝑦𝑟 (0) 𝑦𝑟 (1) ... 𝑦𝑟 (𝐿) 
 
 𝑦𝑟 (1) 𝑦𝑟 (2) ... 𝑦𝑟 (𝐿 + 1) 
(𝑌𝑟 ) =   (4.5)
 ... ... ... ... 
 
 𝑦𝑟 (𝐾 − 𝐿 − 1) 𝑦𝑟 (𝐾 − 𝐿) ... 𝑦𝑟 (𝐾 − 1) 

where 𝑌𝑟 is of dimensions (𝐾 − 𝐿) × (𝐿 + 1) and 𝐿 is the Pencil Parameter. 𝑅𝑛 can
be obtained from the following equation upon knowing terms 𝑁 and 𝑧 𝑛 :
 𝑦(0)   1 1 ... 1  𝑅 
    1
     
 𝑦(1)   𝑧1 𝑧2 ... 𝑧 𝑁   𝑅2 
 ...  =  ... (4.6)
    
   ... ... ...   ... 
 
  (𝑘−1) (𝑘−1)
... 𝑧 𝑁(𝑘−1) 
   
 𝑦(𝑘 − 1)  𝑧 𝑧2 𝑅𝑁 
   1  
Note that the impulse response is represented by 𝑁 poles and residues respectively.

4.3 Conclusion
In this chapter, high resolution methods for feature extraction are discussed. In the
next chapter, we show the findings of simulations and experiments.
 16
Chapter 5

SIMULATION RESULTS

5.1 Introduction
We describe the simulation and experimental results from this effort in this chapter.
We start by presenting the simulation findings of the GPR signal echo data that was
gathered. The findings of feature extraction using the Matrix Pencil Method are
then displayed.

Figure 5.1: Simulation Flowchart

 17
5.2 Analysis of collected data
In this section , the collected GPR data analysis is presented through MATLAB
software which has been used to compute. The Data has been collected by Univer-
sitè Gustave Eiffel with a Stepped Frequency Ground Penetrating Radar.

The parameters of the collected data is given below

Data collection iterations: 3392

𝑑𝑥 : 0.0025 metre
𝑑 𝑦 : 0.0025 metre
𝑑𝑡 : 5.89664e-12 secs
Number of Steps: 50
Number of 𝑡𝑥 : 1
Number of 𝑟 𝑥 : 1
Number of 𝑟 𝑥 𝑏𝑜𝑥 : 0
𝑡𝑥 : Number 1
𝑡𝑥 : Initial position 30 98 (cell coords)
𝑡𝑥 : ID source
𝑡𝑥 : Delay 0 secs
𝑡𝑥 : Removed at 2e-08 secs
𝑟 𝑥 : 50 98 (cell coords)
𝑡𝑥 𝑠𝑡𝑒 𝑝𝑠 : 0, 0 (cell coords)
𝑟 𝑥 𝑠𝑡𝑒 𝑝𝑠 : 2, 0 (cell coords)
 18
The Signals in Time Domain
The Electromagnetic field signals are given in Time Domain. We will next analyse
and trim the data to increase the efficiency of the input before MPM processing.

The Signals in Frequency Domain

Signals in the time domain can be input directly into the Matrix Pencil Method
(MPM) for feature extraction, but the output effectiveness of MPM is increased by
data in the frequency domain, thus we must first translate the time domain signals
into the frequency domain. The time domain signals can be used to generate the
frequency domain signals. We must first convert the signal using the fast Fourier
transformed (FFT) method in order to get the frequency response from it.

Figure 5.2: Plot of collected Data in Time Domain and Frequency Domain(after
FFT)

Discussion on Fig. 5.2

This graphic shows a clear comparison of Signals in the Time and Frequency Do-
mains. The raw signal in Time Domain consists of three echos, the second and third
of which may be ignored during simulation because they carry insignificant infor-
mation in comparison to the first echo. Thus, this data is trimmed and transformed
into Frequency Domain by FFT, and we obtain a new data chart drawn on Frequency
Domain from which features will be retrieved via MPMFD.
 19
5.3 Matrix Pencil Method in Frequency Domain
Matrix Pencil Method in Frequency Domain (MPMFD) [21] is applied on the SEM
model presented in section 2.3. MPMFD is conducted upon the data in frequency
domain taking all parameters into account . We take only three poles for simulation.

Simulation Results are given below

Figure 5.3: Evaluation of Singular Values

 20

Figure 5.4: Evaluation of error rates in percentage

 21

Figure 5.5: AScan as a function of wavelength

 22

Figure 5.6: AScan reconstructed according to the wavelength

 23

Figure 5.7: Real part of reduced poles

 24

Figure 5.8: Imaginary part of reduced poles

 25

Figure 5.9: Real part of residues

 26

Figure 5.10: Imaginary part of residues

 27
5.4 Conclusion
The signals are generated in the frequency domain. The findings of feature extraction
were then given. We picked MPM as our primary approach for feature extraction
since it is simpler than other high resolution methods. Furthermore, it produced
outstanding results and is less susceptible to noise.The nature of signatures for three
poles may be determined from the figures above. It can also be seen that the Buried
Object is at an angle with the earth layers. Furthermore the length of the buried
object can be obtained from the graph taking 𝑑𝑥 = 0.0025 meter and 𝑑 𝑦 =0.0025
meter with initial position 30(𝑋)and 98(𝑌 ).
 28
Chapter 6

CONCLUSION

In this thesis, we explored using Matrix pencil feature extraction on UWB radar
data to automatically characterise an unknown item buried beneath. Automatic
Target Categorization (ATC) is comprised of two major processes: feature extrac-
tion and categorization. By comparing the resulting target characteristics to the
numerous target traits already recorded in a database, the target was recognised.
For discriminating between distinct targets, a correct set of attributes was required.
Understanding the physical behaviour of a radar system in its surroundings had
been primarily investigated in order to handle feature extraction more effectively.
Characteristics must then be numerically described in compliance with the defined
standards based on this knowledge.
Our aim in Chapter 2 was to comprehend the physical behaviour of electromagnetic
waves as they struck the target (scattering). Based on this knowledge, characteristics
must then be mathematically characterised in accordance with the specified spec-
ifications. We discovered that several models for scattering have been developed,
employing either resonances or scattering centres.
In Chapter 3, we discussed the benefits of employing UWB radar instead of con-
ventional radars. The majority of the benefits of UWB radars are related to the
acquired bandwidth. Essentially, the greater the bandwidth of a system, the better it
operates. Then we discussed the various approaches utilised in UWB radar. We can
employ either temporal or frequency approaches. Finally, we discussed the various
ATC phases.
We discussed the strategies we used for feature extraction in Chapter 4. The Singu-
larity Expansion model used to apply high resolution approaches in the frequency
domain.
In Chapter 5, we reported simulation and experiment findings. In terms of high
resolution approaches, we discovered that MPM is appropriate for the rest of our
work due to its simplicity and lower susceptibility to noise in contrast to other meth-
ods. Through the simulated findings, we also acquired the signature pattern via pole
and residues on real and imaginary regions.
The simulation results demonstrated the significance of feature extraction in target
categorization. The results show that MPMFD-extracted features offer a viable
 29
approach to automated target categorization. The division between early and late
time is not necessary with this strategy. MPMFD may be used to create a database
containing object signatures.
It might be interesting to build Classifiers based on Convolutional Neural Networks
or Machine Learning in the future. It would be intriguing to integrate the created
ATC approach with localization and communication algorithms and apply them on
electrical devices such as FPGA.
 30

References

[1] N.de Beaucoudrey J.Chauveau and J.Saillard.Characterization of perfectly con-

ducting targets in resonance domain with their quality of resonance. Progress In
Electromagnetics Research,74:69–84, 2007.

[2] E. J. Rothwell and K. M. Chen. A hybrid e-pulse/least squares technique for

natural resonance extraction. Proceedings of the IEEE, 76:296–298, 1988.

[3] R. Kumaresan and D. W. Tufts. Estimating the parameters of exponentially

damped sinusoids and pole-zero modeling in noise. IEEE Transactions on Acous-
tics, Speech, and Signal Processing, 30:833–840, 1982.

[4] T. K. Sarkar and O. Pereira. Using the matrix pencil method to estimate the
parameters of a sum of complex exponentials. IEEE Ant. and Prop. Mag, 37:4855,
1995.

[5] N. V. Z. Shuley. A review of uncooperative target identification using uwb

resonance based radar techniques. ECTI Transactions on Electrical Engineering,
Electronics, and Communications, 4:118–126, 2006.

[6] G. Turhan-Sayan. Real time electromagnetic target classification using a novel

feature extraction technique with pca-based fusion. IEEE Transactions onAntennas
and Propagation, 53:766–776, 2005.

[7] H. Moon W. Lee, T. K. Sarkar and M. Salazar-Palma. Computation of the

natural poles of an object in the frequency domain using the cauchy method. IEEE
Antennas andWireless Propagation Letters, 11:1137–1140, 2012.

[8] J.A. Stratton. Electromagnetic Theory. Mc Graw-Hill, London, 1941.

 31
[9] Odile Picon et coll. Les antennes : Th´eorie, conception et applications.
Dunod, Paris, 2009.

[10] C. E. Baum. The singularity expansion method. Transient Electromagnetic

Field, pages 129–179, 1976.

[11] J. B. Keller. Geometrical theory of diffraction. J. Opt. Soc. Amer, page

116130, Jan. 1962.

[12] M. Skolnik. Radar Handbook. 3rd Edition,The McGraw-Hill Companies,

2008.

[13] C-C Chen. Electromagnetic resonances of immersed dielectric spheres. IEEE

Transaction on Antennas and Propagation, 46:1047–1083, July 1998.

[14] J. D. Taylor. Ultra-Wideband Radar Technology, chapter 13, page 343378.

CRC Press, 2001.

[15] I. Chernyi S. Lukjanov, R. Stepanov and O. Stukach. Use of the ground pen-
etrating radar methods for paleontology on example of the mammoth fauna investi-
gation. Proceedings of the 37th European Microwave Conference, page 17471750,
2007.

[16] C. Nguyen J. S. Lee and T. Scullion. A novel, compact, low-cost, impulse

ground-penetrating radar for nondestructive evalulation of pavements. IEEE Trans-
actions on Instrumentation and Measurement, 53:15021509, 2004.

[17] V. Utsi. Design of a gpr for deep investivations. 4th International Workshop
on Advanced Ground Penetrating Radar, page 222225, 2007.
 32
[18] O. Brandt K. Langley, R. Storvold and J. Kohler. Use of c-band ground pen-
etrating radar to determine backscatter sourceswithin glaciers. IEEE Transactions
on Geoscience and Remote Sensing, 45:12361246, 2007.

[19] W. Kofman. Radar techniques to study subsurfaces and interiors of the solar
system objects. 19th International Conference on Microwave Radar and Wireless
Communications, page 409 412, 2012.

[20] Fa. Wenzhe. Modeling and simulation for ground penetrating radar study
of the subsurface structure of the moon. 14th International Conference on Ground
Penetrating Radar, page 922 926, 2012.

[21] M. Khodjet-Kesba K. El Khamlichi Drissi, S. Lee and coll. Comparison of

matrix pencil extracted features in time domain and in frequency domain for radar
target classification. International Journal of Antennas and Propagation, 2014,
Article ID 930581, 9 pages, 2014.

[22] Xindong Wu ., V. Kumar , J. Ross Quinlan , J. Ghosh , Q. Yang , H. Motoda ,

G. J. McLachlan , Angus Ng , B. Liu , Philip S. Yu , Zhi-Hua Zhou , M. Steinbach
, D. J. Hand and D. Steinberg, “Top 10 algorithms in data mining,” Knowledge and
Information Systems, V.14 n.1, p.1-37,December 2007.

[23] Mitchell, T., Machine Learning, McGraw Hill, 1997.

[24] Cover, T. and P. Hart, “Nearest neighbor pattern classification,” IEEE Trans-
actions on Information Theory, 13, 21–27, 1967.

[25] Kecman, V., Learning and Soft Computing, MIT Press, Cambridge, MA, 2001.