Citation
Modulation, coding and detection for satellite and space communications

Material Information

Title:
Modulation, coding and detection for satellite and space communications
Creator:
Etellisi, Ehab A
Place of Publication:
Denver, CO
Publisher:
University of Colorado Denver
Publication Date:
Language:
English
Physical Description:
viii, 101 leaves : ill. ; 28 cm.

Subjects

Subjects / Keywords:
Signal processing -- Digital techniques ( lcsh )
Artificial satellites in telecommunication ( lcsh )
Wireless communication systems -- Security measures ( lcsh )
Modulation (Electronics) ( lcsh )
Coding theory ( lcsh )
Artificial satellites in telecommunication ( fast )
Coding theory ( fast )
Modulation (Electronics) ( fast )
Signal processing -- Digital techniques ( fast )
Wireless communication systems -- Security measures ( fast )
Genre:
bibliography ( marcgt )
theses ( marcgt )
non-fiction ( marcgt )

Notes

Thesis:
Thesis (M.S.)--University of Colorado Denver, 2011. Electrical engineering
Bibliography:
Includes bibliographical references (leaves 99-101).
General Note:
Department of Electrical Engineering
Statement of Responsibility:
by Ehab A. Etellisi.

Record Information

Source Institution:
|University of Colorado Denver
Holding Location:
|Auraria Library
Rights Management:
All applicable rights reserved by the source institution and holding location.
Resource Identifier:
747019919 ( OCLC )
ocn747019919

Downloads

This item has the following downloads:


Full Text
A
MODULATION, CODING AND DETECTION FOR SATELLITE AND SPACE
COMMUNICATIONS
by
Ehab A. Etellisi
B.S., A1 Fateh University, 2005
A thesis submitted to the
University of Colorado Denver
in partial fulfillment
of the requirements for the degree of
Master of Science
Electrical Engineering
2011


This thesis for the Master of Science
degree by
Ehab A. Etellisi
has been approved
by
Papantoni


Etellisi, Ehab A. (Master of Science, Communications Engineering, Electrical Engineering, University
of Colorado Denver)
Modulation, Coding and Detection for Satellite and Space Communications
Thesis directed by Professor Titsa Papantoni
ABSTRACT
Communication and signal processing applications can be linked to our current and historic
events in the social, political, economic, and cultural dimensions in the human world. At the same
time, prominent performance metrics in various such applications include accuracy, transmission
power, bandwidth efficiency, security and cost effectiveness. In this thesis, we present and analyze
three applications of current interest, each of which invests heavily on one of the following
performance metrics: accuracy, bandwidth/energy efficiency and security. In the first four chapters,
we give some background on communication and signal processing such as modulation and coding.
Then in chapter five, we start introducing some communications system applications. For instance, to
address accuracy; we selected the problem of oil-spill detection: we show how system accuracy may
minimize remediation costs and limit dangerous impacts to the environment. To address bandwidth/
energy efficiency, we selected a voice activity detection problem. To address security, we selected the
problem of timely and accurately detecting the presence of cyber exploits in the communication
transmission, where a new detection method is introduced and analyzed. In our developments, cost
effectiveness is of high priority as well.
This abstract accurately represents the content of the candidates thesis. I recommended its publication.
Signe


DEDICATION
I would like to dedicate my thesis to my belovedfamily, especially...
to Dad and Mom for instilling the importance of hard work and higher education;
to my brother, Gaith, for his encouragement and support;
to my friends and relatives, and supporters who have made this happen;
finally, many thanks go out to my country for providing me this scholarship.


ACKNOWLEDGEMENTS
This work would not have been possible without the scholarship provided to me by my
country. I would like to thank my family and friends for their understanding, encouragement and
support in my pursues. I am indebted to all my faculty members who have strongly affected so much
on my educational experience like Dr. Yiming J. Deng. I offer my regards and blessings to all of those
who supported me in any respect during the completion of this thesis.
I am heartily thankful to my supervisor, Prof. Titsa Papantoni, whose encouragement,
guidance and support from the initial to the final level enabled me to develop an understanding of the
subjects that I covered in this thesis. She is an excellent role model, as someone who really knows how
to balance scientific research and teaching. Despite her busy schedule, she always finds the time to
discuss anything from the first ideas and designs to simulation results. I can only express that her
dedication and commitment to science and education is truly inspiring and remarkable.
The partial support of the contract AFOSR FA9550-05-1-0388 is acknowledged.
Finally I would like to express my appreciation to whomever touched my life directly or indirectly.


TABLE OF CONTENTS
Figures....................................................................................ix
Tables....................................................................................xii
CHAPTER
1. Introduction to Digital Communication Systems..........................................1
2. Signal Processing......................................................................9
2.1 Analog to Digital Conversion..........................................................10
2.2 Anti-aliasing Filter..................................................................10
2.3 Quantization (Uniform Quantization)...................................................10
2.4 Sampling..............................................................................10
3. Modulation Techniques.................................................................16
3.1 Introduction..........................................................................16
3.2 Modulation Techniques.................................................................16
3.3 Frequency Shift Keying (FSK)..........................................................18
3.3.1 Binary Frequency Shift Keying (BFSK)................................................18
3.3.2 M- Frequency Shift Keying (MFSK)....................................................18
3.4 Phase Shift Keying (PSK)..............................................................19
3.5 Differential Phase Shift Keying (DPSK)............................................... 23
3.6 Quadrature Amplitude Modulation (QAM).................................................24
3.7 Signal Constellation................................................................. 25
4. Channel Coding....................................................................... 26
vii


4.1 A Novel Trellis-Coded Modulation
26
4.1.1 Trellis-Coded Modulation Background................................................ 26
4.1.2 Fundamentals and Concept of TCM................................................... 27
4.1.3 Mapping and Trellis Diagram....................................................... 29
4.1.4 Viterbi Algorithm................................................................. 33
4.1.5 TCM MFSK, MPSK, DPSK, and MQAM Constellation Mapping.................................... 34
5. Communication Systems Applications....................................................41
5.1 Oil Spill Detection by Satellite Image using Sequential Detection of Change Test........... 41
5.2 Sequential Tests for the Detection of Voice Activity and the Recognition of Cyber Exploits.61
5.3 Detect Cyber Attacks using BER-Threshold........................................... 88
6 Conclusion............................................................................. 94
Appendix
A. Derivations....................................................................... 96
Bibliography............................................................................. 99
viii


LIST OF FIGURES
Figure
1.1 Block diagram of a typical digital communication system.............................. 5
1.2 Formatting and transmission of baseband signals...................................... 7
2.1 Typical analog to digital conversion process........................................ 10
2.2 Quantized images (a, b, c, and d).................................................. 11
2.3 Original and Quantized image (a, b, c, and d)....................................... 14
2.4 Original and Quantized voice (a, b, c, and d)...................................... 15
3.1 Uncoded- BER for different modulation schemes....................................... 19
3.2 Signal Constellation for Binary, Quadrature, 8, and 16 PSK...........................20
3.3 Un-coded BER for DPSK, 2,4, 8, 16, 32-PSK.......................................... 22
3.4 Signal Constellations for 7r/4 and 7r/4 DQPK........................................ 23
3.5 Signal Constellations for 16, 32, 64, 128, and 256-QAM............................. 25
3.6 Uncoded BER Signal Constellations for MQAM......................................... 25
4.1 Source and channel coding............................................................27
4.2 Trellis Coded Modulation............................................................ 28
4.3 Constellation doubling in TCM. A QPSK and 32-QAM signal transmitted................. 29
4.4 General trellis coded modulation (a) BPSK, code rate 1/2, output QPSK and QFSK (b) QPSK,
code rate = 2/3, output 8PSK and 8FSK (c) 8PSK, code rate = 3/4, output 16PSK and
16QAM.............................................................................. 30
4.5 Uncoded BER vs. Eb/NO in AWGN environment............................................31
4.6 Coded BER vs. Eb/NO in AWGN environment..............................................32
4.7 TCM MFSK, MPSK, DPSK, and MQAM Constellation Mapping.................................34
4.8 Multiple trellis coded Modulation with different Rates (a) Frequency Shift Keying (FSK) (b)
(QPSK) (c) 16QAM (d) 64QAM..........................................................35
4.9 8PSK Set-partitioning according to Ungerbock.........................................37
4.10 Conventional versus multiple trellis coded mod with Rate (a) 2/3 (Rate #1) and (b) Rate 2/5
(Rate #2)...........................................................................38
4.11 Conventional versus multiple trellis coded mod with Rate (a) 2/6 (Rate #1) and (b) Rate 2/6
(Rate #2)...........................................................................39
IX


4.12
4.13
5.1.1
5.1.2
5.1.3
5.1.4
5.1.5
5.1.6
5.1.7
5.1.8
5.1.9
5.1.10
5.1.11
5.1.12
5.1.13
5.1.14
5.1.15
5.1.16
5.1.17
5.1.18
5.1.19
Conventional versus multiple trellis coded mod with Rate 2/8 (Rate #5)..........39
Multiple trellis coded Modulation with different Rates Frequency Shift Keying (FSK) (b)
(QPSK) (c) 16QAM (d) 64QAM......................................................40
Original image. NASA captured this image of the Gulf of Mexico on May 24, 2010 at 16:50
UTC MOD1S............................................................................43
Block diagram of the implemented system.............................................44
Block diagram of the whole system using different scenarios......................... 45
Monitoring System operates for many different scenarios..............................46
The effect of AWGN on transmitted and received Gray-Coded 256-QAM Signal
Constellation image.................................................................47
BER to measure the performance of the channel for 256 QAM (Simulation Vs Theoretical Vs
Simulation with extra AWGN)..........................................................48
Original image after transmitted and received through the satellite. After modulated and
demodulated.........................................................................48
Original image after transmitted and received through the satellite. After modulated and
demodulated.........................................................................48
Original image after transmitted and received through satellite and adding some extra AWGN.
After modulated and demodulated....................................................49
Original image after transmitted and received through satellite and adding some extra AWGN.
After modulated and demodulated....................................................49
Received image.......................................................................50
Received noisy image More AWGN ....................................................50
Make a decision in the first time the threshold: detect oil spills..................51
Make a decision in the first time the threshold: detect: detect noise............... 51
Detecting oil leak by using RSDCT....................................................54
Detecting oil leak by using RSDCT more AWGN........................................54
The method of scanning an image......................................................55
One row is used to calculate false alarm and power curve............................55
Performance Characteristics of Sequential Algorithms using two pairs (q=0.3, p=0.075) and
four different thresholds (t =200,100,50,and 25). Detect from q to p................57
x


5.1.20 Performance Characteristics of Sequential Algorithms using two pairs (q=0.3, p=0.075) and
four different thresholds (t = 150,90,60,and 30). Detect from p to q................57
5.1.21 Results of choosing wrong thresholds.................................................58
5.1.22 Performance Characteristics of Sequential Algorithms using two pairs (q=0.3, p=0.075) and
(q=0.2, p=0.055); Detect from q to p................................................59
5.1.23 Performance Characteristics of Sequential Algorithms using two pairs (q=0.3, p=0.075) and
(q=0.2, p=0.055); Detect from p to q................................................59
5.2.1 VAD integrated in a telecommunication system......................................... 62
5.2.2 The general operation flowchart of the VAD algorithm................................. 63
5.2.3 Distributions Voice Signals.......................................................... 64
5.2.4 Actual noiseless voice signal silence + active voice............................... 64
5.2.5 Actual Voice Noisy Voice Signal...................................................... 64
5.2.6 Making the final decision in the first crossing to the threshold; detecting a change (a) from
f0(x)tof1(x) (b) from fx(x) to f0(x)................................................ 66
5.2.7 Signal-to-Noise Ratio SNR............................................................ 69
5.2.8 Power and false alarm curves Vs different thresholds................................. 71
5.2.9 Making the final decision in the first crossing to the threshold, (a) Cyber Exploits Present,
denoted H1 (b) Cyber Exploits Absent, denoted HO.................................... 73
5.2.10 Original Voice Signal.................................................................75
5.2.11 Various noisy environments........................................................... 76
5.2.12 Original signal corrupted by AWGN SNR=25,20,15,10 and 5 dB........................ 76
5.2.13 Results of adding noise to the original speech signal (5dB SNR), (a) Clean speech, (b)
Wind Noise, (c) Babble Noise, (d) Computer Fan Noise, (e) Flowing Traffic, (f) Train
Passing Noise, (g) Noisy Signal Noise: AWGN. (h) Noisy Signal Noise: Wind, (j)
Noisy Signal Noise: Computer Fan, (k) Noisy Signal Noise: Flowing Traffic. (1)
Noisy Signal Noise: Train Passing Cut............................................. 76
5.2.14 SDCT-VAD results Original Voice Signal corrupted by AWGN (SNR=25dB)............... 77
5.2.15 SDCT-VAD results Original Voice Signal corrupted by AWGN (SNR=5dB)................. 77
5.2.16 Pe's of the proposed RSDCA-VAD various environmental conditions...................... 80
5.2.17 Pc's of the proposed RSDCA-V AD various environmental conditions..................... 80
xi


5.2.18 (a) Original Signal (b) Noisy signal with SNR=15dB (c) Noisy signal with SNR=1 OdB (d)
Noisy signal with SNR=5dB......................................................... 82
5.2.19 Cyber Detection during speech activity detected periods using the CA-SDCT (a) Sequential
Test: SNR=15. (b) Sequential Test: SNR=10. (c) Sequential Test: SNR=5.......... 83
5.2.20 Cyber Detection during speech activity detected periods using RSDA-CA (a) Alarms
SNR=15. (b) Alarms SNR=10. (c) Alarms SNR=5.................................... 84
5.3.1 Model: Detect Cyber Attack using BER-Threshold.................................... 86
5.3.2 QPSK. Modulation Vs Cyber Attack using BER-Threshold (a) No cyber attack detected
(Green alarm) (b) and (c) minor warning (Yellow Alarm) (d) Cyber attack detected (Red
Alarm)............................................................................ 89
5.3.3 64QAM Modulation Vs Cyber Attack using BER-Threshold (a) No cyber attack detected
(Green alarm) (b) and (c) minor warning (Yellow Alarm) (d) Cyber attack detected (Red
Alarm)............................................................................90
5.3.4 The received noisy speech signal..................................................92
5.3.5 Two-Dimensional (2-D) Time-vs-Frequency Spectrogram.............................. 92
Fig. A Evaluate the whole updating step algorithm g(£) ................................. 96
xii


LIST OF TABLES
Table
2.1 Show the differences of the quantization step......................................13
2.2 Show the differences of the quantization step......................................14
2.3 Quality of the quantized Signal....................................................15
3.1 Abbreviation of different modulation schemes.......................................17
4.1 Coding Gain........................................................................32
5.2.1 Comparing the Starting and Ending detection time instances of the Noisy Active Voice
Messages Using the Manual and Proposed meth.......................................78
5.2.2 Pc's and PF's OF the Proposed RSDCA-VAD for Various Environmental Conditions.......79
5.2.3 Pc's Of The Proposed RSDCT-VAD, and Different VAD Approaches for Various
Environmental
Conditions........................................................................
81
5.3.1 Frequency Bands....................................................................92
xiii


1. Introduction to Digital Communication Systems
The rapid progression of satellite communication technology has made available a new
variety of services. Satellite systems are no longer dedicated solely to military transmission; they are
also offered to public, resulting in increased frequency spectrum demand. In addition, to maximize the
services offered through satellite communications, the available frequency spectrum must be used
efficiently. Reliable transmissions of data and integrated services also have gained a lot of importance
in today's information era. There is great demand for reliability and speed in the transmission of
information. Spectrally efficient linear modulation schemes increase the effective rate of information
transmission, without spectral increase. Several widely used linear modulation schemes are spectrally
efficient, simple to implement and characterized by moderately good bit-error-rate (BER) performance.
However, a tradeoff between power and spectral efficiency arises.
Power efficiency is a major issue in satellite communications. To keep the size of the HPA
(High Power Amplifier) at each terminal station somehow limited, the power consumption must be
minimized. To minimize the power required for achieving a certain BER, error control techniques may
be applied. The combination of Trellis encoding and Viterbi decoding techniques may be very helpful
in this case. Furthermore, considerable gains may be attained if various modulation and coding
schemes may be incorporated into the trellis-coded modulation (TCM) encoding. TCM was first
discover by Ungerboeck [1] who showed that, in the case of Additive White Gaussian Noise (AWGN)
transmission channel, gains of up to 6 dB were achievable without any bandwidth expansion, and that
most of the coding gain could be obtained by doubling the constellation.
Digital modulation techniques are essential to many digital communication systems, whether
a fiber system, a wireless communication system, or a satellite communication system is considered. In
the last decade, research and development in digital modulation techniques have been very active with
many promising results. There exist countless research efforts in digital communications [2], [3], [4],
[3], and [6]. Each such effort includes deployment of digital modulation techniques.
This thesis provides up-to-date information of most modulation techniques in digital
communication systems. It presents principles of most currently used digital modulation techniques
and their applications, as well as new techniques currently being developed. For each modulation
1


scheme, the following topics are covered: historical background, operation principles, bit error rate
performance (power efficiency), bandwidth efficiency, constellations, comparisons, and applications.
After the modulations and their performances in the AWGN channel are presented, system
performance is studied and evaluated when an encoder and a decoder are added.
The principal advantage of the approach deployed in this thesis is that the same
encoder/decoder may be employed for a wide variety of trellis codes. The system may generate
different trellis-coded signals using the same rate convolutional encoder. The adaptive nature of the
approach allows for lowering the coding rate while keeping the same symbol transmission rate during
deep fades. Most communication systems fall into one of three categories: Bandwidth efficient, power
efficient, or cost efficient. Bandwidth efficiency shows the ability of a modulation scheme to
accommodate data within a limited bandwidth. Power efficiency describes the ability of the system to
reliably send information at the lowest practical power level. In most systems, there is a high priority
on bandwidth efficiency. Each particular system can be designed and optimized, as dictated by the
demands of its application. In this thesis, we will evaluate different modulation techniques in terms of
their combined bandwidth, power and cost efficiency.
We specify three design stages: algorithmic, architectural and implementational. The
conceptual links between these stages represent design constraints, such as throughput, bit error rate,
bandwidth efficiency, flexibility, latency, power-savings, and complexity and so on. These constraints
determine the choice of the architectural and ultimately the implementational platforms. A comparison
is carried out between flexible designs, which decode both uncoded and TCM coded data and provide
two or three transmission rates. It is concluded that the larger number of rates is more beneficial from a
cost-flexibility viewpoint. In this thesis, the Viterbi algorithm is chosen for decoding, since it provides
a good trade-off between achievable coding gain and implementation complexity.
Towards the end of the 1970s, Ungerbock [1] addressed the issue of bandwidth expansion by
combining coding and modulation. According to him, redundancy is now provided by using an
expanded signal set and the coding is done directly on the signal sequences. In this thesis, a scheme
consisting of trellis-coded (Multiple Phase Shift Keying) MPSK, (Multiple Frequency Shift Keying)
MFSK, (Differential Phase Shift Keying) DPSK and (Multiple Quadrature Amplitude Modulation)
2


MQAM modulated signals, AWGN fading transmission channels and the corresponding Viterbi
decoder is proposed. The scheme realizes a family of different- rates codes, applicable to various
channel conditions. During poor channel conditions, TCM is good choice in these cases to be
employed. Theoretical bounds for the error performance and throughput of the proposed adaptive
scheme are derived. Simulations have also been performed to measure the performance of the scheme
for different parameter values and non-ideal conditions. It is shown that TCM results in considerable
improvement in bit-error-rate (BER) performance of MPSK, MFSK, DPSK and MQAM signals.
Under ideal conditions, gains in the range of 3 10 dB are achieved over conventional fixed rate
trellis-coded schemes.
After showing how the digital communication systems perform using various modulation
techniques and some channel coding, selected satellite and communication systems applications will
be presented. Images and speech signal processing will also be used. To evaluate systems robustness,
AWGN channel will be employed. An efficient and effective monitoring and detection of change
algorithm is introduced. An innovative satellite- detecting framework is demonstrated, including
satellite communication link configuration, satellite images transmission, enhancement and
segmentation. Finally the sequential detection of change algorithm (SDCA) is proposed to detect al
changes. A novel Voice activity detection algorithm is introduced and analyzed. Then, a new method
to detect cyber attacks using bit-error rate- Threshold BERT is developed.
This Thesis is organized into five chapters. Chapter 1 provides a review of digital communication
systems. Chapters 2, 3, and 4 describe the development, design, analysis and simulation of the different
modulation techniques considered and the TCM encoding method. In Chapter 5, some Satellite and
communication systems applications are introduced.
Chapter 1 is an introduction to digital communication systems, Information format, Data
compression, Baseband Signal, Channel Coding and Bandpass Signals.
Chapter 2 discusses signal processing methods with applications for images and voice. We
demonstrate how analog information sources can be transformed into digital sources through the use of
sampling and quantization. We also discuss the importance of baseband modulation due to its use in
short distance data communications, as well as due to it being the front end of bandpass modulations.
3


Chapter 3 covers baseband signal modulation schemes. Various different modulation techniques
and their performances are discussed, while simulation setups and results are also included.
Chapter 4 is a brief review of trellis-coded modulation. Different TCM encoder examples are
given for illustration. Applying appropriate code design criteria, optimum trellis coded schemes for
various modulation techniques are presented and the enhancement of the BER performance due to the
TCM is measured. This chapter has been dedicated to the detailed error performance analysis of the
pragmatic schemes deployed in adaptive systems, both in AWGN and Rayleigh fading transmission
channels. The details of simulation programs are explained, and the results of simulations are reported.
Simulations have been run for various parameters to take into account the effects of non-ideal
conditions. The simulation results show that the performance gain between the BER-Uncoded and the
BER-TCM-Coded is approximately 5-6 dB.
In Chapter 5, the ideas and the scope presented in this thesis will be used on real life applications.
1.1 Digital Communication Systems
Figure 1.1 is the block diagram of a typical digital communication system. The message to be
sent may be from an analog source (e.g., voice) or from a digital source (e.g., computer data).
Figure 1.1 Block diagram of a typical digital communication system
4


An information source can be either analog or digital. To ensure that the source signal is
compatible with digital processing, we transform analog information sources into digital sources via
the use of sampling and quantization. These techniques are called either formatting or source coding.
The digital sources are considered being represented in the logical format of binary ones and zeros.
Figure 1.2 exhibits the formatting and transmission of baseband signals, where,
1. Data is already in a digital format.
2. Textual information is transformed into binary digits by use of an encoder.
3. Analog information is formatted using three separate processes: sampling, quantization and
coding.
The results for all three cases are binary digits, or bit streams. The bit streams are then transformed
into pulse wave sequences via pulse modulation and the bit streams are recovered at the receiver via a
demodulator.
The analog-to-digital (A/D) converter samples and quantizes the analog signal and represents
the samples by binary sequences (bits 1 or 0). The source encoder accepts these binary sequences
(digital signal) and encodes them into generally shorter sequences. The latter process is called source
encoding or data compression; it compresses the signal by reducing redundancy, hence reducing the
transmission speed, and thus reduces the signal bandwidth. The channel encoder accepts the output to
the source encoder compressed signal and encodes it into a longer digital signal, adding error
correction bits. Additional bits are intentionally added into the compressed encoded digital signal, so
that some of the errors caused by the noise during transmission through the channel can be corrected at
the receiver. Frequently, the transmission is in a high frequency passband, the modulator thus
impresses the encoded digital symbols onto a carrier. Usually there is a power amplifier following the
modulator. For high-frequency transmission, modulation and demodulation are usually performed in
the intermediate frequency (IF). If this is the case, a frequency up-converter is inserted between the
modulator and the power amplifier. For wireless and satellite systems, an antenna is the final stage of
the transmitter. The transmission medium is usually called the channel, where, for satellite
transmissions, noise is added to the signal, where fading and attenuation effects appear as a complex
multiplicative factor on the signal. The term noise is used to represent a variety of random electrical
disturbances caused from within and outside the system sources. The channel noise normally possesses
limited frequency bandwidth, so that it can be viewed as a filter. At the receiver, virtually the reverse
5


signal processing happens. First the received weak signal is amplified (and down-converted if needed)
and demodulated. Then the added due to error bits redundancy is removed by the channel decoder, and
the source decoder recovers the signal to its original form before being sent to the user. A digital to
analog (D/A) converter is needed for analog signals.
Digital
Information
Figure 1.2 Formatting and transmission of baseband signals
The digital modulator maps the digital information sequences to corresponding analog radio
waveforms. In baseband representation, information sequences are mapped to a complex signal
constellation and then transmitted from either a single antenna or multiple antennas. The information
source, source encoder, channel encoder and modulator are collectively known as the transmitter. The
physical medium over which the signals are transferred from the transmitter to the receiver is known as
the channel. An ideal channel has no fading or other channel perturbations. The only concern for the
receiver operating on an ideal channel is the disturbance caused by the presence of thermal noise
primarily due to the receiver amplifier. The thermal base band noise is modeled as additive white
Gaussian noise (AWGN).
In a satellite communication system, the channel is usually more complicated than the simple
AWGN model. For example, the fixed-access line of a sight digital microwave radio channel is a
multi-path fading channel. In such a channel, the received signal is the linear combination of
6


components arriving via multiple channel paths reflected by obstacles such as trees, buildings or
radio signals typically travel over several uncorrelated propagation paths and arrive at different receive
antennas with different phases and signal levels.
The block diagram in Figure 1.1 is just a typical system configuration. For a multi-user
system and a multi-station system, a multiplexing and multiple access control stage is added; before
the modulator for a multi-user system case; and before the transmitter for multi-station system case.
Therefore, a real system configuration could be more complicated. In addition, the system shown in
Figure 1.1 can also be simpler if Source and Channel coding may be unnecessary. In the latter case,
only the source, modulator, channel, demodulator, amplifier, and antennas, for wireless
communication systems, are necessary. The fundamental objective of the communication system
design is the effective delivery of information from transmitter to receiver, with acceptable information
distortion, as dictated by the application. The channel model used most often is the AWGN channel. In
an AWGN channel, independent identically distributed noise samples are added to the transmitted
information symbols. The noise samples have a Gaussian distribution, i.e., the conditional density of
the channel output y given the input x is given by,
According to Shannon [7], reliable communication with arbitrarily low bit error rate (BER) in the
AWGN channel can be achieved for transmission rates below
atmospheric disturbances. If multiple transmit antennas and receive antennas are used, the reflected
e~(y-x)/2a2
(1.1)
(1.2)
where W is the bandwidth occupied by the information bearing signal, S is the signal power and is the
Gaussian noise variance.
7


2. Signal Processing
2.1 Analog to Digital Conversion
Information can be categorized in two forms: digital or analog. An analog signal is such that
its amplitude may take any value within an open interval; thus the number of possible amplitude values
is then infinite. Voice is analog and can take any number of volume levels within its dynamic-range.
Digital devices convert analog voice to a digital signal (A/D) by the process of sampling and
quantization. The analog signal is fust sampled and then quantized in a finite number of levels. Each
level is then converted into a binary sequence. For example, we may quantize voice in 16 levels, where
each of these levels can be represented by four bits. The same process can be performed for image and
video signals [8].
Nearly everything nowadays is digital. The medium is the environment that the signal travels
through. It can be air, space or various types of wires. Each medium offers its own unique set of
advantages and distortions that determine what will be used as a carrier. A signal through space, as in
satellite transmission, may require a very high frequency carrier that can overcome space and other
atmospheric losses. The carrier frequency may be otherwise light, as in optical fiber, or microwave, as
in mobile communications. Most mediums dictate the type of carrier (its frequency, amplitude) that
can propagate effectively through it and the type of distortions that the carrier will be affected by.
Wireless carriers are always analog, while wired carriers can be both analog and digital.
Communications inside a computer are examples of purely digital representations: digital data over
digital medium. LAN communications are digital data over analog medium. The AM and FM radios
are examples of analog data over analog medium. To convert an analog signal to a digital signal, the
following three steps are required. First, the signal is passed through a lowpass filter to prevent
aliasing. Second, the signal is sampled by a sample-and-hold circuit. Finally, the samples are quantized
by an analog to digital converter (ADC) in order to be represented in digital form as shown in Figure
.2.1.
x(l) Anti-aliasing Sample-and-Hold i(n) A/D Converter i(0
Analog Signa; Filter Circuit Discrete- Digital Signal
Figure 2.1 Typical analog to digital conversion process
8


2.2 Quantization (Uniform Quantization)
There are many different kinds of quantization techniques available. Many quantization
methods are explored such as the linear quantization, the nonlinear quantization, the delta modulation,
and the sigma-delta modulation. Figure 2.1 shows the block diagram of a speech coding system. The
continuous time analog speech signal from a given source is digitized by an ordinary connection of
filter (eliminates aliasing), sampler (discrete-time conversion), and analog-to digital converter (uniform
quantization is assumed). The output is a discrete-time digital speech signal. This signal is referred to
as the digital speech. The encoded digital speech data is further processed by the channel encoder,
providing error protection to the bit-stream before transmission to the communication channel, where
various types of noise and interferences can damage the reliability of the transmitted data. In this
thesis, we will use uniform quantization for both image and voice signals. The results show the effect
of increasing number of quantization levels on the quality of both voice and picture signals,
Correlation, Histogram Error, Signal-to-Noise Ratio, and Rate Distortion.
2.3 .1 Image Quantization
Figures 2.2 show a comparison between the four quantized images to the original, exhibiting
how the quality of the image changes as different quantization levels are used. It can be noticed that
some artifacts (errors) appear in the images, as the number of bits per signal sample is lowered. As the
latter number decreases, the quality of the quantized picture worsens and the artifacts increase
dramatically.
9


Figure 2.2.a Original Image Colorado Convention Center Denver
Colorado Convention Center Denver
^__________________________i___________li____________i___________i_
100 200 300 400 500 600
Figure 2.2.b Quantized Image (7 b/pel)
10


Colorado Convention Center Denver
600
600
50
100
150
200
250
300
350
400
450
500
100 200 300 400 500
Figure 2.2.c Quantized Image (4 b/pel)
Colorado Convention Center Denver
50
100
150
200
250
300
350
400
450
500
100 200 300 400 500
Figure 2.2.d Quantized Image (2 b/pel)
ll


Colorado Convention Center Denver
50
100
150
200
250
300
350
400
450
500
100 200 300 400 500 600
Figure 2.2.e Quantized Image (1 b/pel)
Table.2.1 Show the differences of the quantization step
K Quantization Levels Xmin Xmax Q Step
1 2 -254 254 508
2 4 -254 254 169.334
4 16 -254 254 33.8667
7 128 -254 254 4
The image quality noticeably deteriorates when the number of bits equals I and 2 b/pel. The figures
below show the difference between the original and the quantized images. Figures 2.3.(c and d) show
how the image quality is completely destroyed due to the wrong choice of bit numbers per signal value
representation.
12


n n
Figure 2.3.a Original Vs Quantized Image (7 b/p) Figure 2.3.b Original Vs Quantized Image (4 b/p)
n fl
Figure 2.3.c Original Vs Quantized Image (2 b/1) Figure 2.3.d Original Vs Quantized Image (1 b/p)
The same scenario is repeated again, where the information signal is voice instead of image. Table.2.2
shows the quality of the voice signal is decrease, as the number of b/sample is reduced.
Table.2.2 Show the differences of the quantization step
k Quantization Levels Xmin Xmax Q Step
1 2 -254 254 1.5857
2 4 -254 254 0.5286
4 16 -0.7928 0.7928 0.1057
7 128 -0.7928 0.7928 0.0125
For each signal, there is a point that the signal quality deteriorates drastically and becomes
incomprehensible, as shown in Table 2.3.
13


Table.2.3 Quailty of the Quantized Signal
number of b/sample 7 b/sample 4 b/sample 2 b/sample 1 b/sample
speech.au Clear A Little bit distortion Not Good So bad
music.au Clear Clear Not Good So bad
The signal quality deteriorates drastically when the number of levels is less than 16, and it becomes
completely incomprehensible when the number of b/sample equals 1 (2 Levels).
Figure 2.4.a Original and Quantized
(Speech.au) (7 b/pel)
Figure 2.4.b Original and Quantized (4 b/pel)
n
Figure 2.4.a Original and Quantized
(Speech.au) (2 b/pel)
original -quantized
_
7220 7210 7260 7260 7300 7320 7310 7360 7380 7100
Figure 2.4.b Original and Quantized (1 b/pel)
The signal (speech.au) deteriorates as the number of quantization levels decreases and it can
be recognized clearly when the number of quantization levels equal 16, as exhibited in Table.4.
14


3. Modulation Techniques
3.1 Introduction
Digital modulation techniques are necessary for many digital communication systems, as
necessitated by the information transmission medium. In this, chapter, we discuss several modulation
techniques that are applicable to digital communication systems. We present principles and
applications information of most currently used digital modulation techniques, as well as new
techniques that are currently being developed. We briefly discuss the role of modulation in a typical
digital communication system, basic modulation methods, and criteria for choosing modulation
schemes. For each modulation scheme, the following topics are covered: historical background,
operation principles, bit error rate performance (power efficiency), bandwidth efficiency, block
diagrams of modulator, demodulator, and constellation for different modulation schemes, comparison,
and applications. After presenting modulation schemes and their performances in the AWGN channel,
we discuss their performances when an encoder and decoder are added to the overall system.
3.2 Modulation Techniques
To provide an overview, we list the abbreviations and descriptive names of the various digital
modulation schemes that are listed in Table 3.1. Among the listed schemes, (Amplitude Shift Keying)
ASK, (Phase Shift Keying) PSK, and (Frequency Shift Keying) FSK are basic, while (Minimum Shift
Keying) MSK, (Gaussian Minimum Shift Keying) GMSK, (Quadrature Amplitude Modulation) QAM,
etc. are advanced schemes. The advanced schemes are variations and combinations of the basic
schemes. Constant envelope modulations such as FSK and GMSK offer not only enhanced spectral
efficiency, they also provide an inherent transmitted power advantage. All constant envelope
modulations allow transmitters power amplifiers to operate at or near saturation levels. The constant
envelope class is generally suitable for communication systems such as (Amplitude Shift Keying) ASK
and (Binary Shift Keying) BSK; however, the generic (Frequency Shift Keying) FSK schemes in this
class are inappropriate for satellite application since they have very low bandwidth efficiency in
comparison to the (Phase Shift Keying) PSK schemes. Binary FSK is used in the low-rate control
channels of first generation cellular systems. The PSK schemes, including (Binary Phase Shift Keying)
BPSK, (Quadrature Phase Shift Keying) QPSK, (Offset Quadrature Phase Shift Keying) OQPSK, and
15


(Minimum Shift Keying) MSK have been used in satellite communication systems. tt/4-QPSK is
worth special attention due to its ability to avoid 180" abrupt phase shift and to enable differential
demodulation. It has been used in digital mobile cellular systems, such as the United States digital
cellular (USDC) system. MSK has excellent power and bandwidth efficiency. Its modulator and
demodulator are also not too complex. ASK is generally not suitable for systems with nonlinear power
amplifiers. QAM has been widely used in modems used in telephone networks, such as computer
modems because it can achieve extremely high bandwidth efficiency. QAM can even be considered for
satellite systems [9].
Table 3.1 Abbreviation of different modulation schemes
Abbreviation Alternate Abbr Descriptive Name
Frequency Shift Keying (FSK)
BFSK FSK Binary Frequency Shift Keying
MFSK M-ary Frequency Shift Keying
Phase Shift Keying (PSK)
BPSK PSK Binary Phase Shift Keying
QPSK 4PSK Quadrature Phase Shift Keying
OQPSK Offset QPSK
jt/4-QPSK 7t/4 Phase Shift Keying
MPSK M-ary Phase Shift Keying
Amplitude and Amplitude /phase modulation
ASK Amplitude Shift Keying
QAM Quadrature Amplitude Modulation
We begin our discussion of digital modulation by starting with the three basic forms of digital
modulation techniques: frequency shift keying (FSK), amplitude shift keying (ASK), and phase shift
keying (PSK). In all these techniques, the transmitted information modifies a single parameter of a
sinusoidal waveform: either frequency, or amplitude, or phase. The sinusoidal waveform, called the
carrier, travels then through the corresponding medium, where the latter may be wire, air, water and
space. The transmission medium generally introduces corruptions to the traveling sinusoidal
waveform, and thus to the transmitted information. Below, we discuss each of the above three basic
modulation techniques and their performance.
16


3.3 Frequency Shift Keying
We first describe the binary FSK scheme. Then, we generalize it to the M-ary FSK (MFSK)
3.3.1 Binary FSK Signal
In the most general form of FSK, the frequency of the carrier is modified to represent the
transmitted information. In other words, the binary FSK scheme uses two sinusoidal signals
possessing two different frequencies to represent bits 1 and 0. Bit 1 is transmitted by a sinusoidal
carrier of one particular frequency, while, to transmit bit 0, the frequency of the carrier changes to a
different specified frequency. In particular, bits 1 and 0 are respectively represented by the waveforms
SI(t) and S2(t), below.
5i(t) = A cos(2rr/it + ) ; kT < t < (k + 1)7\/or 1 (3.1)
S2(t) = A cos(2n f2t + 0) ; kT < t < (k + l)T,for 0 (3.2)
Where and d> is the initial phase at t = 0, A is the amplitude and T is the bit period .
3.3.2 MFSK Signal
In M-ary FSK modulation, the information binary data stream is divided into n-tuples of
n = log2M bits. We denote all M possible n-tuples the M distinct messages: /; = 1,2, ...,M. M
sinusoidal waveforms, with M distinct frequencies, represent then, each of the Af messages. The
waveform for the ith message is:
Si(t) = A cos(2nfr+ 0J ; kT < t < (k + 1)T, for lt (3.3)
where T is the per message period corresponding to the transmission of n bits. If the initial phases are
the same for all i, then the scheme is called coherent. As with the binary case, we can always assume
transmitted signal set may be non-coherent, where the demodulation sheme must be then non-coherent.
17


Figure 3.1 Uncoded- BER for different modulation schemes
3.4 Binary Phase Shift Keying (BPSK)
In BFSK, bits are represented by two sinusoidal waveforms possessing two distinct phases.
Typically, these two phases are 0 and n. Let us denote the two binary sinusoidal representations
andS2. Then,
5x(t) = A cos(27rft + 0) ; kT 52(t) = A cos(27rft + 180) ; ; kT < t < (fc + l)T,for 0 (3.5)
In the above representation, the information bit is represented by a 180 degrees phase-change in a
sinusoidal signal and the two bit representations are then called antipodal. This phase choice
corresponds to a signal design that minimizes the probability of error in AWGN transmission, inducing
a correlation coefficient of (-1).
18


In BPSK, the unit circle is 2-quantized. As a generalization, M -quantized levels of 2n may be
deployed, to create a variety of PSK modulation schemes. Given Af, let i be a number from 1 to M.
The allowed phases are then given by the following modulating angles.
0i =
2 ni
(3.6)
In the above expression, M stands for the order of the modulation. M = 2, results in a BPSK scheme, M
= 4 represents a QPSK scheme, and so on. We note that all PSK signals may be graphically
represented by a signal constellation in a two-dimensional coordinate system. The following diagram
shows some of the MPSK modulation schemes and their constellations.
Const lban Cor2PSK
CandtfeaiBn fat QPSK
Figure 3.2 Signal Constelation for Binary, Quadrature, 8, and 16 PSK
19


As compared to the BPSK, for M larger than 2, the MPSK decreases the signal bandwidth. Indeed, in
BPSK, a single bit is represented by a single sinusoidal waveform, while such a waveform represents
n = log2 M bits in MPSK.
3.4.2 Quadrature PSK
Among all the MPSK schemes, QPSK is the most frequently used because it does not suffer
from Bit Error Rate (BER) degradation, while the bandwidth efficiency is sufficiently increased, as
compared to the BPSK. Other MPSK schemes, on the other hand, increase bandwidth efficiency at the
expenses of BER performance [9], In this section we will study QPSK in great detail.
If the transmission rate of the symbols is the same in QPSK and BPSK, it is intuitively
obvious that BPSK transmits data half as fast as QPSK does. At the same time, we observe that the
distance of adjacent points in the QPSK constellation is less than that of the BPSK. In comparison to
the BPSK, this causes demodulation problems, where the distinction of symbols worsens and the per
symbol error performance thus degrades and so consequently does the bit error rate. However, as
shown in the figure below, the bit error probability remains the same.
Over the last two decades, communication networks have evolved from the analog signal-
transmitting telephone and radio transmissions to the modem digital communication systems. These
digital networks are required to carry large amounts of data at high rates. The need for increased
bandwidth has caused an evolution in the media from wires and air. Service providers must guarantee
data integrity to their clients, where bit-error rate measurements are used to measure error resistance
performance. The pertinent criterion is called bit-error rate (BER) and it plays an important role in
measuring the quality of service delivered by a network. The BER definition is given below.
BER =
Number of changed bits
Total Number of bits
(3.7)
Under ergodicity conditions, the BER converges to the bit error probability induced by the channel of
transmission, and it may be then monitored to detect changes in the quality of the communication link.
The BER performance is affected by many factors such as power, noise, or the deployed modulation
method. If we assume specific system power, known noise and given modulation technique, we can
20


predict how the BER performance. The BER measurement is not complicated; it just requires the
transmission of a bit stream through the communication system and the comparison between input and
output bits. Noise is the main enemy of BER performance. The noise introduced by an atmospheric
transmission medium is frequently described with a Gaussian probability density function, while the
signal path is usually described with a Rayleigh probability density function. A Rayleigh, or fading,
signal path is not noise in the intuitive sense of the familiar hissing sound of white noise, but it is
a random process that is analyzed in the same manner as Gaussian noise. Without going into details,
the mathematical representations of these functions represent a system model and allow for the system-
analysis and performance prediction. As compared to PSK, the QPSK increases the data rate at the
gain of decreased BER.
As compared to the QPSK, in 16-PSK, the signal space is subdivided into smaller regions. 16
sinusoidal signals or symbols are then available, where each symbol represents 4 bits. The bit rate is
now four times that of the BPSK for the same symbol rate. Figure 3.3 shows the 16-PSK signal at
various stages during modulation.
Figure 3.3 Un-coded BER for DPSK, 2,4, 8, 16, 32-PSK
21


3.5 n /4-QPSK a variation on both QPSK and 8-PSK
n/4-DQPSK has been designated as the American standard of the second-generation cellular
mobile communications. It is a variation of the QPSK that mimics 8-PSK. Like QPSK, 7T/4-QPSK
transmits two bits per symbol. So only four carrier signals are needed but this is where the twist comes
in. In QPSK we have four signals that are used to send the four two- bit- length symbols. In n/A-
QPSK, we have eight signals, instead: every alternate symbol is transmitted using a tt/4- shifted
pattern of the QPSP constellation. As shown below, a symbol at (45, 135, 225, -45) uses a signal
on this path, while, even if the pattern remains unchanged, the next symbol uses path
(0, 90, 180, 270). Thus, a phase shift always occurs, even when adjacent symbols are identical.
The constellation diagram looks similar to the 8-PSK. Note that a 8-PSK constellation can be broken
into two QPSK constellations as show below. In 7T/4-QPSK, one symbol is transmitted on the first
type constellation and the next one is transmitted using the second type constellation. Even though the
constellation looks like 8-PSK, on the network analyzer, this modulation is strictly a form of QPSK
with same BER and bandwidth. Although the symbols move around, they always convey just 2 bits per
symbol.
Constellation for (pi/4>DPSK Constellation for (pifl)-DPSK
Figure 3.4 Signal Constellations for ir/4 and rc/8 DQPK
22


3.6 Quadrature Amplitude Modulations
At this point, all the passband modulation schemes we have studied, MFSK, MPSK, and
DPSK are constant envelope schemes. The constant envelope property of these schemes is especially
important to systems with power amplifiers which must operate in the nonlinear region of the input-
output characteristic for maximum power efficiency. Such are the satellite transponders. For some
other communication systems, constant envelope may not be a crucial requirement, whereas bandwidth
efficiency is more important. Quadrature Amplitude Modulation (QAM) is such a class of non-
constant envelope schemes that can achieve higher than the MPSK bandwidth efficiency, for the same
average signal power. QAM is widely used in modems designed for telephone channels. The telephone
circuit modem standards are all based on various QAM schemes ranging from uncoded 16-QAM to
trellis coded 128-QAM. The research of QAM applications in satellite systems, point-to-point wireless
systems, and mobile cellular telephone also systems have been very active.
Conti item to 1604M
CwtiM*nito32CM
Corattltajori tofeiQAM
CvniHtMilu 128QAM
Contietono to QAM
15 -10 -5
Figure 3.5 Signal Constellations for 16, 32, 64, 128, and 256-QAM
23


QAM is a combination of amplitude and phase modulation and its development may be
attributed to the following logic: In MPSK schemes, signals have the same amplitude but different
phases. It may be natural to consider both amplitude and phase modulations (QAM) as the next
development step, where the transmitted signals are:
Sj(t) = Ai cos(2nft + d>j) ; i = 1,2,..., Af (3.8)
where At is the amplitude and 0t is the phase of the ith signal in the M-ary signal set. Similarly to
the MPSK, a geometric representation called constellation is a very clear way of describing a QAM
signal set. A QAM signal is represented by a point or vector, or phasor. The two axes sometimes are
simply labeled as I-axis and Q-axis. Figure 3.6 shows different types of QAM constellations.
Figure 3.6 Uncoded BER Signal Constellations for MQAM
24


4. Channel Coding
4.1 Trellis Coded Modulation (TCM)
4.1.1 Trellis Coded Modulation Background
Trellis Coded Modulation (TCM), introduced by Ungerboeck [1], is a very effective method
for reducing the required power without any increase in the bandwidth requirement. The innovative
aspect of TCM is the concept that encoding and modulation should not be treated as separate entities,
but rather, as a unique operation. We usually consider coding and modulation as two separate stages in
a communication connection, while in TCM the two stages are united. Trellis Coded Modulation
(TCM) is a relatively complex concept, especially due to the nonlinear nature of its operation. TCM
belongs in the class of convolutional codes and has been applied for transmissions through telephone,
satellite and microwave digital radio channels, where coding gains of the order of 3-6dB may be
obtained with no loss in bandwidth or data rate [8] and [9], Generally, the Hamming distance between
binary representations of two signals does not possess a direct translation to their distance in the
signal/symbol space (after modulation). It may be concluded, therefore, that the Hamming distance is
not the correct distance representation between different symbols. On the other hand, the geometric
distance between the signals or their Euclidean distance (ED) may be the appropriate measure.
Figure 1.1 is a block diagram of a communication link. Here, we assume that the output of the
source is time-discrete. The output of the source is first encoded for error correction after transmission
through a distortion-inflicting medium. This encoding induces the addition of some redundant
symbols to a group of raw source symbols. The encoded data stream is then modulated and sent over
the channel (using Trellis).
The objective of modulation is the transformation of encoded (for error correction) symbols
into a signal-form that is suitable for transmission through the available channel. For AWGN channels
with fading, noise is added to the latter signal-form, while other artifacts are also inflicted upon it. At
the receiver, the received noisy signal is first demodulated and the encoded symbols are recovered with
some error. Then, the decoder (Viterbi Decoder) attempts to correct the errors using the extra
information available due to the redundancy bits added by the channel encoder (Trellis Encoder).
25


Figure 4.1 Source and channel coding
As of the results in the information theory of Shannon, the best system performance can be
obtained when codes for long message sequences are designed, as long as the transmission rate
remains below the channel capacity. The receiver decides then among different long message/symbol
sequences rather than making per symbol decisions. The induced probability of error is inversely
proportional to the length of the symbol sequences decoded. TCM follows the Shannon principle: At
the digital level, it collects a block of bits and encodes them by inserting extra error-correcting bits.
The so extended binary sequences are then modulated for conversion to a analog form via the use of a
sinusoidal carrier.
4.1.1 Fundamentals and Concept of TCM
The functions of a TCM consist of a Trellis code and a constellation mapper as shown in
Figure 4.2. TCM combines the functions of a convolutional coder of rate R = k / k + 1 and a M-
ary signal mapper that maps M = 2k input points into a larger constellation of M = 2k + 1
constellation points.
26


Trellis Code
Kbits f
-------:
Constellation Mapper
K+i bits f
Convolutional 1 Encoder i Rate k/(k+l) 1 L. MPSK
i 1 . MASK

1 H MQAM
j MFSKorDPSK
T
Carrier!
Symbols
Figure 4.2 Trellis Coded Modulation
For k = 2, we have a code of rate 2/3 that takes a 4PSK signal (M = 4) and produces a 8-PSK signal (M
= 8). Thus, instead of expanding the bandwidth, as the signal transforms from 4PSK to 8PSK, it
doubles the constellation points. The same scenario applies if we select k=4, and we choose a code
rate 4/5 that takes a 16-QAM signal (M=16) and produces a 32-QAM signal (M = 32). Thus, instead of
expanding the bandwidth, as the signal transforms from 16-QAM to 32-QAM, the system is upgraded
to one with a larger number of constellation points as shown in Figure 4.3 below.
27


Constellation for QPSK
Constellation for8PSK
In-Phase
In-Phase
Constelation (or 16QAM
ConstoUation for 32 QAM
3
2
1
V
lo
3
o
-1
-2
3
-3-2-10123
In-Phase
Figure 4.3 Constellation doubling in TCM. A (QPSK and8PSK); (16 and 32-QAM) signal
transmitted
4.1.3 Mapping and Trellis Diagram
The figure above shows Trellis Coded Modulation with different code rates. The coding
adds just one extra bit to the symbol bit size. The symbol size increases from k bits to k + 1 bits which
means that the constellation size doubles. Notice that the code rate may not be only 1/2, 2/3, and 3/4,
but it could be anything: 3/8, for instance. Assuming that the original signal is BFSK, then, a TCM
encoder will produce a QFSK, a QFSK will become 8FSK, and, if 8PSK was selected, a 16QAM
signal will result.
28


Trellis Code Constellation Mapper
1 | 1
| J "l.; From BFSK to 4FSK
' i 1 4FSK i
04 K-l ! Convolutional 1 2 bib i 1
J 1 n Encoder i i - i
i Rate 1/2 | i 1
i Oi OS i i 1 i 4PSK i i hi From BFSK to QPSK
i 1 i i
T
Carrier:
1 Trellis Code | Constellation Mapper ;**V
b| 1 K-2 ; Convolutional 1 1 3 bib -hJ i 8FSK 1 i H* * From 4FSK to 8FSK
1 1 i Encoder Rate 2/3 1 i i i i ; i -*i i i 8FSK 1 i i * * From QPSK to 8PSI
T
Carrier;
#
* *
K-3 f
Trellis Code
Constellation Mapper

1 1 1
Convolutional Encoder Rate 2/3 4 bib 1 1 1 y
i 1 i ii

1
16PSK
16QAM
* *
J*
rom 8FSK to 16PSK
* *
* * 1- *
From 8PSKto 16QAM
*
Carrier i
Figure 4.4 General trellis coded modulation
(a) BPSK., code rate 1/2, output QPSK and QFSK (b) QPSK, code rate = 2/3, output 8PSK and 8FSK
(c) 8PSK, code rate = 3/4, output 16PSK and 16QAM
29


As constellation expands while the signal energy is kept the same and the distance between the
symbols decreases, Let us assume that we are given some amount of power and we want to calculate
the needed signal BER. Figure 4.5 shows the improvement in BER and power performance if TCM is
deployed.
Figure 4.5 Uncoded BER vs. Eb/NO in AWGN environment
Let us assume the transmission of a BER of 10-4 encoded QPSK signal. If this power level is not
available, another option is to add a code of rate 2/3 to reduce the BER and thus the subsequent
Eb/NQ requirement. However, another problem arises then. If we keep the same bit rate for the
information bits and allow the coded bit rate increase to accommodate the overhead bits, then the
bandwidth requirement will increase by an amount inversely proportional to the code rate increase.
Thus, addition of coding increases the bandwidth by 3/2. If bandwidth change is not allowed, then the
information rate will have to decrease by the same proportion.
From Figure 4.6, we observe that due to the use of TCM, 4-6 dB can be saved. For example,
let us pick QPSK and compare Figures 4.5 and 4.6 It can be observed that due to the use of TCM,
with a BER of 10~2,a 5 dB gain could be achieved, meaning that less power is required for the
delivery of the same BER. From the same figures, we also notice that, with a BER of 10-2, instead of
QPSK, we can instead use 16-QAM with the same power, at the expense of information rate increase.
30


Table 4.1 Coding Gain
Modulation scheme Coding gain at 10-2 over Uncoded System Coding gain at 10-2 over Coded System
QPSK 4dB 1.5dB
16QAM 8.2dB 2.3dB
32QAM 1 ldB 5.5dB
256QAM 16.3dB 11.5dB
Figure 4.6 Coded BER vs. Eb/NO in AWGN environment
A rate of a 2/3 convolutional encoder is used as an example to analyze the trellis diagram and mapping
method. Figure 4.7 illustrates a convolutional encoder. The input of this TCM encoder is a 4-bit
symbol, and the output is a 6-bit symbol. The output of this encoder consists of n bits. These bits are
used to choose one of 2n partitions in the constellation. This means that the constellation has been
31


partitioned into 2n subsets. After the constellation is chosen, we need to select the modulation
technique. Figure 4.7 shows some the results of this process.
4.1. Viterbi Algorithm
The TCM encoder is illustrated using a trellis whose branches are associated with transitions
between encoder states and codeword transmitted over the channel. The primary task of the TCM
decoder is to estimate the path that the codeword sequence traverses through the trellis. In this manner,
TCM decoder is a reverse process of TCM encoder. In addition to the convolutional decoding, the de-
mapping algorithm is a reverse function of the mapping logic function and the differential decoder
performs the reverse function of the differential encoder. The decoder algorithm used in this thesis is
based on the Viterbi algorithm.
Andrew Viterbi proposed an algorithm in 1967, to decode convolutional codes and this
became the Viterbi Algorithm [10]. This algorithm is an application of dynamic programming that
finds shortest paths, (maximum likelihood sequences) widely used in solving minimization problems.
A critical feature of this algorithm is the complexity of the decoding process grows linearly with the
number of symbols being transmitted, rather than exponentially with the number of the transmitted
symbols. The Viterbi algorithm uses a metric and tracks this metric for several trellis paths at once.
The path with larger metric is dropped when it merges with another. In hard-decision Viterbi decoding,
this is done using the Hamming distance as a metric. In TCM the decoding is done with soft-decision
algorithm and Euclidean distance is used as the metric. The objective is to track n possible sequences,
keep track of cumulative MSEDs. When paths merge at a state, follow only the one with the smallest
metric.
32


4.2 TCM MFSK, MPSK, DPSK, and MQAM Constellation Mapping
Figure 4.7 The block diagram of a coded-modulation system.
4.3 Set Partitioning
From now on, two-dimensional orthogonal (//Q) signal constellations, such as MPSK or
MQAM, are assumed. Also, transmission rates R are now expressed in bits per two- dimensional
channel use. The error rate performance of the trellis-coded signals in the presence of additive
Gaussian noise can be evaluated for convolutional codes by summing these error event probabilities to
obtain a union bound on the first-event error probability. At high SNR, the first error probability is
well approximated as
33


Pe a Nfed Q
(4.1)
/ Mfed'
lJ2N,
Where Nfed denotes the number of signals sequences with distance Dfed that diverge at any
state and remerge at the state after one or more transitions. Recall that increasing dmin lowers the
BER. In trellis-coded modulation (TCM), this increase is achieved by partitioning a constellation into
subsets. This split is done according to Ungerbocks rules Figure, every split of a uniform 16-QAM
and 8-PSK constellation increases dmin in the subset by a factor of V2. If the lattice at the root has
distance 1 between constellation points, the distance in the subsets at the lowest branch is 2y[2. This
distance is the ultimate limit on how well a TCM code can perform.
O O
0*0*
0*0
o o
o o
o o
0*0*
o o
dc=2dA
oooo
o o
o o o o
o o
0*0
o o o o
o o
o o o o
o o
o o o o
o o
o o o o
o o o o
o o
o o o o
o o
dD = 2dc
oooo
0*00
o o o o
o o o
o o o o
o o o
o o o o
o o o
o o o
o o o o
0 0*0
o o o o
0 0*0
o o o o
o o o
o o o o
o o o
o o o o
o o o
o o o o
o o o
o o o o
o o o
o o o o
o o o o
0 0*0
o o o o
o o o
o o o o
O O O
O O O O aF
o o o
2 dD
oo oo
oo oo
00 oo
0001
OO O O
0*00
oo o o
oooo
oo o o
oo o o
oooo
oooo
oooo
oooo
oo o o
oo oo
oooo oooo
oooo oooo
oooo oooo
oooo oooo
oooo oooo
oo oo oooo
oooo oooo
oooo oooo
oooo oooo
oooo oooo
oo o o oooo
oooo oo oo
oooo oooo
oooo oooo
oooo oo oo
oooo oooo
oooo
oooo
oooo
oooo
oooo
oooo
oooo
oooo
oo o o
oooo
oooo
oooo
oooo
oooo
oooo
oooo
Figure 4.8 16QAM Set-partitioning according to Ungerbock.
34


Ungerboeck [1] showed that using (Euclidean distance) ED as the metric in systems
employing Trellis coding and Viterbi decoding, and designing codes to achieve maximum ED between
all possible sequences of channel signals, leads to significant coding gains without any sacrifice in
bandwidth or the effective information rate. In fact, it was shown [ 1 ] that nearly all the possible gain in
performance is achieved by doubling the size of the signal space. This led to the design of rate
n/n + 1 codes with signal spaces of size M = 2n + 1.
Figure 4.8 shows a 64-QAM constellation, which is divided into 16 subsets with 2 signal points per
subset. This division stems from successive splits as in Figure 4.8, and the bit-to-constellation point
mapping is derived from these splits.
If the at the lowest distance is 2a with (Square-Euclidean Distance) SED equal to 4a2. After that the
same processes is repeated until the final stage (Last Level), 4a2 -* 8a2 -> 16a2 - 32a2 and so on,
which means we are doubling each time. As shown in figure 4.8.
The figure below shows how the 8 points of 8PSK are successively portioned into disjoint cosets such
that the SEDs are increasing at each level. There is a total of four partitions, counting the first
unpartitioned set. At the top-most level, the MSED (Minimum Squared Euclidean Distance) is 0.586.
At the next level, where there are only four points in each of the two cosets, the MSED has increased
to 2.0 and at the last level, the MSED is 4.0. Each subset is also called a coset and by the lattice
terminology, we can show the partition with its coset generators in this way.
In general, if M denotes the size of the master constellation, there are log2 M 1 possible
splits to finally produce M/2 subsets with 2 signals. Of course, one could split one more time to
arrive at one signal per subset. This transforms the set-partition problem back to mapping code
symbols efficiently to constellation points. However, it is not considered a set-partition code in the
traditional sense, and will not be discussed further.
Each Version of TCM as created by Ungerboeck, requires a different rate code. It is assumed that for
each input sequence of n bits, the trellis encoder produces a sequence of n + 1 coded bits which are
then mapped into a single symbol chosen from an M-ary alphabet where M = 2n+1. The modulation
35


sets from which these M-ary symbols were chosen were typically multiple-frequency-shift-keying
(MFSK), multiple-phase-shift-keying (MPSK) or quadrature amplitude modulation (QAM). In
particular, they proposed an encoder with b binary input bits and s binary output symbols, which were
mapped into k-Mary symbols (k is referred to as the multiplicity) in each transmission interval where s,
k, and M obey the relation s = k log M.
d;--2
d; = 4
o o
I
r
T
*
o .
I
<-*- I
i
T
1
+ ...
k d?= 0584 1 *
t
o | . i
-t T
O 1 0
T -4
O ' 0 1 e i
j 4-- T
e e l / r
T
4 4
1 * * *
- T +<*- 7 4T""<'

i "
T
Figure 4.9 8PSK Set-partitioning according to Ungerbock.
The main concept in multi-dimensionality is increasing the number of symbols created in one
processing period. The transmitted symbols are generated together and this co-generation creates
dependence and allows better performance. The term multi-dimensionality does not mean anything
other than a form of multi-processing. The advantage of multi dimensional TCM are so many for
instance we can transmit fractional information rates. Instead of the effective code rate being 2/3 as it
is in 1 x 8PSK, here it can be higher. We can use code rates like 2/5, 2/6, 2/7and 2/8. Perhaps the most
significant aspect of M-TCM with regard to its application in a fading environment is its ability to
provide improved diversity (the rate of descent of error probability with average bit energy-to-noise
36


ratio) on such a channel. Figures 4.10-4.12 show the whole process of mapping two input bits to 2+n
bits using different convolutional encoder rates. The performance of each rate versus different
modulation techniques such as Multidimensional frequency shift keying (M-FSK), Multidimensional
phase shift keying (M-PSK), and quadrature amplitude modulation M-QAM is shown in figure 4.13.
Figure 4.10 Conventional versus multiple trellis coded mod with Rate (a) 2/3 (Rate #1) and (b) Rate
2/5 (Rate #2)
Figure 4.11 Conventional versus multiple trellis coded mod with Rate (a) 2/6 (Rate #1) and (b) Rate
2/6 (Rate #2)
37


Figure 4.12 Conventional versus multiple trellis coded mod with Rate 2/8 (Rate #5)
It can be distinguished from the results that as the rate increase the shape of the bit-error rate
BGR enhances more, but the disadvantage here is that the bandwidth also increases. Therefore, we
should know what the requirements of our system are during the design process to make the best
selections of the needed rate.
EbNo EbNo
38


8ER
EbNg EbNo
Figure 4.13 Multiple trellis coded Modulation with different Rates
(a) Frequency Shift Keying (FSK) (b) (QPSK) (c) 16QAM (d) 64QAM
39


5. Communication Systems Applications
5.1.011 Spill Detection by Satellite Image using Sequential Detection of Change Test
Oil spills on the sea surface might happen without any previous caution and are seen relatively often.
Efficient and effective oil spill monitoring and detection accelerates response time, thus minimizing
remediation costs and limiting dangerous impact to the environment. An innovative satellite-based oil
pollution detecting framework is demonstrated in this section, including satellite communication link
configuration, satellite images transmission, enhancement and segmentation. Finally the sequential
detection of change algorithm (SDCA) is proposed to detect oil spills on the ocean surface from the
enhanced remote sensing data. MODIS images of the Gulf of Mexico accident from NASA between
May and June 2010 are adopted and the results of this research show that the proposed algorithms can
effectively distinguish the spills covering vast areas of the marine environment even with severe
additive noise, and have good separation properties against complex signatures, e.g. the vicinity to the
irregular coast or foggy and cloudy weather conditions.
5.1.1 Introduction
Large scale oil pollution is one of most universal problems in the ocean. For instance, the oil
spill transportation is increasing every year, and it is hard to detect because the amount of oil spill from
these see ships is not recognizable. Therefore, the risk of large-scale oil pollution catastrophe increases
substantially during the coming years. Various efforts to manage the spill with controlled burning,
dispersal and plugging the leak have so far been ineffective. As a result, the long term influences on
fisheries, wildlife, human health, and tourism would occur [11] and [12]. The idea behind this section
started actually after April 20th, when Transocean Ltd. reported an explosion and subsequent fire
aboard the semi-submersible drilling rig Deepwater Horizon. This has led to the largest oil spill in
American history and that because of three measure breaks in the connected pipe. No one knew for
cretin how much oil has been released, but there is no disagreement that the spill was massive.
Usually, oil spills occurred at the ocean service, at this time it initiated at the ocean floor and rising up
to the service of the water. As the oil reaches the service of the water, it begins to spread and moves
duo to the winds in random directions. Therefore, it is very important to detect these oil leaks as soon
as possible. As with any catastrophic spill it is difficult to predict the extent of the damage. How much
40


oil has been spilled, and how far will spread. After a series of failed efforts to plug the leak,
government and company officials said oil will likely continue flowing until a relief well cut off the
gusher. This event has fmished closing the leaking in middle of July, 2010.The incident has resulted in
a massive oil spill and has been announced an incident of national significance by the government.
The Deepwater Horizon Unified Command has expected that between 1.5 million and 2.5 million
gallons of oil were leaking into the Gulf of Mexico every day [22], [23], and [24]. That would mean
between 94 million and 185 million gallons of oil had leaked into the Gulf of Mexico. In this section,
the catastrophic explosion that caused an oil spill from a BP offshore drilling rig in the Gulf of Mexico
is adopted. The deployment real-time oil spill detection algorithm will facilitate and largely simplify
the cleanup process. In the effort to locate oil spills, satellite images are utilized. Such images are
valuable, since they cover large areas. A sequential detection of change algorithm (SDCA) is
introduced for the detection of oil spills in the Gulf of Mexico. Additive White Gaussian Noise
(AWGN) is superimposed, however, for the study of the SDCAs robustness against noise present in
satellite channels. As will be explained in detail later in this section, the proposed SDCA algorithm is
robust, allowing for effective oil spills detection even within the near-beach noisy environment.
The section is organized as follows: In Section 5.1.2, we explain our approach and present the source
of the data and the steps involved in our oil spill detection methodology. In Section 5.1.3, we discuss
how the image may be transmitted using a 256-QAM modulation technique, while the transmission is
through a noisy and fading satellite channel. In Section 5.1.4, we discuss some image processing
components, such as image enhancement and segmentation. In Section 5.1.5, the deployed sequential
detection of change algorithm is discussed and experimental simulation results are presented.
5.1.2 Data Source and Methodology
5.1.2.1 Study Area
The study area is located aboard the Deepwater Horizon, a drilling rig working on a well
for the oil company BP one mile below the surface of the Gulf of Mexico. It lies between longitude
-89.0000 and +29.0000 latitude. On April 20th, 2010, Transocean Ltd. reported an explosion and
subsequent fire on board the semi-submersible drilling rig Deepwater Horizon as shown in Figure
41


5.1.1. This has led to the largest oil spill in American history. The incident has resulted in a gigantic
oil spill and has been declared an incident of national significance by the government. For more
information one may look into references [22], [23] and [24].
5.1.2.2 Data Source
NASA's Aqua satellite captured these images of the Gulf of Mexico between May-June 2010
using its Moderate Resolution Imaging Spectroradiometer (MODIS) instrument. A number of images
are handpicked each day and posted on [14] as soon as possible after data acquisition. The satellite
may also observe real-time sections, to view the latest MODIS imagery available, within minutes after
it is automatically processed by the MODIS Rapid Response System [13] and [14].
Gutf of M*ico
UASA ' irf th* Gulf & My ?4, JO'Q M If 91 l.fTC MOOT9
Figure 5.1.1 Original image. NASA captured this image of the Gulf of Mexico on May 24, 2010 at
16:50 UTC MODIS
5.1.23 System Operation
In this section, we develop a real-time oil spill detection process depicted by the system in
Figure 5.1.2. The detection process follows the steps below.
1) Images of the interested region are captured via any of the scenarios depicted in Figure 5.1.3.
2) Images are transmitted to the ground receiving station via satellite channels. The data will be
corrupted by additive White Gaussian Noise and some fading loss.
42


3) Depending on the scenario used, different modulation techniques are applicable, before
transmission. In all cases, Bit Error Rate (BER) will be calculated to measure the performance of the
link.
Fig 5.1.2 Block diagram of the implemented system
4) The ground receiving station is a Very Small Aperture Terminal (VSAT), receiving data from all the
different satellites that may cover the area of interest.
43


5) At the receiving end, several image processing techniques, such as image enhancement and
segmentation, are used in this study, to remove any noise and artifacts. This step is crucial to the
objective of correct oil spill detection.
6) The detection of change sequential algorithm is deployed for the real-time detection of oil spills.
7) The results from the deployment of the algorithm in Step 6 separate the oil from the water. In
particular, this step will help the analyst to make the final decision.
5.1.3. Satellite Communication Link Configuration
5.1.3.1 Different scenarios:
The significant difference between the three scenarios lies in the levels of superimposed to the
transmitted images noise and fading. The applicable scenario should be known to the analyst for the
subsequent deployment of the appropriate image processing techniques.
Figure 5.1.3 Block diagram of the whole system using different scenarios
44


Scenario #1:Image tracking and monitoring directly from the satellite.
Scenario #2:The boat takes multi-images from the interested area around it and transmits them via a
Local Area Network (LAN) to a Wide Area Network (WAN) (such as a VSAT). Then, from the boats
VSAT, the images are transmitted to the ground receive station that processes them and detects the
possible presence of oil spills.
Scenario #3: A helicopter takes multi-images from the interested area and transmits them to the
ground receive station for processing and subsequent oil spill detection.
I-------------------
~ ~l
Tx-Antenna
t
Modulation
' t
Capture images
sequentially
SceriMrio #3
Helicopters images
--------f---------
£
AWGN +
Fading Channel
Capture images
sequentially
Scenario #2
Booths images
Capture images
sequentially
Scenario #1
_______1_____
AWGN +
Fading Channel
Image
Enhancement
I
Segmentation
*
Sequential Test
Decision
Figure 5.1.4 Monitoring System operates for many different scenarios.
5.1.3.2Modulation, Constellation, and the affect of AWGN
Figure 5.1.5 shows the constellation of a Gray-coded 256- Quadratic Amplitude Modulation
(QAM) transmitted and received image signal. Notice that the received signal image constellation plot
does not appear exactly like the transmitted signal constellation. The reason from this is that the
45


received signal constellation will generally have a small cluster of points around the 256 exact points,
as a result of superimposed transmission channel noise.
As shown in Figure 5.1.4, excessive White Gaussian Noise (AWGN) results in 256 clusters of points
whose positions are far away from the exact constellation points.
Received and Transmitted Signal
Received and Transmitted Signal
15
10
. 5
I 0
TO
3
-5
-10
-15
+ + + + +
+ + + + +
+ + + + +
+ + + + +
+ + + + +
+ + + + +
+ + + + +
+ + + + +
+ + + + +
+ + + + +
+ + + + +
+ + + + +
+ + + + +
+ + + + +
+ + + + +
+ + + + +
Received Signal
+ Transmitted Signal
+ + + +
+ ++ +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + + +
+ + +
+ + +
+ + +
+ + +
+ + +
+ + +
+ + +
+ + +
+ + +
+ + +
+ + +
+ + +
+ + +
-15 -10 -5 0 5
In-Phase
10 15
20
15
10
TO TO 0
o -5
-10
-15
-20
Received Signal
+ Transmitted Signal
++++4-+++++++++++
++++++++++++++++
++++++++++++++++
++++++++++++++++
++++++++++++++++
++++++++++++++++
++++++++++++++++
++++++++++++++++
++++++++++++++++
++++++++++++++++
++++++++++++++++
++++++++++++++++
++++++++++++++++
++++++++++++++++
++++++++++++++++
-20
-10
0
In-Phase
10
20
Figure 5.1.5 The effect of AWGN on transmitted and received Gray-Coded 256-QAM Signal
Constellation image
5.1.33 Measure the performance of the link using BER
In satellite digital transmission, the Bit Error Rate (BER) is an important performance. It is
defined as the ratio of the number of information bits over the number of bits transmitted, over a fixed
time period, where error correction bits are added to information bits before transmission for error
protection. The BER is used as a metric of satellite channel/modulation scheme evaluation
[19],[20],[21],
In this section, the BER of a variety of modulation schemes over an AWGN channel is
calculated for use in the oil leak detection case. The x-axis is the ratio of energy to noise power
spectral density (E^No), in dB. The y-axis represents the BER for the 256-QAM scheme. In Figure
5.1.6, theoretical results for the 256-QAM scheme are compared with simulation results before and
after the addition of AWGN.
46


Figure 5.1.6 BER to measure the performance of the channel for 256 QAM (Simulation Vs Theoretical
Vs Simulation with extra AWGN)
Figures 5.1.7-5.1.10 are subset images from the Gulf of Mexico, where the incident has happened.
From these captured images, it can be recognized that the spilled oils are close to beach and that the
weather is also somewhat cloudy, deterring the oil spill identification.
Figure 5.1.7 and 5.1.8 Original image after transmitted and received through the satellite. After
modulated and demodulated
47


All the pictures were purposely selected to include a mixture of clear water, oil silk, beach,
and clouds as well as added AWGN, to evaluate the proposed sequential detection algorithm in
unfavorable given conditions. We include additional supported images at the end of this section, to
exhibit the efficiency of our proposed algorithm on many different satellite images.
Figures 5.1.9 and 5.1.10 are the same as Figures 5.1.7 and 5.1.8; the only difference being that
there are many dots on the both images. These dots are generated because of the effect of the AWGN.
As mentioned earlier, the amount of AWGN depends on the chosen operative scenario.
Figure 5.1.9 and 5.1.10 Original image after transmitted and received through satellite and adding
some extra AWGN. After modulated and demodulated
5.1.4. Image processing
Figures 5.1.11 and 5.1.12 show how the clear water and the oil silk are enhanced. Then, it will
be shown how the images can be segmented, via the use of a certain threshold. The darker region
represents the oil silk, and the brightness region represents the clear water. It is clear from the Matlab
color bar that intensity of clear water is less than 150, while the intensity of oil silk is more than 150.
The margin of the separated threshold should be chosen within the range of 140-170. To
detect a spilled region of oil from a satellite image, everything was coded in Matlab. Both Figures
5.1.11 and 5.1.12 are converted to binary images. It can be distinguished the effect of adding the
AWGN on the binary image where Figure 5.1.11 has more white dots than Figure 5.1.12.
48


Figure 5.1.11 Received image Figure 5.1.12 Received noisy image More AWGN
5.1.5 Sequential Detection of Change Algorithm
The sequential test for detection of change was first introduced by Page in 1954 for
memory less processes. Lorden proved its asymptotic optimality for such processes. Then, Bansal and
Papantoni-Kazakos extended the test for processes with memory and proved the asymptotic optimality
of the extension for general processes possessing mixing conditions. The algorithm for Bernoulli
memoryless processes was used by Papantoni Kazakos for the detection of faulty links in networks.
In this section, it is assumed that the monitoring system will be permitted to receive many
images in the base station from the satellite and decide to make an alarm if the oil spill exits. Mainly
the system take the image and examine it, by going through the whole pixels in the image and decide
when to stop observing as well as what change in the underlying data distribution has occurred Make
an alarm showing that oil spill exists.
5.1.5.1.1. Tests for detecting a change in distribution:
Let f0(xn) and fi(xn) denote the n-dimensional density function of two well known, distinct,
discrete-time, and mutually independent stochastic processes at the vector point xn. Let it be known
that the data sequence is intitally generated by the density function f0. Let it then be possible that,
instead, at some point in time the density function fj may become active and remain so form that point
on [18]. Then, given an infinite data sequence x = {xj; i > 1} the possibilities are:
49


1) The f0 and ft change never occurs. Thus, the total sequence x is generated by the density
function f0.
2) The f0 and fj change occurs before the sequence x starts being observed; Thus, the total sequence x
is then generated by the density function fx.
3) The f0 and fi change occurs just after the datum xm;m > 1. Thus, the subsequence, x is then
generated by the density function f0 and the remaining sequence x+1 = {xk; k > m + 1), is then
generated by the density function fj [8],
In our situation, it is assumed ft represents the oil spill, while f0 represents the noise, clear
water, clouds and winds, and AWGN. Where x = {xj; i > 1} represents the observed data, scanning
all the pixels row by row. Given the finite sequence x = i > 1} and the density functions
f0(x")and fi(xn), the objective is to detect a possible f0 and fj change are reliably and quickly as
possible. To detect the possible occurrence of an f0 and fj change. Select some positive thresholds,
and scan the given image, and decide that the f0 and fj change has occurred at the first time n such
that T(xn) > 6, where
T(0) = 0
T(xn) = maxjO,T(xn J) + gnO^l&giCx*) = lgj
l h
bCxiK1))
(5.1)
The last equation operates sequentially via the use of two thresholds, 0 and S where 0 represents a
reflecting barrier and S represents an absorbing barrier [18].
Oil Spills : Decide H ,
'
n-data selected _______*-*
sequentially
----- ;
0
0
Noise : Decide H n
-----------------------p- s,
n-data selected :'
sequentially ----- /
A
0
Figure 5.1.13 Making the final decision in the
first crossing to the threshold.
Figure 5.1.14 Reflect the final decision when
it crosses the 0
50


There are going be two thresholds in this case, one detects the change from not oil spill
exists to oil spill exists has occurred, and another one to detect the opposite case. When the two
stochastic processes represented by the density functions f0 and /j are memoryless, then Let both the
non-composite hypotheses H0 and H1 be described by Bernoulli process that generates binary
sequences with elements zero and one, and it can be any kind of distribution. Let the element one occur
with probability q under H0, and probability p under H1. Let it be desirable to detect a possible change
from the Bernoulli process. We then apply the sequential test in (5.1.1) while we assume q > p Then
we can start the derivation to obtain
log
/iQi)
/o(*i)
pXi(l -p)1-*1
q*i(l q)1-xi
P
= Xj log-+(1-Xj)log
q
(1-p)
(i q)
log

Xj log
p(i q)
q(i p)
+ log
(i-p)
(i-q)
(5.1.2)
(5.1.3)
Now, select some positive thresholds. Observe data sequentially and decide that the changes has
occurred at the first instant n such that T(xn) > 5, where
T(0) = 0
T(xn) = max{0,T(xn *) + gn(xn)}
w p(i-q)v (1-p)
= l0e5w)2. x + 8{l^q)
i=l
p(l q) (1 pi
T(xn) = T(xn J) + log-----xn + log-
q(i -p)
(1-q)
(5.1.4)
Let us donate
y(q.p) =
log
(i-p)
(l-q)
log
pri-q)
q(i-p)
(5.1.6)
Then the test will take the following equivalent form.
51


T (xn) = max{0,T (xn_1) + [xn + y(q, p)]}
(5.1.7)
After changing the image to binary, we assume p represents oil spill, and q represents the
effects that look like oil spill. The Bernoulli distribution has been chosen in this case, and any kind of
distribution can be used for such case. The only difference will be the general formula of the algorithm
which means the complexity might increase or decrease. To compute both p and q, we need to train
our system using many different images. In this way, the range of each p and q can be known and
saved in our data base as the following:
1) Select many different images.
2) Compute p by counting the number of oil spill pixels divided by the number of background pixels.
3) Then, compute q using this formula =>
Number of affects that look like oil
Total number of pixels-Number of oil spill pixels
4) Results: after repeating the same procedure on many images, it has been found that the range of q is
approximately (0.2 to 0.4) whilep is between (0.045 to 0.08).
The affects of using different pairs of p and q has some affect on the results of detecting the
oil spill. A threshold on the other hand is also a very important factor; therefore, the choice of the
threshold will be determined from specific false alarm and power requirements and will depended on
where the emphasis of power Vs false alarm is placed. The false alarm and power curves will be
further discussed in section 5.1.2. Figures 5.1.15 and 5.1.16 show the steps of how the oil spill can be
extracted. It is obvious that there are a lot of small dot spots because of adding up AWGN in the
satellite communication link between the interested region and the ground receiving station. In fact, the
whole satellite communication system model was built in this section to show the effect of increasing
AWGN levels on the final decision. We thus did not apply processing operations on these dots, to
allow the evaluation of the robustness of the sequential algorithm against the AWGN.
52


If the process Hn ,(Ho), causes the stopping of the algorithm, it is decided that a q - p shift
has occurred. If, on the other hand, the Tn ,(H[). process causes the stopping, it is decided that a p -* q
shift has occurred instead. For any given (q, p) pair, every threshold t and h selection is characterized
by a pair of statistical curves: the probability of false alarm and power curves.
The flowing steps below show how the false alarm and power curves can be computed. Figure 5.1.17
and 5.1.18 show how the image can be scanned row by row.
Cofcoms
Figure 5.1.17 Method of scanning images
Figure 5.1.18 one row to calculate false alarm and power
In our case, one row has been used that contains about 600 columns using some certain
thresholds t and h to compute the probability of false alarm and power curves. The derivation of the
Markov Chain expression in [18] and [25] is used in order to generate a matrix. This matrix actually is
function of (q, p, t, s, 1, v) for q > p case, and function of (q, p, h, s, 1, v) for p > q case and the
results are given below for the q > p case, and it is similar for p -> q with different symbols.
Case # 1: Detect from q > p
Given some threshold t > 0, the mode not-oil-spill (q) to mode oil-spill (p) change monitoring
algorithm is basically characterized by two time curves: the power and false alarm curves, denoted
Pt(n) and Ot(n); respectively. Where n denotes the number of samples and t represents the case
q > p where,
54


pt(n): The probability that the q > p mode change monitoring algorithm crosses its threshold t
before or at time nt, given that the operation mode is p throughout..
Ot(n): The probability that the q > p mode change monitoring algorithm crosses its threshold t
before or at time nt, given that the operation mode is q throughout.
Case #2: Detect from p > q
Given some threshold h > 0, the oil-spill (p) mode to not-oil-spill (q) mode change monitoring
algorithm is basically characterized by two curves: the power and false alarm curves, denoted
respectively Ph(n) and ah(n). Where n denotes the number of samples and h represents the case
p > q where,
Ph(n): The probability that the p -> q mode change monitoring algorithm crosses its threshold t
before or at time nh, given that the operation mode is q throughout.
ah(n): The probability that the p -> q mode change monitoring algorithm crosses its threshold t
before or at time nh, given that the operation mode is p throughout.
In Figure 5.1.19 and 5.1.20 below, we depict the Pt(n), dt(n), Ph(n) and ah(n)curves, to
discuss qualitative behavior. We plot these curves for different threshold values and (q,p) pairs. The (q,
p) pairs (0.3, 0.075) has been considered. These two pairs are most representative of the quality shifts
that are of practical interest. For the above pairs the false alarm and test power curves have been drawn
as functions of the sample size and also using some threshold t and h. The Markov chain model has
been used for calculation the Pt(n), Ot(n), Ph(n) and ah(n) probabilities [25].
The first two curves in the Figure.5.1.19 are the pt(n) and Ot(n); detect from q -> p where
the threshold that has been used t = 25,50,100, and 200; y = 0.1674, and the stopping time for
monitoring a change algorithm when it crosses its threshold t before or at time nt equal to 262. On the
other hand, the second two curves in the Figure 5.1.20 below are the Pt(n) and Ot(n); detect from p -
> q where the threshold that has been used h = 30,60,90, and 150; y = 0.1674, and the stopping
55


time for monitoring a change algorithm when it crosses its threshold h before or at time nh equal to
413. Figure 5.1.21 shows the affect of changing the threshold on the oil spills image.
Perfomance Characteristics of Sequential Algorithms
Figure 5.1.19 Performance Characteristics of Sequential Algorithms using two pairs (q=0.3, p=0.075)
and four different thresholds (t =200,100,50,and 25). Detect from q to p.
Perfomance Characteristics of Sequential Algorithms
Figure 5.1.20 Performance Characteristics of Sequential Algorithms using two pairs (q=0.3, p=0.075)
and four different thresholds (t =150,90,60,and 30). Detect from p to q.
56


Figure 5.1.21 results of choosing wrong thresholds
The results from chosen different pairs (q=0.2, p=0.055) is shown in Figure 5.1.22 and 5.1.23.
The stopping time in each case is shown where nt equal to 262 and nh equal to 413. Different
thresholds have been tested for the pair (q = 0.2, p = 0.055) .The thresholds have been chosen to
achieve certain false alarm probability a within different sample sizes n. In particular, in this case,
y2 = 0.1143 was used. It is clear from Figure 5.1.22 and 5.1.23 that the a curves are sensitive to the
variation of the threshold t and h. we note also that as the value of the decision threshold increases, the
false alarm and power curves decrease. Different thresholds have been tested for the pairs (p =
.075 q = 0.3 ) and (p = .055, q = 0.2 ). The thresholds have been chosen to achieve a certain false
alarm probability a within different sample sizes n., in this case, y = 0.1674 and y2 = 0.1143 were
57


used. It is clear from figure 5.1.19 and 5.1.20 that the acurves are sensitive to the variation of the
threshold, while the P curves are stable.
Perfomance Characteristics of Sequential Algothms
Figure 5.1.22 Performance Characteristics of Sequential Algorithms using two pairs (q=0.3, p=0.075)
and (q=0.2, p=0.055); Detect from q to p.
Perfomance Characteristics of Sequential Algorithms
Figure 5.1.23 Performance Characteristics of Sequential Algorithms using two pairs (q=0.3, p=0.075)
and (q=0.2, p=0.055); Detect from p to q.
58


5.2 Conclusion
In this study, many areas of study are combined together and applied for one chosen
application which is oil spills detection. These areas are communication systems, image processing,
and detection and estimation area. It was first shown the data source of the images which were taken
from the Gulf of Mexico between May and June 2010. Then, they have been sent through a satellite
communication channel; then, the effect of the AWGN on various images was applied. After that it
was shown how the image signal process contributed significantly to enhance the received noisy
image. Different results have illustrated how the image enhancement assisted the analyst to make the
final decision. Finally, the implanted robust sequential detecting of change test was derived for
Bernoulli case, and the results have shown the efficiency of this algorithm against different
environment. The obtained results took approximately two minutes to obtain instead of hours.
Therefore, if all the points that have been covered in this section are considered, a gigantic time and
money can be saved to decrease the pollution and the harm on our planet.
59


5.2. Sequential Tests for the Detection of Voice Activity and the Recognition of Cyber Exploits
We consider the problem of automated voice activity detection (VAD), in the presence of
noise. To attain this objective, we introduce a Sequential Detection of Change Test (SDCT), designed
at the independent mixture of Laplacian and Gaussian distributions. We analyse and numerically
evaluate the proposed test for various noisy environments. In addition, we address the problem of
effectively recognizing the possible presence of cyber exploits in the voice transmission channel. We
then introduce another sequential test, designed to detect rapidly and accurately the presence of such
exploits, named Cyber Attacks Sequential Detection of Change Test (CA-SDCT). We analyse and
numerically evaluate the latter test. Experimental results and comparisons with other proposed
methods are also presented.
5.2.1 Introduction
Voice Activity Detection (VAD) is deployed extensively, including the Global System for
Mobile Communications (GSM), as well as several satellite and radar military and civilian
applications, (see in Figure 5.2.1). Thus, VAD is an important component of most systems that
incorporate digital voice transmissions. During real time voice transmission, periods of voice activity
are followed by silence, where both voice and silence periods are imbedded in background noise.
Since voice is generally transmitted through fixed bandwidth links, the transmission of the silence
periods induces severe bandwidth waste. Voice Activity Detection (VAD) allows for the compression
of the silence periods and may result in up to 30 to 40 percent of bandwidth savings.
To detect voice activity versus silence periods, the starting and ending points of continuous
speech activity must be detected. Several research efforts have been invested in this area, [28] [30].
In this section, we propose a novel VAD algorithm, named Voice Activity Detection using a
Sequential Detection of Change Test (SDCT-VAD). The algorithm is designed at an independent
mixture of Laplacian and Gaussian distributions; it is tracking effectively the boundary points between
continuous voice activity and silence time periods, where during silence there is only noise, while
during speech there is speech plus noise. The noise and noisy speech are modelled by a Gaussian
versus Laplacian plus independent Gaussian distributions. Results are included for the cases where
the SDCT-VAD is applied to detect voice activity, in both the presence and the absence of noise. The
60


algorithm is also tested within a real-time scenario, to exhibit its robustness and low complexity
properties.
Miliary p
V fi \
m ^ > j
'j/M V / civilian applications
Voice Activity
VAD
Figure 5.2.1 VAD integrated in a telecommunication system
Considering the possibility of cyber exploits during voice transmission, we also present a
novel cyber attack-sequential detection of change Test (CA-SDCT). The CA-SDCT algorithm is
deployed during voice activity periods, as detected by the SDCT-VAD algorithm. The proposed CA-
SDCA algorithm is designed at the Additive White Gaussian Noise (AWGN) cyber attack model and
is fully analysed and numerically evaluated in various environments. This section is organized as
follows: In Section 5.2.2, the SDCT-VAD algorithm is presented. In Section 5.2.3, the CA-SDCT
algorithm is developed. In Section 5.2.4, experimental results are included. In Section 5.2.5, we draw
some conclusions.
5.2.2 Voice Activity Detection Algorithm
The general operation flowchart of the VAD algorithm is depicted by Figure 5.2.2. The
problem to be solved here is the effective distinction between active and inactive voice periods.
However, the variety of both the active voice and the ambient noise make this problem quite
complicated in real life. As shown in Figure 5.2.2, first the speech signal is generated and is then
corrupted by Additive White Gaussian Noise (AWGN). The AWGN affects the shapes of both the
active voice and silence periods. The SDCT-VAD is then applied to detect voice activity periods. As
61


will be further discussed in Section 5.2.3, the SDCT-VAD operates on various Signal-to-Noise Ratios
(SNRs).
Figure 5.2.2 The general operation flowchart of the VAD algorithm
5.2.2.1 Speech and Noise Probability Distributions
Throughout this section, we consider two distinct probability density functions (pdfs) which
represent the voice and noise amplitude distributions of the proposed model. The two distributions for
speech and noise are assumed to be Laplacian and Gaussian, respectively, as in [26], where
different speech distribution models are shown in Figure 5.2.3. Figure 5.2.4 and 5.2.5 show noiseless
and noisy actual speech signals, respectively.
62


ProbeMly Derrely Function
Figure 5.2.3 Distributions Voice Signals
To decrease algorithmic design complexity, we assume statistical independence between
successive voice periods, as well as between signal and noise. We then derive the noisy speech
distribution via the convolution of the Laplacian and Gaussian densities. Assuming that the corrupting
noise is AWGN, the noisy speech signal is represented below,
Y = %signal + NAWGN (5.2.1)
where Y denotes the noisy speech, Xsignal stands for the clean speech and Nawcn represents the noise,
and where Xsignal and Nawgn are statistically mutually independent.
rtsw* '*eo 5 io !5 S X X £ C
Figure 5.2.4 Actual noiseless voice signal
silence + active voice
Original Signal
Figure 5.2.5 Actual Voice Noisy
Voice Signal
63


5.2.2.2 The Sequential Test for the Detection of Change
The sequential test for the detection of change, in its general form, was introduced and
analysed in [16] and [18]. It is assumed that an automated system will be monitoring the signal
activity to decide when the voice is active versus not, whose design is based on the detection of change
in the data generating stochastic process. The automated system will be implemented via the
deployment of the SDCT-VAD algorithm which will be tracking voice to silence and silence to voice
shifts, where voice and silence are modelled by two distinct stochastic processes. Below, we first
present the general model considered in references [16] and [18],
Let f0(xn) and fx(xn) denote the n-dimensional density functions of two well known, distinct,
mutually independent discrete-time stochastic processes at the vector point xn = {x1(x2, ,xn}.
Let it be known that the active process is initially generated by the density function f0 [16]. For the
problem addressed, it is assumed that fx represents the noisy voice process, while f0 represents the
AWGN noisy silence. Given the finite sequence x = {xj; 1 > 1} and the density functions f0(xn)
and fa(xn), the objective is to detect a possible f0 to change as reliably and as quickly as possible. To
detect a possible such change, select some positive threshold 5. Then, observe data points sequentially
and decide that the f0 to f| change has occurred the first time n such that T(xn) > 6, where
T(0) = 0 ; T(xn) = max jo,T(x
*) + gn(xn);gi(x) = log
fotoK1))
(5.2.2)
The above algorithm operates sequentially via the use of two thresholds, 0 and S, where 0
represents a reflecting barrier and S represents an absorbing or decision barrier [16].
When the two stochastic processes represented by the density functions f0 and ^ are
memoryless, then the conditioning in the log likelihood in (5.2.2) drops and the algorithmic operations
are memoryless as well. A symmetric algorithm that detects a shift from fj to f0, instead, can be easily
derived. In Figure 5.2.6, the time-evolution of both algorithms is depicted.
64


Active Voice: Decide HI
N -Data
sequentially.
Observed
/
r
\
£1
0
rO~ I Threshold
Reflector
Silence: Decide HO
N-Data
sequentially.
Observed
/ \
/ 1 i*
0
Reflector
0di
/
Threshold
Figure 5.2.6 Making the final decision in the first crossing to the threshold; detecting a change (a)
from f0(x) to fj(x) (b) from ft(x) to f0(x).
S.2.2.3 The SDCT-VAD using Laplacian and Gaussian Distributions
Let ft(x) and f0(x) represent the density functions of a single datum x from noisy voice
versus just noise, respectively. Let it be desirable to detect a possible change from f0(x) to fa(x),
where f0(x) is Gaussian and fa(x) is Laplacian plus Gaussian. The assumption here is that the observed
voice process is stationary, memoryless and Laplacian, while the noise process is independent from the
voice process and AWGN. Then,
fi(*) = [ fs(x y) fN(y)- *y (5.2.3)
J-T
where fs(x) is the Laplacian distribution that represents the voice speech, and fN(x) is the Gaussian
distribution representing the AWGN.
a
fsGO = 2exP(-l*l)
(5.2.4)
(5.2.5)
65


where a denotes the standard deviation of the Gaussian distribution. We now derive the
expression fi(x):
AW = f [|exp(-a|x y|)] £

A
i W = g exp(-a|x yl)) [^P (-£r)
dy
AW = | exp (~J~j {exP(- Alternatively, the distribution AW can be expressed as follows,
AW
a (a2a2\
= 2 CXP \ 2 /
(/i(x) + /i(-x)}
where
/t(x) = exp(-ax) O (- tra)
/i(x) = exp(ax) O (-- We then compute the log likelihood ratio updating step in the sequential test:
(5.2.6)
1 1
log/-w = 'og
§exp (~r) (ft w +
M;)
A(x) acrypHi a2a2 x2
,0gbo = ln_2~ ^+ "1'+ 2^+ ,n{,l(x) + fc("*)}
The algorithmic updating step can be written as:
66


(5.2.7)
Where
/?V2tt B2 f2
9(0 = In^y- + y + y + MHO + H~0)
h(0exP(-p0W-p)
The algorithmic step may be subsequently modified as follows
where
9(0 =
+ Y + Mh (O + K-O}
p a KOexp(-pOm-P)
(x 1
O(^x) = | -(p(u)du
(5.2.8)
and where the Signal to Noise Ratio (SNR) is:
SNR=^ = J2 ^
From the last equation, it can be recognized that the Signal-to-Noise Ratio (SNR) is a function
of both the Laplacian constant a and the standard deviation a. of the noise. To develop a robust
method for tracking the noise and speech signals, in Section 5.2.4, we will test the use of different
SNRs in the design of the algorithm. In Figure 5.2.7, we plot the SNR as a function of a, for various
values of the standard deviation a of the noise.
67


Figure 5.2.7 Signal-to-Noise Ratio SNR
Let us now select some positive threshold 5 and defme:
8 =
5
We may then modify the algorithm in (5.2.2), as manifested by the distributions derived in
this section, via scaling, resulting in the following operation: Observe data sequentially and decide that
the change from noise to voice activity has occulted the first time instant n such that T(xn) > S, where
T(0) = 0 ; T(x") = max{0,T(xn J) + gn(xn)}
/?V2rr p2] -1 \f2 r J
T(xn) = max 0,T(xn_1) + 1 + 2 2 Y + HKO + K-O) (5.2.10)
We note that the algorithm in (5.2.10) detects change from silence to active voice. The algorithm
that detects change from active voice to silence, instead, is similarly derived, where its recursively
derived algorithmic values are given by the expression in (5.2.12) below and where its decision
threshold is generally different than that of the algorithm in (5.2.10).
68


1 foto 1
,oewr'g
Mf)
f exp [h(x) + h(-x)}
T(xn) = max 0,T(xn-1) 1 [ pyfa B2] ,n 2 +2 -l Y + ln{fc(0 + ft(-0}|

(5.2.11)
(5.2.12)
5.2.2.4 Power and False Alarm Curves for Threshold Values Selections
In this section, we present algorithmic performance criteria and their use in the selection of
the decision thresholds. We specifically evaluate power and false alarm curves induced by the two
algorithms in Section 5.2.2.3 for several given decision thresholds. We then compare such curves for
different threshold values, to subsequently decide on the values of the operational algorithmic
thresholds. Let us define,
fnl(^)d^: The probability that at time n the algorithm has not crossed the threshold, 8, and its
value lies in (Ij, \ + d£), given that the acting pdf is fj.
where, the recursive expression below can be derived,
fnj(9= / fn-l.l(x).fSnft-x)-dx (5.2.13)
x=0 The probabilities {Pn; n > l} represent a power set.
The probabilities {an; n > 1} represent a false alarm set.
The main objective here is to find the threshold 8 that induces low false alarm and high power
for small sample sizes. To compute the power and false alarm curves, as induced by the probability
sequences {pn; n > l} and {an; n > 1}, respectively, we need to analyse the characteristics of the
69


updating step shown in equation (5.2.8). This process is explained in the Appendix, where the
expressions for the computation of the sequences {Pn; n > 1} and {a,,; n > 1) are also derived.
Given threshold S, the silence mode to active voice mode change detecting algorithm is basically
characterized by two time curves: the power and false alarm curves, denoted respectively pn and an,
respectively, where n denotes time instant and where,
Pn: The probability that the silence to active voice mode change detecting algorithm crosses its
threshold before or at time n, given that the operation mode is active voice mode throughout [32].
an: The probability that the silence to active voice mode detecting algorithm crosses its threshold
before or at time n, given that the operational mode is silence mode throughout [32],
When the algorithm that monitors change from mode silence to mode voice is considered, the
threshold <5 may be selected based on the following principle: At given time n have the powers
induced by the parallel algorithms be above a predetermined lower bound, while the false alarm
induced by each algorithm remains below a predetermined upper bound. The threshold for the
algorithm that monitors change from voice to silence, instead, is selected similarly.
8
XI
2
0.
Number of Samples
Figure 5.2.8 Power and false alarm curves Vs different thresholds
In Figure 5.2.8, we depict the pn and an representative curves, to observe and discuss
qualitative characteristics. We plot these curves for two different threshold values. From the Figure
70


5.2.8, we note that as the value of the decision threshold increases, the false alarm curve decreases, but
so does the power curve. The threshold selection for the silence to active voice change monitoring
false alarm, at a given time instant n. A similar criterion may be adopted in the threshold selection for
the active voice to silence monitoring algorithm.
5.2.3 An Algorithm for Detecting Cyber Attacks during Speech Activity
In this section, we consider the case where the voice transmission channel may be vulnerable to cyber
exploits. We then focus on developing an automated system that, in concurrence with voice activity
detection, also detects cyber exploit activities. We thus develop a Cyber Attack- Sequential Detection
of Change Test (CA-SDCT), designed to detect cyber attacks during voice activity periods, as the latter
are detected by the SDCT-VAD algorithm in Section 5.2.2. The block diagram of the overall system is
depicted in Figure 5.2.2, Section 5.2.2. As shown in Figure 5.2.2, first the speech signal is generated
and is then corrupted by additive white Gaussian noise (AWGN). The SDCT-VAD is then deployed to
distinguish between voice activity and silence periods. We finally wish to detect possible cyber attacks
during voice activity periods.
As in Section 5.2.2.2, let f0(xn) and fi(xn) denote the n-dimensional density functions of two
well known, distinct, mutually independent discrete-time stochastic processes at the vector point x" =
{x1(x2, ,xn). For the problem addressed here, fj represents the process of cyber exploits
superimposed on noisy voice activity, while f0 represents the noisy voice activity process in the
absence of cyber attacks. Given the infinite sequence x = {xi(- i > 1}, let the n-dimensional density
functions be denoted f0(xn) and fi(xn). The objective is to detect a possible f to ft change as reliably
and as quickly as possible, utilizing the observed data sequences. As in Section 5.2.2.2, we first select
some positive threshold S0. Subsequently, we observe noisy voice data sequentially, during voice
activity periods detected by the SDCT-VAD, and decide that the f0 to change has occurred, the first
time n such that T(xn) > <50, where
algorithm may be based on a required lower bound for the power and a required upper bound for the
(5.2.14)
71


A similar algorithm may be devised for the detection of shifts from ft to f0, instead. The sequential
operation of the two algorithms is depicted in Figure 5.2.9.
Cyber Exploits Present Decide HI
---------------------------o----*0
N -Data
sequentially ^ /
Observed '
Threshold
Cyber Exploits Absent: Decide HO
---------------------------------0----\
N -Data
sequentially.
Observed
/
Threshold
/
0
Reflector
/*.
1
0
Reflector
Figure 5.2.9 Making the final decision in the first crossing to the threshold, (a) Cyber Exploits Present,
denoted H| (b) Cyber Exploits Absent, denoted H0.
We model the presence of cyber exploits by Additive White Gaussian Noise (AWGN) that is
superimposed on the transmission channel AWGN, resulting in relatively excessive cumulative white
noise. When the two stochastic processes represented by the density functions f0 and fj are
memoryless, the conditioning in the log likelihood in (5.2.14) drops and the algorithmic operations are
memoryless as well. As directly deduced from Section 5.2.2.3, in the present case we have:
fiO) = + VC*)} exp (5.2.15)
foto = f (VM + exP (~22~) (5.2.16)
where,
K(*) = exp(-ax) (To: Standard deviation of the transmission noise, in the absence of cyber exploits.
ax \ Standard deviation of the cumulative noise when cyber exploits are added to the transmission noise.
Then,
72


fxQ) = f {VM + hat(-x)} exp (^Y~)
gfoW f (M*) + M"*)} e*P pT")
Or,
fxW a2
logfoW = TK -ffo] + ,n
{fegl 00 +
(V> 00 + /la0(-^)}
(5.2.17)
The implementation of the cyber exploits detection algorithm is then as follows:
During voice activity periods, as detected by the SDCT-VAD algorithm, observe data sequentially and
decide that the change from absence to presence of cyber exploits has occurred, the first time instant n
such that T(xn) > T(n) = max
0, T(n 1) + y ki2 Oo2] + In
{ft^OO + hai (-*)}
{ha0 to + ha0(-x)}
(5.2.18)
We note that the algorithm in (5.2.18) detects a f0 to fj change. The algorithm that detects a
fj to f0 change, instead (from presence to absence of cyber exploits), is similarly derived, where its
recursively derived algorithmic values are given by the expression in (5.2.19) below and where its
decision threshold, 51( is generally different than that of the algorithm in (5.2.18), as shown in Figure
5.2.9.
T(n) = max
0, T(n 1) + y [ob2 oi2] + In
[hao{x) + hao(-x)}
+ hCTi(-x)}
(5.2.19)
5.2.4 Experimental Results
5.2.4.1 Testing the SDCT-VAD
In this section, we state the steps involved in the numerical evaluation of the SDCT-VAD
algorithm. First, we select the pertinent involved parameters and deploy the resulting SDCT-VAD
algorithm, to detect any voice activity in the communication link. Then, the SDCT-VAD is evaluated
73


in various noisy environments. In our simplified model, the silence plus noise mode of operation is
assumed to be represented by a Gaussian distribution, while the noisy voice signal is represented by a
mixture of Laplacian and Gaussian distributions, as shown in Figure 5.2.10. The pertinent parameters
to be chosen in the SDCT-VAD design are the Laplacian parameter, the standard deviation of the
Gaussian noise and the two algorithmic thresholds: A threshold 8 used by the algorithm in (5.2.10);
for the detection of change from noise to noisy voice activity, and a threshold 8j used by the algorithm
in (12); for the detection of change from noisy active voice to just noise. We used the power and false
alarm curves discussed in Section 5.2.2.4, to decide on the values of these two thresholds. In
particular, we selected the (80, Sl5 a, a) values (0.3,0.05,0.98,0.0523).
We used the design parameter values stated above and tested the robustness of the resulting
SDCT-VAD algorithm in the presence of various noisy environments. Various noises were mixed with
the clean speech signals. Six different noises were used in our evaluations, including white noise,
wind, computer fan, babble, flowing traffic and train passing, as shown in Figure 5.2.11, with different
SNRs (25,20,15,10 and 5), as shown in Figure 5.2.12 and 5.2.13.
Original Signal
0 05 1 1.5 2 25 3
Time(Sec)
Figure 5.2.10 Original Voice Signal
74


Noise A WON
Noise Wind
Original Signal
Original Signal $NR=25
Q. <
0 20 40 60 80 0 20 40 60 00 Noise Computer Fan Noise Babble
'o 20 40 60 80 "'o 20 40 60 80
Noise Flowing Traffic
Time (sec)
Noise Tram Passing Cut
0 20 40 60 80
Time (sec)
Time (sec)
Original Signal SNR=1Q
Time (sec)
Time (sec)
Original Signal SNR=15
-LXi
111
I ------ Mil->
0 12 3
Time (sec)
Onginal Signal SNR=S
Time (sec)
Figure 5.2.11 Various noisy environments Figure 5.2.12 Original signal corrupted by AWGN
1
E O
<
-1
l
O 5
£ 0
-O 5
I
5
f
-5
I
5
£
"50 10 20 30 40 50
(a) Original Signal
10 20 30 40 50
(b) Noise: Wind
10 20 30 40 50
(c) Noise: Babble
im ii m* mMHUiml
10 20 30 40 50
(d) Noise: Computer Fan
£
(e) Noise: Flowing Traffic
5 '---------------------------------------------1
0 10 20 30 40 50
(f) Noise: Train Passing
2 |-----------------------------
20 10 20 30 40 50
Time(Sec)
(g) Noisy Signal "AWGN" for SNR= 5 dB
20 10 20 30 40 50
(h) Noisy Signal Wind" for SNR= 5 dB
£ 0
-1
0 10 20 30 40 50
(i) Noisy Signal "Babble" for SNR= 5 dB
5
CL.
5
-5
0 10 20 30 40 50
(l) Noisy Signal Computer Fan for SNR* 5 dB
10
£ 0
-10
0 10 20 30 40 50
(k) Noisy Signal Flowing Traffic for SNR= 5 dB
5
£ 0
-5
0 10 20 30 40 50
(I) Noisy Signal "Train Passingfor SNR= 5 dB
2 i------------------------------------------1
'20 10 20 30 40 50
Time(Sec)
Figure 5.2.13 Results of adding noise to the original speech signal (5dB SNR), (a) Clean speech, (b)
Wind Noise, (c) Babble Noise, (d) Computer Fan Noise, (e) Flowing Traffic, (f) Train Passing Noise,
(g) Noisy Signal Noise: AWGN. (h) Noisy Signal Noise: Wind, (j) Noisy Signal Noise:
Computer Fan, (k) Noisy Signal Noise: Flowing Traffic. (1) Noisy Signal Noise: Train Passing
Cut.
75


For comparison, the same voice and noise environments are also tested by the approach
presented in [29] and the G.729 VAD algorithm in [31]. The results are summarized in Tables 5.2.1
and 5.2.3. The noise data are obtained from http://www.freesound.org/index.php and are added to the
clean speech signal at SNRs varying from 5dB to 25dB.
To empirically evaluate the SDCT-VAD algorithm, many audio messages were used, with
different lengths, (3 sec and 50 sec), with both male and female speakers and with different SNRs, (5
dB -25dB). The effect of these SNRs on the audio messages is exhibited in Figure 5.2.12, where
Figure 5.2.10 exhibits the original speech signal. Figures 5.2.14 and 5.2.15 below show results for the
SDCT-VAD algorithm with operating parameters as those stated in this Section, when the SNR is
25dB and 5dB, respectively. The accuracy of the results depends on the level of the SNRs and the type
of the noise environment, as shown in Tables 5.2.1 and 5.2.2.
Sequential Detecting a Change Te$t
y o TimefSec)
Figure 5.2.14 SDCT-VAD results Original
Signal corrupted by AWGN (SNR=25dB)
Sequential Delecting a Change Tesl
0 05 I 15 K17m 25 3
TimefSec) y o
Figure 5.2.15 SDCT-VAD results Original
Signal corrupted by AWGN (SNR=5dB)
The efficiency of the SDCT-VAD algorithm was evaluated for various noisy voice signals. In
the first experiment, we tested the efficiency of the proposed method using the same audio recording
discussed above, tracing the speed and the accuracy of the algorithm in detecting the silence mode to
active mode change and vice verse.
To comparatively evaluate the performance of the proposed SDCT-VAD algorithm, we
compared its induced results with those of the manual segmentation. Figure 5.2.10 exhibits the hand-
marked results of manual segmentation. Figures 5.2.14 and 5.2.15 exhibit the automated
76


segmentation induced by the SDCT-VAD proposed algorithm. We evaluated the algorithmic
probability of error (Pe), using the formula below:
re(Av)
Y [ \AVSP(manual) AVSP(RSDC VAD)\2
-w 1 L
AVSP(manual)
\AVEP(manual) AVEP(RSDC VAD)\2
+
AVEP(manual)
(5.2.20)
A V: Active Voice;
A VSP: Active Voice Starting Point;
A VEP: Active Voice Ending Point;
N: Number of Active Voice regions.
The performance of the SDCA-VAD is evaluated in terms of probability of false and correct
decisions, where Pc, is the probability of correct speech classification and where Pe is the probability
of false speech classification, computed as in (5.2.20).
To compute Pe, we start with known voice activity and the starting and ending points of voice
activity marked. We then superimpose AWGN voice contamination with various SNRs (25,20,15,10
and 5dB) and deploy the corresponding SDCT-VAD for various noisy environments, as shown in
Figures 5.2.10, 5.2.11 and 5.2.12. Results comparing the SDCT-VAD with the manual approach are
shown in Table 5.2.1.
Table 5.2.1 Comparing the Starting and Ending detection time instances of the Noisy Active Voice
Messages Using the Manual and Proposed method.
77


After finding the starting and ending points of voice activity, as shown in Table 5.2.1, we
used the formula Pe(Av) in (5.2.20) to evaluate the accuracy of the algorithm. The results are
summarized in Table 5.2.2. From the experimental results, it is clear that the proposed SDCT-VAD
algorithm outperforms the manual method.
Table 5.2.2 Pc's and Pe's OF the Proposed RSDCA-VAD for Various Environmental Conditions.
SNR SNR SNR SNR SNR
Noise/SNR(dB) 25 20 15 10 5
White Pc(%) 99.958 99.957 99.946 99.92 99.91
Pe(%) 0.0418 0.0423 0.0540 0.076 0.081
Pc(%) 99.976 99.972 99.945 99.89 99.88
Wind Pe(%) 0.0238 0.0279 0.0545 0.107 0.117
Computer Pc(%) 99.976 99.979 99.977 99.97 99.97
Fan pe(%) 0.0234 0.0210 0.0227 0.025 0.026
Pc(%) 99.979 99.971 99.969 99.96 99.95
Babble Pei%) 0.0208 0.0285 0.0303 0.030 0.037
Flowing Pc(%) 99.996 99.982 99.953 99.94 99.83
Traffic Pei%) 0.0037 0.0178 0.0468 0.058 0.162
Train Pc(%) 99.979 99.977 99.975 99.96 99.95
Passing Pe(%) 0.020 0.0224 0.0243 0.038 0.048
Pc(%) 99.977 99.973 99.96 99.94 99.91
Average Pe(%) 0.0225 0.0266 0.0387 0.0556 0.078
In Figures 5.2.16 and 5.2.17 we plot the Pc's and Pe's results included in Table 5.2.2. To
further validate the effectiveness of the proposed SDCT-VAD, we compared its probabilities of correct
speech activity detection with those of other approaches. Table 5.2.3 shows a comparison between the
SDCT-VAD and two different VAD approaches: the G.729 in [31] and the proposed method in [29],
78


The Propsed RSDCA-VAD For Various Environmental Conditions
Figure 5.2.16 Pe's of the proposed RSDCA-VAD various environmental conditions.
The Propsed RSDCA-VAD For Various Environmental Conditions
Figure 5.2.17 Pc's of the proposed RSDCA-VAD various environmental conditions.
79


The clean speech signal that has length of 50 sec, 60.05% speech and 39.95% silence, was used to
evaluate the proposed algorithm against various environmental conditions and to compare it with the
ITU standard G.729Annex B [31 ] and the proposed in [29] approaches.
From Table 5.2.3, it can be recognized that even with environmental challenging conditions, the
proposed SDCT-VAD outperformed the G729B VAD and the method in [29],
Table 5.2.3 Pc's Of The Proposed RSDCT-VAD, and Different VAD Approaches for Various
Environmental Conditions.
Environment G.729 VAD [31] and [29] Proposed Method in [29] Proposed RSDCA-VAD
Noise SNR Pc(%) Pc(%) Pc(%)
5 87.46 75.62 97.563
White 15 97.83 93.42 99.136
25 99.69 99.07 99.328
5 97.29 97.83 98.131
Vehicle 15 99.77 99.46 99.711
25 100.00 100.00 99.781
5 92.96 86.38 97.201
Babbie 15 98.45 94.89 99.504
25 99.77 99.38 99.664
4.2 Testing the CA-SDCT
As stated in Section 5.2.3, to detect shifts from absence to presence of cyber exploits and vice
versa using the CA-SDCT algorithm, two algorithmic decision thresholds are needed: A threshold 60
used by the algorithm in (5.2.18); for the detection of change from Cyber exploits absent to Cyber
exploits present, and a threshold used by the algorithm in (5.2.19); for the detection of change from
Cyber exploits present to Cyber exploits absent. It is assumed that the cyber attack is represented by
AWGN. Using the power and false alarm curves, as with the SDCT-VAD algorithm, we selected
80


thresholds. In particular, we selected the (60,6l5 a,0i) design values (0.07,0.01,0.98,0.0773) and
tested the a2 values (0.0798,0.0849, 0.1034), to evaluate the robustness of the resulting CA-SDCT
algorithm.
Figure 5.2.18 (a) Original Signal (b) Noisy signal with SNR=15dB (c) Noisy signal with SNR=10dB
(d) Noisy signal with SNR=5dB
From Figure 5.2.19, parts (a), (b) and (c), we may observe the evolution of the deployed CA-
SDCT algorithm for different SNR values, where the latter values reflect the cumulative effect of
normal channel noise and cyber noise. Each time the threshold is crossed, an alarm is activated.
81


Amplitude
(c)
Figure 5.2.19 Cyber Detection during speech activity detected periods using the CA-SDCT (a)
Sequential Test: SNR=15. (b) Sequential Test: SNR=10. (c) Sequential Test: SNR=5.
Figure 5.2.20 shows alarm activation scenarios regarding cyber attacks, where in (a) no alarm
is activated, where in (b) one alarm is activated and where in (c) several alarms are activated.
82


Cyber Attack alarms SNR=6dB
Cyber Attack alarms SNR=5dB
1 2 | 0 SNR=15 L | SNR=10 [
1 1 < >
oe 06 -
e c i
5 06 3 06
0 4 04 -
0 2 02
01 02 03 04 05 06 C 01 02 03 04 05 06
Tffrw(sc) Time(sec)
(b)
(b)
Cyber Attack alarms SNR=6dB
(c)
Figure 5.2.20 Cyber Detection during speech activity detected periods using RSDA-CA (a) Alarms
SNR=15. (b) Alarms SNR=10. (c) Alarms SNR=5.
5.2.5 Conclusions
A novel voice activity detection (VAD) approach was presented. The approach uses the
Sequential Detection of Change Algorithm (SDCT-VAD), designed at the Laplacian-Gaussian
distributions additive mixture. We analysed and evaluated the robust sequential algorithm in the
presence of Additive White Gaussian Noise. Several different speech messages were chosen for the
effectiveness evaluation of the SDCT-VAD, regarding its accurate detection of changes from voice
83


activity to silence and vice versa. The experimental results have shown that the algorithm is effective
in various noisy environments and outperforms other existing voice activity detection methods.
A novel cyber attack- sequential detection of change algorithm (CA-SDCA) was also
presented, deployed to detect cyber attacks during speech activity periods. The proposed algorithm is
preceded by the Voice Activity Detection algorithm Sequential Detection of Change Test (SDCT-
VAD). We considered the case where voice messages are transmitted through the communications
system, while a cyber attack may occur at any point in time. The proposed algorithm was analysed and
evaluated. The latter algorithm detects cyber attacks effectively; during speech activity periods
detected-by the SDCT-VAD algorithm. We modelled the cyber attacks by Additive White Gaussian
Noise. The experimental results have shown how the algorithm can be implemented with effective
detection results, in a variety of different environments.
84


5.3 Detect Cyber Attacks Using BER-Threshold Tool
The risk of cyber attacks has grown tremendously in the recent years. Many of these cyber attacks can
be linked to current and historic events in the social, political, economic, and cultural dimensions in the
human world. The goal of this section is to detect cyber attacks using a Bit Error Rate Threshold
(BERT) tool. The system administrator may get an early warning, the, regarding system hacking,
allowing time for possible recovers and future protection. A study case related to the previous section
is presented in this chapter. It is a combination of the Robust Sequential Detection of Change Voice
Activity Detection Algorithm (RSDC-VAD) and detection of Cyber Attacks using Bit Error Rate -
Threshold (CA-BERT). The model is analyzed and some experimental results show how the proposed
model may conserve bandwidth and energy, while increasing the level of security against cyber
attacks.
5.3.1 Introduction
Cyber attacks impose a major threat to our communication networked systems. The motivations vary,
while, the victims of cyber attacks may be commercial organizations, governments or military
infrastructures. The attack agent may be a hacker or a terrorist. The opportunity of these attacks to hack
the system increases as the weaknesses of the system decreases.
To detect cyber attacks, we monitor the traffic and the performance of the network. After detecting that
the network has been hacked, the system administrator starts investigating and troubleshooting. In this
section, we only monitor the traffic performance of the network, to detect if any change in the
communication link has occurred. The proposed method for cyber attack detection is called the Cyber
Attacks using Bit-Error Rate -threshold (CA-BERT) tool, which may be low cost, rapid and efficient,
for different cyber attack scenarios. The proposed simulation methodology is designed to test
information fusion systems for cyber security that are under development or already built. The BER-
Threshold method may be used for conducting experiments to analyze computer network
vulnerabilities and evaluate efficiency and effectiveness of security policies. Systems administrators
normally deal with countless daily warning that is hard to track each one of them. Alternatively, an
85


automated system may transmit warning signals, when unusual anomalies occur, as detected by CA-
BERT traffic monitor. The complete model is shown in figure 5.3.1.
X Cyber X
\ Attack '
CAForm#3 j ^ CAForm#2 , CAForm#!
i
Random Noise
i i
i i
i r
'V "
(Distribution Not i i N i r
1 !'1.
i
i 1 r
i 1 i
Start
.X
X
Voice Signal #(1) i
Silence + Voice j
Signal #(2) ,
i
A/D Converter |
r :
Known) " T " \ / Yes - j Mdulator i ! Demodulnlar
f -Xo / r -NV if > N / 1 Yes i i ' BER- i 1 Threshold i To; : . No Attack i 1 Detected i i i "X
i Yes ' r : i Demodulator 1
o
o
BER- i
Threshold i*
r Attik' ' ...... r i i
i rvZTj >' D/A Converter RSDC-VAD
Detected : j i i
Voice Signal #1
& #2 & Noise
i D/A Converter 1
' r ~ ^ ~
i RSDC-VAD
T.
r *-------? i---------
j Voice Signal #(2) ^ RSDC-VAD
i
fill
Save BW and
Energy
Inverse
Fourier
i"^T Transform
! Only for Voice
i Signal #(1)
Separate
* *
Fourier
Transform
and noise 1
using filters i
Start Processing
Mi to recover the 1
original signal 1
i Voice Signal #2
SaveBW (
and |
Energy i
Ignore the rest
J
Figure 5.3.1 Model: Detect Cyber Attack using BER-Threshold
The present work motivated by the need for testing situational awareness tools, developed to
detect and analyze attacks on computer networks. The simulation approach requires knowledge of the
network operation, which must be captured by the simulation model. However, as discussed briefly in
the introduction, the level of detail included in the model will depend on the objective of the
simulation. In this case, the packet level information and computer network traffic details are not
86


needed, so the simulation may be constructed at a higher level, to produce alerts caused by cyber
attacks on information carrying network traffic.
Let us assume that an organization uses the model shown in figure 5.3.1, utilizing an encoded
communications link to relay secure messages among its members. Our objective here is to keep the
encoded link secure and use it to detect possible cyber attacks. It is then implied that the transmitted
messages through the secure link are confidential and it can be recovered regardless of possible cyber
attacks. Let the transmitted voice message to one of the members be I am completely operational
and all my circuits are functioning perfectly ", where the latter is the signal that we wish to recover.
Let the signal generated by the attackers be "I cannot take any of this seriously unless I know who I
am talking to ", which is the signal that causing interference.
5.3.2 Modeling Cyber Attacks
The term "Cyber Attacks = hacker" is refers to anyone who is authentically interested in
pushing the limits of software and hardware and is eager to do so in a fully supportive, low cost and
sharing network environment. The progress that a hacker can make in an attack depends upon the
hackers capabilities and the vulnerabilities of the network. The methods for modeling and simulating
the initiation and progression of cyber attacks through a computer network included in this model are
based on AWGN, random noise or interference.
5.3.3 Bit-Error Rate -Threshold (BERT)
Over the last two decades, communication networks have evolved from all analog telephone
and radio transmissions to the modem digital communications systems. These digital networks are
required to cany large amounts of data at high rates. The need for increased bandwidth has caused an
evolution in the media from wires and air. Service providers must guarantee data integrity to their
clients, where the bit-error rate measurements are used to measure enor resistance performance. The
pertinent criterion is called bit-error rate (BER) and its definition is given below,
87


Number of changed bits
BER = ^ ,v,, * . (5.3.1)
Total Number of bits
Thus, bit-error rate monitoring plays an integral role in detecting changed error performance and
quality degradation of communication links. An important aspect of the bit-error rate monitoring that
we have designed is that it may operate at a variety of modulation techniques and rates. The BER
performance is affected by many factors such as power, noise, or modulation method. If we assume
that our system has a specific power, known noise and given modulation technique, the BER
performance can be predicated, thus, its change from predicted value may point to the presence of
cyber attacks. The BER measurement is not so complicated; it is affined by sending a stream of bits
through our communication systems and compares the output to the inputs using the formula in (5.3.1),
where the input stream is known to the receiver via its transmission through a secure link.
Our proposed monitoring method is named Bit-Error Rate -Threshold (BERT); we will present and
analyze it, in detail. The complete model is shown in figure 5.3.1.
To start applying the BERT, let us make some assumptions:
1- The system is already installed, while we wish to increase the level of the security by
detecting effectively cyber attacks.
2- Generate a well known message at the receiver with a specific length.
3- The cyber attack is modeled as Additive White Gaussian Noise (AWGN), as well as
additional distortion and interferences.
4- The optimum BER system performance is assumed well known.
The BERT tool starts operating as an investigator. If, for example the BER at some EbNo is normally
equal to BER = 0.0001, then we assume that a minimum and a maximum threshold define derivation
tolerance; that is 0.001>BER>0.00001, represents normal performance, where an alarm is initiated if
either threshold is crossed. To demonstrate this method in more detail, let us assume that the
communication link uses some modulation method such as, QPSK. shown in figure 5.3.2, and normal
optimal BER performance is known (Assumption #4) and is represented by the Blue curve. Then
two thresholds are created by black curves.
88


This operation presents a simulation model for the intrusion detection system, depicted by
simulated alerts (Red, Yellow or Green), where these colors level represent the cyber attacks, and the
level should be assigned by the users specifications. The users specifications represent design
requirements for the system, operation, such as bandwidth, speed, type of information (voice or data),
security level, distance between each Local or Wide Area Network (LAN or WAN), quality of the
communication link, type of modulation techniques, etc. There are 4 subplots in figure 5.3.2, where
each one represents a specific behavior for the communications link; i.e.;
1) Plot (a) demonstrates that no attack has been detected because the evaluated BER is still between
the two thresholds; so, the alarm in this case will be green.
2) Plot (b) and (c) illustrate that there is something wrong in the communication link, but it is not
critical, since the evaluated BER is still between the two thresholds; so, the alarm in this case is
yellow.
4) Plot (d) demonstrates that an attack has been detected, since the evaluated BER is not between the
two thresholds; so, the alarm in this case is red.
The same previous steps can be repeated again if we use another modulation technique such as
64QAM. Figure.5.3.3 shows the some results..
89


Full Text

PAGE 1

MODULATION, CODING AND DETECTION FOR SATELLITE AND SPACE COMMUNICATIONS by Ehab A. Etellisi B.S., AI Fateh University, 2005 A thesis submitted to the University of Colorado Denver in partial fulfillment of the requirements for the degree of Master of Science Electrical Engineering 2011 A

PAGE 2

This thesis for the Master of Science degree by Ehab A. Etellisi has been approved by 7 J. Deng J Jan Bialasiewicz Date

PAGE 3

Etellisi, Ehab A. (Master of Science, Communications Engineering, Electrical Engineering, University of Colorado Denver) Modulation, Coding and Detection for Satellite and Space Communications Thesis directed by Professor Titsa Papantoni ABSTRACT Communication and signal processing applications can be linked to our current and historic events in the social, political, economic, and cultural dimensions in the human world. At the same time, prominent performance metrics in various such applications include accuracy, transmission power, bandwidth efficiency, security and cost effectiveness. In this thesis, we present and analyze three applications of current interest, each of which invests heavily on one of the following performance metrics: accuracy, bandwidth/energy efficiency and security. In the frrst four chapters, we give some background on communication and signal processing such as modulation and coding. Then in chapter five, we start introducing some communications system applications. For instance, to address accuracy; we selected the problem of oil-spill detection: we show how system accuracy may minimize remediation costs and limit dangerous impacts to the environment. To address bandwidth/ energy efficiency, we selected a voice activity detection problem. To address security, we selected the problem of timely and accurately detecting the presence of cyber exploits in the communication transmission, where a new detection method is introduced and analyzed. In our developments, cost effectiveness is of high priority as well. This abstract accurately represents the content of the candidates' thesis. I recommended its publication.

PAGE 4

DEDICATION I would like to dedicate my thesis to my beloved family, especially ... to Dad and Mom for instilling the importance of hard work and higher education; to my brother, Gaith, for his encouragement and support; to my friends and relatives, and supporters who have made this happen; finally, many thanks go out to my country for providing me this scholarship.

PAGE 5

ACKNOWLEDGEMENTS This work would not have been possible without the scholarship provided to me by my country. I would like to thank my family and friends for their understanding, encouragement and support in my pursues. I am indebted to all my faculty members who have strongly affected so much on my educational experience like Dr. Yiming J. Deng. I offer my regards and blessings to all of those who supported me in any respect during the completion of this thesis. I am heartily thankful to my supervisor, Prof. Titsa Papantoni, whose encouragement, guidance and support from the initial to the final level enabled me to develop an understanding of the subjects that I covered in this thesis. She is an excellent role model, as someone who really knows how to balance scientific research and teaching. Despite her busy schedule, she always finds the time to discuss anything "from the first ideas and designs to simulation results". I can only express that her dedication and commitment to science and education is truly inspiring and remarkable. The partial support ofthe contract AFOSR FA9550-05-I-0388 is acknowledged. Finally I would like to express my appreciation to whomever touched my life directly or indirectly.

PAGE 6

TABLE OF CONTENTS Figures ..................................................................................................................................................... ix Tables ..................................................................................................................................................... xii CHAPTER l. Introduction to Digital Communication Systems ............................................................................. I 2. Signal Processing ............................................................................................................................... 9 2.I Analog to Digital Conversion ......................................................................................................... I 0 2.2 Anti-aliasing Filter .......................................................................................................................... 10 2.3 Quantization (Uniform Quantization) .......................................................................................... IO 2.4 Sampling ......................................................................................................................................... IO 3. Modulation Techniques ................................................................................................................ I6 3 .I Introduction .................................................................................................................................... 16 3.2 Modulation Techniques .................................................................................................................. 16 3.3 Frequency Shift Keying (FSK) ...................................................................................................... I8 3.3.1 Binary Frequency Shift Keying (BFSK) .................................................................................... 18 3.3.2 M-Frequency Shift Keying (MFSK) ..................................................................................... 18 3.4 Phase Shift Keying (PSK) ............................................................................................................. I9 3.5 Differential Phase Shift Keying (DPSK) ...................................................................................... 23 3.6 Quadrature Amplitude Modulation (QAM) .................................................................................. 24 3.7 Signal Constellation ...................................................................................................................... 25 4. Channel Coding .............................................................................................................................. 26 vii

PAGE 7

4.1 A Novel Trellis-Coded Modulation .............................................................................................. 26 4.1.1 Trellis-Coded Modulation Background ...................................................................................... 26 4.1.2 Fundamentals and Concept of TCM ........................................................................................... 27 4.1.3 Mapping and Trellis Diagram .... .................................................... ...... .. ...... ............................. 29 4.1.4 Viterbi Algorithm ....................................................................................................................... 33 4.1.5 TCM MFSK, MPSK, DPSK, and MQAM Constellation Mapping.......................................... 34 5. Communication Systems Applications ............................................................................................ 41 5.1 Oil Spill Detection by Satellite Image using Sequential Detection of Change Test.. ................... 41 5.2 Sequential Tests for the Detection of Voice Activity and the Recognition of Cyber Exploits ...... 61 5.3 Detect Cyber Attacks using BER-Threshold ................................................................................. 88 6 Conclusion ......................................... ........ ................................................................... ................... 94 Appendix A. Derivations.............................................................................................................................. 96 Bibliography .......................................................................................................................................... 99 viii

PAGE 8

LIST OF FIGURES Figure 1.1 Block diagram of a typical digital communication system ....................................................... 5 1.2 Formatting and transmission ofbaseband signals .................................................................... 7 2.1 Typical analog to digital conversion process. ...................................................................... I 0 2.2 Quantized images (a, b, c, and d) .......................................................................................... 11 2.3 Original and Quantized image (a, b, c, and d) ..................................................................... 14 2.4 Original and Quantized voice (a, b, c, and d) ....................................................................... 15 3.1 UncodedBER for different modulation schemes. ................................................................ 19 3.2 Signal Constellation for Binary, Quadrature, 8, and 16 PSK ................................................. 20 3.3 Un-coded BER for DPSK, 2, 4, 8, 16, 32-PSK. .................................................................... 22 3.4 Signal Constellations for n/4 and n/4 DQPK. ..................................................................... 23 3.5 Signal Constellations for 16, 32, 64, 128, and 256-QAM ..................................................... 25 3.6 Uncoded BER Signal Constellations for MQAM ................................................................. 25 4.1 Source and channel coding ...................................................................................................... 27 4.2 Trellis Coded Modulation. ............................. ...... .................................................................. 28 4.3 Constellation doubling in TCM. A QPSK and 32-QAM signal transmitted ......................... 29 4.4 General trellis coded modulation (a) BPSK, code rate l/2, output QPSK and QFSK (b) QPSK, code rate = 213, output 8PSK and 8FSK (c) 8PSK, code rate = 3/4, output 16PSK and 16QAM .................................................................................................................................. 30 4.5 Uncoded BER vs. Eb/NO in A WGN environment. ................................................................. 31 4.6 Coded BER vs. Eb/NO in A WGN environment ...................................................................... 32 4. 7 TCM MFSK, MPSK, DPSK, and MQAM Constellation Mapping ........................................ 34 4.8 Multiple trellis coded Modulation with different Rates (a) Frequency Shift Keying (FSK) (b) (QPSK) (c) 16QAM (d) 64QAM ........................................................................................ 35 4.9 SPSK Set-partitioning according to Ungerbock ...................................................................... 37 4.10 Conventional versus multiple trellis coded mod with Rate (a) 2/3 (Rate #I) and (b) Rate 2/5 (Rate #2) ................................................................................................................................. 38 4.11 Conventional versus multiple trellis coded mod with Rate (a) 2/6 (Rate #I) and (b) Rate 2/6 (Rate #2) ................................................................................................................................. 39 ix

PAGE 9

4.12 Conventional versus multiple trellis coded mod with Rate 2/8 (Rate #5) ............................ 39 4.13 Multiple trellis coded Modulation with different Rates Frequency Shift Keying (FSK) (b) (QPSK) (c) 16QAM (d) 64QAM ............................................................................................ 40 5.1.1 Original image. NASA captured this image of the Gulf of Mexico on May 24, 2010 at 16:50 UTC MODIS ................................................................................................................. 43 5.1.2 Block diagram of the implemented system ............................................................................ 44 5.1.3 Block diagram of the whole system using different scenarios ............................................... 45 5.1.4 Monitoring System operates for many different scenarios ................................................... 46 5.1.5 The effect of A WGN on transmitted and received Gray-Coded 256-QAM Signal Constellation image ................................................................................................................. 47 5.1.6 BER to measure the performance ofthe channel for 256 QAM (Simulation Vs Theoretical Vs Simulation with extra A WGN) .............................................................................................. 48 5.1.7 Original image after transmitted and received through the satellite. "After modulated and demodulated" .......................................................................................................................... 48 5.1.8 Original image after transmitted and received through the satellite. "After modulated and demodulated" .......................................................................................................................... 48 5.1.9 Original image after transmitted and received through satellite and adding some extra AWGN. "After modulated and demodulated" ..................................................................................... 49 5.1.1 0 Original image after transmitted and received through satellite and adding some extra A WGN. "After modulated and demodulated" ...................................................................................... 49 5.1.11 Received image ....................................................................................................................... 50 5.1.12 Received noisy image "More AWGN" ................................................................................. 50 5.1.13 Make a decision in the first time the threshold: detect oil spills ............................................ 51 5.1.14 Make a decision in the first time the threshold: detect: detect noise ............................. 51 5 .1.15 Detecting oil leak by using RSDCT ....................................................................................... 54 5.1.16 Detecting oil leak by using RSDCT "more A WGN" ............................................................. 54 5.1.17 The method of scanning an image .................................................................................... 55 5.1.18 One row is used to calculate false alarm and power curve ..................................................... 55 5.1.19 Performance Characteristics of Sequential Algorithms using two pairs (q=0.3, p=0.075) and four different thresholds (t =200, I 00,50,and 25). Detect from q to p ................................... 57 X

PAGE 10

5.1.20 Performance Characteristics of Sequential Algorithms using two pairs (q=0.3, p=0.075) and 5.1.21 5.1.22 5.1.23 5.2.1 5.2.2 5.2.3 5.2.4 5.2.5 5.2.6 5.2.7 5.2.8 5.2.9 5.2.10 four different thresholds (t = 150,90,60,and 30). Detect from p to q ...................................... 57 Results of choosing wrong thresholds .................................................................................... 58 Performance Characteristics of Sequential Algorithms using two pairs (q=0.3, p=0.075) and (q=0.2, p=0.055); Detect from q top ...................................................................................... 59 Performance Characteristics of Sequential Algorithms using two pairs (q=0.3, p=0.075) and (q=0.2, p=0.055); Detect from p to q ..................................................................................... 59 V AD integrated in a telecommunication system.................................................................... 62 The general operation flowchart of the V AD algorithm ........................................................ 63 Distributions Voice Signals .................................................................................................... 64 Actual noiseless voice signal "silence + active voice" .......................................................... 64 Actual Voice Noisy Voice Signal .......................................................................................... 64 Making the fma1 decision in the first crossing to the threshold; detecting a change (a) from f0 (x) to f1 (x) (b) from f1 (x) to f0 (x) ..................................................................................... 66 Signal-to-Noise Ratio SNR ..................................................................................................... 69 Power and false alarm curves V s different thresholds............................................................ 71 Making the final decision in the first crossing to the threshold, (a) Cyber Exploits Present, denoted HI (b) Cyber Exploits Absent, denoted HO ............................................................ 73 Original Voice Signal .............................................................................................................. 75 5.2.11 Various noisy environments ................................................................................................... 76 5.2.12 Original signal corrupted by AWGN" SNR=25,20,15,10 and 5 dB" ................................... 76 5.2.13 Results of adding noise to the original speech signal (5dB SNR). (a) Clean speech. (b) Wind Noise. (c) Babble Noise. (d) Computer Fan Noise. (e) Flowing Traffic. (t) Train Passing Noise. (g) Noisy Signal "Noise: AWGN". (h) Noisy Signal "Noise: Wind". (j) Noisy Signal "Noise: Computer Fan". (k) Noisy Signal "Noise: Flowing Traffic". (I) Noisy Signal "Noise: Train Passing Cut" .............................................................................. 76 5.2.14 SDCT-VAD results" Original Voice Signal corrupted by AWGN (SNR=25dB)" ............. 77 5.2.15 SDCT-VAD results "Original Voice Signal corrupted by AWGN (SNR=5dB)" ................. 77 5.2.16 Pe's of the proposed RSDCA-VAD various environmental conditions ............................... 80 5.2.17 Pc's of the proposed RSDCAV AD various environmental conditions ............................... 80 XI

PAGE 11

5.2.18 (a) Original Signal (b) Noisy signal with SNR=15dB (c) Noisy signal with SNR=IOdB (d) Noisy signal with SNR=5dB ................................................................................................. 82 5.2.19 Cyber Detection during speech activity detected periods using the CA-SDCT (a) Sequential Test: SNR=15. (b) Sequential Test: SNR=IO. (c) Sequential Test: SNR=5 .......................... 83 5.2.20 Cyber Detection during speech activity detected periods using RSDA-CA (a) Alarms SNR=I5. (b) Alarms SNR=lO. (c) Alarms SNR=5 ............................................................... 84 5.3.1 Model: Detect Cyber Attack using BER-Threshold................................................ 86 5.3.2 QPSK Modulation Vs Cyber Attack using BER-Threshold (a) No cyber attack detected (Green alarm) (b) and (c) minor warning (Yellow Alarm) (d) Cyber attack detected (Red Alarm) ..................................................................................................... 89 5.3.3 64QAM Modulation Vs Cyber Attack using BER-Threshold (a) No cyber attack detected (Green alarm) (b) and (c) minor warning (Yellow Alarm) (d) Cyber attack detected (Red Alarm) ...................................................................................................... 90 5.3.4 The received noisy speech signal ....................................................................... 92 5.3.5 Two-Dimensional (2-D) Time-vs-Frequency Spectrogram ....................................... 92 Fig. A Evaluate the whole updating step algorithm ................................................. 96 xii

PAGE 12

LIST OF TABLES Table 2.1 Show the differences of the quantization step ......................................................................... 13 2.2 Show the differences of the quantization step ......................................................................... 14 2.3 Quality of the quantized Signal. .............................................................................................. 15 3 .I Abbreviation of different modulation schemes ....................................................................... 17 4.1 Coding Gain ............................................................................................................................ 32 5.2.1 Comparing the Starting and Ending detection time instances of the Noisy Active Voice Messages Using the Manual and Proposed meth ..................................................................... 78 5.2.2 Pc's and Ws OF the Proposed RSDCA-VAD for Various Environmental Conditions .......... 79 5.2.3 Pc's Of The Proposed RSDCT-VAD, and Different V AD Approaches for Various Environmental Conditions ................................................................................................................................. 81 5.3.1 Frequency Bands ......................................................................................... 92 xiii

PAGE 13

1. Introduction to Digital Communication Systems The rapid progression of satellite communication technology has made available a new variety of services. Satellite systems are no longer dedicated solely to military transmission; they are also offered to public, resulting in increased frequency spectrum demand. In addition, to maximize the services offered through satellite communications, the available frequency spectrum must be used efficiently. Reliable transmissions of data and integrated services also have gained a lot of importance in today's information era. There is great demand for reliability and speed in the transmission of information. Spectrally efficient linear modulation schemes increase the effective rate of information transmission, without spectral increase. Several widely used linear modulation schemes are spectrally efficient, simple to implement and characterized by moderately good bit-error-rate (BER) performance. However, a tradeoff between power and spectral efficiency arises. Power efficiency is a major issue in satellite communications. To keep the size ofthe HPA (High Power Amplifier) at each terminal station somehow limited, the power consumption must be minimized. To minimize the power required for achieving a certain BER, error control techniques may be applied. The combination of Trellis encoding and Viterbi decoding techniques may be very helpful in this case. Furthermore, considerable gains may be attained if various modulation and coding schemes may be incorporated into the trellis-coded modulation (TCM) encoding. TCM was first discover by Ungerboeck [I] who showed that, in the case of Additive White Gaussian Noise (A WGN) transmission channel, gains of up to 6 dB were achievable without any bandwidth expansion, and that most of the coding gain could be obtained by doubling the constellation. Digital modulation techniques are essential to many digital communication systems, whether a fiber system, a wireless communication system, or a satellite communication system is considered. In the last decade, research and development in digital modulation techniques have been very active with many promising results. There exist countless research efforts in digital communications [2], [3], [4], [5], and [6]. Each such effort includes deployment of digital modulation techniques. This thesis provides up-to-date information of most modulation techniques in digital communication systems. It presents principles of most currently used digital modulation techniques and their applications, as well as new techniques currently being developed. For each modulation

PAGE 14

scheme, the following topics are covered: historical background, operation principles, bit error rate performance (power efficiency), bandwidth efficiency, constellations, comparisons, and applications. After the modulations and their performances in the A WGN channel are presented, system performance is studied and evaluated when an encoder and a decoder are added. The principal advantage of the approach deployed in this thesis is that the same encoder/decoder may be employed for a wide variety of trellis codes. The system may generate different trellis-coded signals using the same rate convolutional encoder. The adaptive nature of the approach allows for lowering the coding rate while keeping the same symbol transmission rate during deep fades. Most communication systems fall into one of three categories: Bandwidth efficient, power efficient, or cost efficient. Bandwidth efficiency shows the ability of a modulation scheme to accommodate data within a limited bandwidth. Power efficiency describes the ability of the system to reliably send information at the lowest practical power level. In most systems, there is a high priority on bandwidth efficiency. Each particular system can be designed and optimized, as dictated by the demands of its application. In this thesis, we will evaluate different modulation techniques in terms of their combined bandwidth, power and cost efficiency. We specify three design stages: algorithmic, architectural and implementational. The conceptual links between these stages represent design constraints, such as throughput, bit error rate, bandwidth efficiency, flexibility, latency, power-savings, and complexity and so on. These constraints determine the choice of the architectural and ultimately the implementational platforms. A comparison is carried out between flexible designs, which decode both uncoded and TCM coded data and provide two or three transmission rates. It is concluded that the larger number of rates is more beneficial from a cost-flexibility viewpoint. In this thesis, the Viterbi algorithm is chosen for decoding, since it provides a good trade-off between achievable coding gain and implementation complexity. Towards the end of the 1970's, Ungerbock [I] addressed the issue of bandwidth expansion by combining coding and modulation. According to him, ''redundancy" is now provided by using an expanded signal set and the coding is done directly on the signal sequences. In this thesis, a scheme consisting of trellis-coded (Multiple Phase Shift Keying) MPSK, (Multiple Frequency Shift Keying) MFSK, (Differential Phase Shift Keying) DPSK and (Multiple Quadrature Amplitude Modulation) 2

PAGE 15

MQAM modulated signals, A WGN fading transmission channels and the corresponding Viterbi decoder is proposed. The scheme realizes a family of differentrates codes, applicable to various channel conditions. During poor channel conditions, TCM is good choice in these cases to be employed. Theoretical bounds for the error performance and throughput of the proposed adaptive scheme are derived. Simulations have also been performed to measure the performance of the scheme for different parameter values and non-ideal conditions. It is shown that TCM results in considerable improvement in bit-error-rate (BER) performance of MPSK, MFSK, DPSK and MQAM signals. Under ideal conditions, gains in the range of 3 -I 0 dB are achieved over conventional fixed rate trellis-coded schemes. After showing how the digital communication systems perform using various modulation techniques and some channel coding, selected satellite and communication systems applications will be presented. Images and speech signal processing will also be used. To evaluate systems robustness, A WGN channel will be employed. An efficient and effective monitoring and detection of change algorithm is introduced. An innovative satellitedetecting framework is demonstrated, including satellite communication link configuration, satellite images transmission, enhancement and segmentation. Finally the sequential detection of change algorithm (SDCA) is proposed to detect al changes. A novel Voice activity detection algorithm is introduced and analyzed. Then, a new method to detect cyber attacks using bit-error rateThreshold BERT is developed. This Thesis is organized into five chapters. Chapter I provides a review of digital communication systems. Chapters 2, 3, and 4 describe the development, design, analysis and simulation of the different modulation techniques considered and the TCM encoding method. In Chapter 5, some Satellite and communication systems' applications are introduced. Chapter I is an introduction to digital communication systems, Information format, Data compression, Baseband Signal, Channel Coding and Bandpass Signals. Chapter 2 discusses signal processing methods with applications for images and voice. We demonstrate how analog information sources can be transformed into digital sources through the use of sampling and quantization. We also discuss the importance of baseband modulation due to its use in short distance data communications, as well as due to it being the front end of bandpass modulations. 3

PAGE 16

Chapter 3 covers baseband signal modulation schemes. Various different modulation techniques and their performances are discussed, while simulation setups and results are also included. Chapter 4 is a brief review of trellis-coded modulation. Different TCM encoder examples are given for illustration. Applying appropriate code design criteria, optimum trellis coded schemes for various modulation techniques are presented and the enhancement of the BER performance due to the TCM is measured. This chapter has been dedicated to the detailed error performance analysis of the pragmatic schemes deployed in adaptive systems, both in A WGN and Rayleigh fading transmission channels. The details of simulation programs are explained, and the results of simulations are reported. Simulations have been run for various parameters to take into account the effects of non-ideal conditions. The simulation results show that the performance gain between the BER-Uncoded and the BER-TCM-Coded is approximately 5-6 dB. In Chapter 5, the ideas and the scope presented in this thesis will be used on real life applications. 1.1 Digital Communication Systems Figure 1.1 is the block diagram of a typical digital communication system. The message to be sent may be from an analog source (e.g., voice) or from a digital source (e.g., computer data). Information Digitallnput sor ,---------, ,---------, Source Channel f-Modulator f-encode encode Format Information source Digital output rI : I I M-FSK M-QAM 1 QPSK I DPSK MSK BPF Channel Source decode Format Charmel decode [ Demodulator HL__BP-F ___j[l+- ----' Figure 1.1 Block diagram of a typical digital communication system 4

PAGE 17

An information source can be either analog or digital. To ensure that the source signal is compatible with digital processing, we transform analog information sources into digital sources via the use of sampling and quantization. These techniques are called either formatting or source coding. The digital sources are considered being represented in the logical format of binary ones and zeros. Figure 1.2 exhibits the formatting and transmission of baseband signals, where, I. Data is already in a digital format. 2. Textual information is transformed into binary digits by use of an encoder. 3. Analog information is formatted using three separate processes: sampling, quantization and coding. The results for all three cases are binary digits, or "bit streams". The bit streams are then transformed into pulse wave sequences via pulse modulation and the bit streams are recovered at the receiver via a demodulator. The analog-to-digital (AID) converter samples and quantizes the analog signal and represents the samples by binary sequences (bits I or 0). The source encoder accepts these binary sequences (digital signal) and encodes them into generally shorter sequences. The latter process is called source encoding or data compression; it compresses the signal by reducing redundancy, hence reducing the transmission speed, and thus reduces the signal bandwidth. The channel encoder accepts the output to the source encoder compressed signal and encodes it into a longer digital signal, adding error correction bits. Additional bits are intentionally added into the compressed encoded digital signal, so that some of the errors caused by the noise during transmission through the channel can be corrected at the receiver. Frequently, the transmission is in a high frequency passband, the modulator thus impresses the encoded digital symbols onto a carrier. Usually there is a power amplifier following the modulator. For high-frequency transmission, modulation and demodulation are usually performed in the intermediate frequency (IF). If this is the case, a frequency up-converter is inserted between the modulator and the power amplifier. For wireless and satellite systems, an antenna is the final stage of the transmitter. The transmission medium is usually called the channel, where, for satellite transmissions, noise is added to the signal, where fading and attenuation effects appear as a complex multiplicative factor on the signal. The term noise is used to represent a variety of random electrical disturbances caused from within and outside the system sources. The channel noise normally possesses limited frequency bandwidth, so that it can be viewed as a filter. At the receiver, virtually the reverse 5

PAGE 18

signal processing happens. First the received weak signal is amplified (and down-converted if needed) and demodulated. Then the added due to error bits redundancy is removed by the channel decoder, and the source decoder recovers the signal to its original form before being sent to the user. A digital to analog (D/ A) converter is needed for analog signals. Digital Infonnation : Textual -information '---------Analog infonnation Textual infonnation Digital lnfonnation Format --------Figure 1.2 Formatting and transmission of baseband signals The digital modulator maps the digital information sequences to corresponding analog radio waveforms. In baseband representation, information sequences are mapped to a complex signal constellation and then transmitted from either a single antenna or multiple antennas. The information source, source encoder, channel encoder and modulator are collectively known as the transmitter. The physical medium over which the signals are transferred from the transmitter to the receiver is known as the channel. An ideal channel has no fading or other channel perturbations. The only concern for the receiver operating on an ideal channel is the disturbance caused by the presence of thermal noise primarily due to the receiver amplifier. The thermal base band noise is modeled as additive white Gaussian noise (A WGN). In a satellite communication system, the channel is usually more complicated than the simple A WGN model. For example, the fixed-access line of a sight digital microwave radio channel is a multi-path fading channel. In such a channel, the received signal is the linear combination of 6

PAGE 19

components arriving via multiple channel paths reflected by obstacles such as trees, buildings or atmospheric disturbances. If multiple transmit antennas and receive antennas are used, the reflected radio signals typically travel over several uncorrelated propagation paths and arrive at different receive antennas with different phases and signal levels. The block diagram in Figure l.l is just a typical system configuration. For a multi-user system and a multi-station system, a multiplexing and multiple access control stage is added; before the modulator for a multi-user system case; and before the transmitter for multi-station system case. Therefore, a real system configuration could be more complicated. In addition, the system shown in Figure l.l can also be simpler if Source and Channel coding may be unnecessary. In the latter case, only the source, modulator, channel, demodulator, amplifier, and antennas, for wireless communication systems, are necessary. The fundamental objective of the communication system design is the effective delivery of information from transmitter to receiver, with acceptable information distortion, as dictated by the application. The channel model used most often is the A WGN channel. In an A WGN channel, independent identically distributed noise samples are added to the transmitted information symbols. The noise samples have a Gaussian distribution, i.e., the conditional density of the channel output y given the input xis given by, q(ylx) = _l_e-(y-x)f2u2 ilia (1.1) According to Shannon [7], reliable communication with arbitrarily low bit error rate (BER) in the A WGN channel can be achieved for transmission rates below (1.2) where W is the bandwidth occupied by the information bearing signal, S is the signal power and is the Gaussian noise variance. 7

PAGE 20

2. Signal Processing 2.1 Analog to Digital Conversion Information can be categorized in two forms: digital or analog. An analog signal is such that its amplitude may take any value within an open interval; thus the number of possible amplitude values is then infinite. Voice is analog and can take any number of volume levels within its dynamic-range. Digital devices convert analog voice to a digital signal (AID) by the process of sampling and quantization. The analog signal is first sampled and then quantized in a fmite number of levels. Each level is then converted into a binary sequence. For example, we may quantize voice in 16 levels, where each of these levels can be represented by four bits. The same process can be performed for image and video signals [8]. Nearly everything nowadays is digital. The medium is the environment that the signal travels through. It can be air, space or various types of wires. Each medium offers its own unique set of advantages and distortions that determine what will be used as a carrier. A signal through space, as in satellite transmission, may require a very high frequency carrier that can overcome space and other atmospheric losses. The carrier frequency may be otherwise light, as in optical fiber, or microwave, as in mobile communications. Most mediums dictate the type of carrier (its frequency, amplitude) that can propagate effectively through it and the type of distortions that the carrier will be affected by. Wireless carriers are always analog, while wired carriers can be both analog and digital. Communications inside a computer are examples of purely digital representations: digital data over digital medium. LAN communications are digital data over analog medium. The AM and FM radios are examples of analog data over analog medium. To convert an analog signal to a digital signal, the following three steps are required. First, the signal is passed through a lowpass filter to prevent aliasing. Second, the signal is sampled by a sample-and-hold circuit. Finally, the samples are quantized by an analog to digital converter (ADC) in order to be represented in digital form as shown in Figure .2.1. .t(/) Anti-aliasing Sample-and-Hold .t(rr) ----t AID Converter Analog Filter Circuit Signa; IJisa1:m. Figure 2.1 Typical analog to digital conversion process 8

PAGE 21

2.2 Quantization (Uniform Quantization) There are many different kinds of quantization techniques available. Many quantization methods are explored such as the linear quantization, the nonlinear quantization, the delta modulation, and the sigma-delta modulation. Figure 2.1 shows the block diagram of a speech coding system. The continuous time analog speech signal from a given source is digitized by an ordinary connection of filter (eliminates aliasing), sampler (discrete-time conversion), and analog-to digital converter (Wliform quantization is assumed). The output is a discrete-time digital speech signal. This signal is referred to as the digital speech. The encoded digital speech data is further processed by the channel encoder, providing error protection to the bit-stream before transmission to the commWlication channel, where various types of noise and interferences can damage the reliability of the transmitted data. In this thesis, we will use Wliform quantization for both image and voice signals. The results show the effect of increasing number of quantization levels on the quality of both voice and picture signals, Correlation, Histogram Error, Signal-to-Noise Ratio, and Rate Distortion. 2.3 .1 Image Quantization Figures 2.2 show a comparison between the four quantized images to the original, exhibiting how the quality of the image changes as different quantization levels are used. It can be noticed that some artifacts (errors) appear in the images, as the number of bits per signal sample is lowered. As the latter number decreases, the quality of the quantized picture worsens and the artifacts increase dramatically. 9

PAGE 22

50 100 150 200 250 300 350 400 450 5(l) Figure 2.2.a Original Image "Colorado Convention Center Denver'' Colorado Convention Center Denver Figure 2.2.b Quantized Image (7 b/pel) 10

PAGE 23

50 100 150 200 250 300 350 400 450 Colorado Convention Center Denver 100 200 300 400 Figure 2.2.c Quantized Image (4 b/pel) Colorado Convention Center Denver 50 100 150 200 250 300 400 450 500 600 1(10 200 300 400 500 600 Figure 2.2.d Quantized Image (2 b/pel) II

PAGE 24

Colorado Convention Center Denver 50 100 150 200 250 300 350 400 450 500 100 200 300 400 500 600 Figure 2.2.e Quantized Image (I b/pel) Table.2.1 Show the differences of the quantization step K Quantization Levels X min X max Q Step 1 2 -254 254 508 2 4 -254 254 169.334 4 16 -254 254 33.8667 7 128 -254 254 4 The image quality noticeably deteriorates when the number of bits equals I and 2 b/pel. The figures below show the difference between the original and the quantized images. Figures 2.3.(c and d) show how the image quality is completely destroyed due to the wrong choice of bit numbers per signal value representation. 12

PAGE 25

160 1 i 120 100 0 i 00 l 8) 8 20 1220 12<10 7260 72!ll 1m 7320 73l 7EJ 7EJ 7 m c 160 60 0 <10 20 Figure 2.3.a Original Vs Quantized Image (7 b/p) Figure 2.3.b Original Vs Quantized Image (4 b/p) 20 20 ,, 0 -----------------------72SJ 72!IJ 7Dl 7320 7340 7EJ 7Bl 7.QJ 7220 72 72SJ 72Eil 7DJ 7320 7340 7EJ 7HJ 7<100 Figure 2.3.c Original Vs Quantized Image (2 b/l) Figure 2.3.d Original Vs Quantized Image (l b/p) The same scenario is repeated again, where the information signal is voice instead of image. Table.2.2 shows the quality of the voice signal is decrease, as the number ofb/sample is reduced. Table.2.2 Show the differences of the quantization step k Quantization Levels Xmin Xmax QStep I 2 -254 254 1.5857 2 4 -254 254 0.5286 4 16 -0.7928 0.7928 0.1057 7 128 -0.7928 0.7928 0.0125 For each signal, there is a point that the signal quality deteriorates drastically and becomes incomprehensible, as shown in Table 2.3. 13

PAGE 26

Table.2.3 Quailty of the Quantized Signal number of 7 b/sample 4 b/sample 2 b/sample 1 b/sample b/sample speech. au Clear A Little bit distortion Not Good So bad music.au Clear Clear Not Good So bad The signal quality deteriorates drastically when the number of levels is less than 16, and it becomes completely incomprehensible when the number ofb/sample equals l (2 Levels). .. 06 -onginal I --quanlzed t -04 72:U 7240 72ro 7m 7DJ 7320 7340 7HJ 7m 7400 Figure 2.4.a Original and Quantized (Speech.au) (7 b/pel) 06 II I l lJ il -I -. 1-I ---quantzed I I -02 0 -04 I I I 7m 7240 72ro 7m 7DJ 7320 7340 7:W 7m 7400 Figure 2.4.a Original and Quantized (Speech.au) (2 b/pel) .. 0 06 -onginal 11 quanhzed -A 02 0 -0 4 I 7220 7240 12ro 1m 7DJ 7320 7340 7HJ 7m 7400 Figure 2.4.b Original and Quantized (4 b/pel)
PAGE 27

3. Modulation Techniques 3.1 Introduction Digital modulation techniques are necessary for many digital communication systems, as necessitated by the information transmission medium. In this, chapter, we discuss several modulation techniques that are applicable to digital communication systems. We present principles and applications information of most currently used digital modulation techniques, as well as new techniques that are currently being developed. We briefly discuss the role of modulation in a typical digital communication system, basic modulation methods, and criteria for choosing modulation schemes. For each modulation scheme, the following topics are covered: historical background, operation principles, bit error rate performance (power efficiency), bandwidth efficiency, block diagrams of modulator, demodulator, and constellation for different modulation schemes, comparison, and applications. After presenting modulation schemes and their performances in the A WGN channel, we discuss their performances when an encoder and decoder are added to the overall system. 3.2 Modulation Techniques To provide an overview, we list the abbreviations and descriptive names of the various digital modulation schemes that are listed in Table 3.1. Among the listed schemes, (Amplitude Shift Keying) ASK, (Phase Shift Keying) PSK, and (Frequency Shift Keying) FSK are basic, while (Minimum Shift Keying) MSK, (Gaussian Minimum Shift Keying) GMSK, (Quadrature Amplitude Modulation) QAM, etc. are advanced schemes. The advanced schemes are variations and combinations of the basic schemes. Constant envelope modulations such as FSK and GMSK offer not only enhanced spectral efficiency, they also provide an inherent transmitted power advantage. All constant envelope modulations allow transmitter's power amplifiers to operate at or near saturation levels. The constant envelope class is generally suitable for communication systems such as (Amplitude Shift Keying) ASK and (Binary Shift Keying) BSK; however, the generic (Frequency Shift Keying) FSK schemes in this class are inappropriate for satellite application since they have very low bandwidth efficiency in comparison to the (Phase Shift Keying) PSK schemes. Binary FSK is used in the low-rate control channels of first generation cellular systems. The PSK schemes, including (Binary Phase Shift Keying) BPSK, (Quadrature Phase Shift Keying) QPSK, (Offset Quadrature Phase Shift Keying) OQPSK, and 15

PAGE 28

(Minimum Shift Keying) MSK have been used in satellite communication systems. rr/4-QPSK is worth special attention due to its ability to avoid 180" abrupt phase shift and to enable differential demodulation. It has been used in digital mobile cellular systems, such as the United States digital cellular (USDC) system. MSK has excellent power and bandwidth efficiency. Its modulator and demodulator are also not too complex. ASK is generally not suitable for systems with nonlinear power amplifiers. QAM has been widely used in modems used in telephone networks, such as computer modems because it can achieve extremely high bandwidth efficiency. QAM can even be considered for satellite systems [9]. Table 3.1 Abbreviation of different modulation schemes Abbreviation Alternate Abbr Descriptive Name Frequency Shift Keying (FSK) BFSK I FSK Binary Frequency Shift Keying MFSK M-ary Frequency Shift Keying Phase Shift Keying (PSK) BPSK PSK Binary Phase Shift Keying QPSK 4PSK Quadrature Phase Shift Keying OQPSK Offset QPSK rr/4-QPSK rr/4 Phase Shift Keying MPSK M-ary Phase Shift Keying Amplitude and Amplitude /phase modulation ASK Amplitude Shift Keying QAM Quadrature Amplitude Modulation We begin our discussion of digital modulation by starting with the three basic forms of digital modulation techniques: frequency shift keying (FSK), amplitude shift keying (ASK), and phase shift keying (PSK). In all these techniques, the transmitted information modifies a single parameter of a sinusoidal waveform: either frequency, or amplitude, or phase. The sinusoidal waveform, called the carrier, travels then through the corresponding medium, where the latter may be wire, air, water and space. The transmission medium generally introduces corruptions to the traveling sinusoidal waveform, and thus to the transmitted information. Below, we discuss each of the above three basic modulation techniques and their performance. 16

PAGE 29

3.3 Frequency Shift Keying We first describe the binary FSK scheme. Then, we generalize it to the M-ary FSK (MFSK) 3.3.1 Binary FSK Signal In the most general form of FSK, the frequency of the carrier is modified to represent the transmitted information. In other words, the binary FSK scheme uses two sinusoidal signals possessing two different frequencies to represent bits I and 0. Bit I is transmitted by a sinusoidal carrier of one particular frequency, while, to transmit bit 0, the frequency of the carrier changes to a different specified frequency. In particular, bits I and 0 are respectively represented by the waveforms SJ(t) and S2(t), below. 51 (t) =A cos(2rcf1t + 41) kT :5; t :5; (k + 1)T.[or 1 (3.1) S2(t) =A cos(2rcf2t + ctl) kT :5; t :5; (k + 1)T.[or 0 (3.2) Where and cp is the initial phase at t = 0, A is the amplitude and T is the bit period 3.3.2 MFSK Signal In M-ary FSK modulation, the information binary data stream is divided into n-tuples of n = log2 M bits. We denote all M possible n-tuples the M distinct messages: li = 1,2, ... ,M .. M sinusoidal waveforms, with M distinct frequencies, represent then, each of the M messages. The waveform for the ith message is: kT :5; t :5; (k + 1)T.[or li (3.3) where T is the per message period corresponding to the transmission of n bits. If the initial phases are the same for all i, then the scheme is called coherent. As with the binary case, we can always assume ctli=O for coherent MFSK. The demodulation may be coherent or non-coherent. Otherwise the transmitted signal set may be non-coherent, where the demodulation sheme must be then non-coherent. 17

PAGE 30

----+-4DPSK -&-BFSK ---4FSK ---iii-QPSK 0 5 10 15 Figure 3.1 UncodedBER for different modulation schemes 3.4 Binary Phase Shift Keying (BPSK) In BFSK, bits are represented by two sinusoidal waveforms possessing two distinct phases. Typically, these two phases are 0 and rr. Let us denote the two binary sinusoidal representations 51 and 52 Then, 51 (t) =A cos(2rrft + 0) kT :5 t :5 (k + 1)T,for 1 (3.4) S2(t) =A cos(2rrft + 180) kT :5 t :5 (k + 1)T,for 0 (3.5) In the above representation, the information bit is represented by a 180 degrees phase-change in a sinusoidal signal and the two bit representations are then called antipodal. This phase choice corresponds to a signal design that minimizes the probability of error in A WGN transmission, inducing a correlation coefficient of (-I). 18

PAGE 31

In BPSK, the unit circle is 2-quantized. As a generalization, M -quantized levels of 2rr may be deployed, to create a variety of PSK modulation schemes. Given M, let i be a nwnber from 1 to M. The allowed phases are then given by the following modulating angles. 2xi (J. =1 M (3.6) In the above expression. M stands for the order of the modulation. M = 2, results in a BPSK scheme, M = 4 represents a QPSK scheme, and so on. We note that all PSK signals may be graphically represented by a signal constellation in a two-dimensional coordinate system. The following diagram shows some of the MPSK modulation schemes and their "constellations." C...elatianb2PSf< Consl....,.faJOP'SI< 05 0.5 ; ; :; ofe .. o t 6 6 .()5 .()5 _, _, .05 0 05 _, .()5 0 05 tn-Pha .. Const.._,. ior 8PSI< Cons:l .... ion b' 16PSK 05 05 ; ; i o o ... 6 J .05 .as _, -1 .()5 0 05 tn-Ptt. -1 .()5 0 05 ..__ c....a.aetron 111r J2P'S)( -. 05 ; j of .()5 _, -1 .()5 0 05 ln-PheM Figure 3.2 Signal Constelation for Binary, Quadrature, 8, and 16 PSK 19

PAGE 32

As compared to the BPSK, for M larger than 2, the MPSK decreases the signal bandwidth. Indeed, in BPSK, a single bit is represented by a single sinusoidal waveform, while such a waveform represents n = log2 M bits in MPSK. 3.4.2 Quadrature PSK Among all the MPSK schemes, QPSK is the most frequently used because it does not suffer from Bit Error Rate (BER) degradation, while the bandwidth efficiency is sufficiently increased, as compared to the BPSK. Other MPSK schemes, on the other hand, increase bandwidth efficiency at the expenses ofBER performance [9]. In this section we will study QPSK in great detail. If the transmission rate of the symbols is the same in QPSK and BPSK, it is intuitively obvious that BPSK transmits data half as fast as QPSK does. At the same time, we observe that the distance of adjacent points in the QPSK constellation is less than that of the BPSK. In comparison to the BPSK, this causes demodulation problems, where the distinction of symbols worsens and the per symbol error performance thus degrades and so consequently does the bit error rate. However, as shown in the figure below, the bit error probability remains the same. Over the last two decades, communication networks have evolved from the analog signal transmitting telephone and radio transmissions to the modem digital communication systems. These digital networks are required to carry large amounts of data at high rates. The need for increased bandwidth has caused an evolution in the media from wires and air. Service providers must guarantee data integrity to their clients, where bit-error rate measurements are used to measure error resistance performance. The pertinent criterion is called bit-error rate (BER) and it plays an important role in measuring the quality of service delivered by a network. The BER definition is given below. Number of changed bits BER = -------Total Number of bits (3.7) Under ergodicity conditions, the BER converges to the bit error probability induced by the channel of transmission, and it may be then monitored to detect changes in the quality of the communication link. The BER performance is affected by many factors such as power, noise, or the deployed modulation method. If we assume specific system power, known noise and given modulation technique, we can 20

PAGE 33

predict how the BER performance. The BER measurement is not complicated; it just requires the transmission of a bit stream through the communication system and the comparison between input and output bits. Noise is the main enemy of BER performance. The noise introduced by an atmospheric transmission medium is frequently described with a Gaussian probability density function, while the signal path is usually described with a Rayleigh probability density function. A Rayleigh, or fading, signal path is not "noise" in the intuitive sense of the familiar hissing sound of "white noise," but it is a random process that is analyzed in the same manner as Gaussian noise. Without going into details, the mathematical representations of these functions represent a system model and allow for the system analysis and performance prediction. As compared to PSK, the QPSK increases the data rate at the gain of decreased BER. As compared to the QPSK, in 16-PSK, the signal space is subdivided into smaller regions. 16 sinusoidal signals or symbols are then available, where each symbol represents 4 bits. The bit rate is now four times that of the BPSK for the same symbol rate. Figure 3.3 shows the 16-PSK signal at various stages during modulation . . . . : .......... : ......... ......... : .......... : ........ . . : 1 o8 .__ ____. __ __._ __ _._ __ ....._ ___. __ __,._ __ __._,_ _.__...._ ___. 0 2 4 6 8 10 12 14 16 18 Eb/NO (dB) Figure 3.3 Un-coded BER for DPSK, 2, 4, 8, 16, 32-PSK 21

PAGE 34

3.5 n /4-QPSK-a variation on both QPSK and 8-PSK Il/4-DQPSK has been designated as the American standard of the second-generation cellular mobile communications. It is a variation of the QPSK that mimics 8-PSK. Like QPSK, rr/4-QPSK transmits two bits per symbol. So only four carrier signals are needed but this is where the twist comes in. In QPSK we have four signals that are used to send the four twobitlength symbols. In rr/4QPSK, we have eight signals, instead: every alternate symbol is transmitted using a rr/4shifted pattern of the QPSP constellation. As shown below, a symbol at (45, 135,225, -45) uses a signal on this path, while, even if the pattern remains unchanged, the next symbol uses path (0, 90, 180, 270). Thus, a phase shift always occurs, even when adjacent symbols are identical. The constellation diagram looks similar to the 8-PSK. Note that a 8-PSK constellation can be broken into two QPSK constellations as show below. In rr/4-QPSK, one symbol is transmitted on the first type constellation and the next one is transmitted using the second type constellation. Even though the constellation looks like 8-PSK, on the network analyzer, this modulation is strictly a form of QPSK with same BER and bandwidth. Although the symbols move around, they always convey just 2 bits per symbol. Constellation for (pi/4}-DPSK Constellation for (pi.oti}-DPSK 0.5 0.5 .. .. :; :; t: 0 t: 0 ... "0 "0 '" .. " 0 0 .05 .5 .05 0 0.5 .0.5 0 0.5 In-Phase In-Phase Figure 3.4 Signal Constellations for 1t/4 and 1t/8 DQPK 22

PAGE 35

3.6 Quadrature Amplitude ModulationS At this point, all the passband modulation schemes we have studied, MFSK. MPSK, and DPSK are constant envelope schemes. The constant envelope property of these schemes is especially important to systems with power amplifiers which must operate in the nonlinear region of the input output characteristic for maximum power efficiency. Such are the satellite transponders. For some other communication systems, constant envelope may not be a crucial requirement, whereas bandwidth efficiency is more important. Quadrature Amplitude Modulation (QAM) is such a class of non constant envelope schemes that can achieve higher than the MPSK bandwidth efficiency, for the same average signal power. QAM is widely used in modems designed for telephone channels. The telephone circuit modem standards are all based on various QAM schemes ranging from uncoded 16-QAM to trellis coded 128-QAM. The research ofQAM applications in satellite systems, point-to-point wireless systems, and mobile cellular telephone also systems have been very active. J. I -1. .J. -3 -6 ., 0 2 6 ....... 10 2 I<::::::::::: 0 -2. 0 -10 ............ 0 ....,_ 10 l ; i 0 6. l 0 .., .... IS .o ... ................ 'I(IO e e e. 0 e eo eo e 1 iJi --...... ; ............... j 0: ::: :: :: :: : : : : : .. .......... ................ -W 15 ............. .. -15 -10 ..; (] 5 10 ....,_ Figure 3.5 Signal Constellations for 16, 32, 64, 128, and 256-QAM 23

PAGE 36

QAM is a combination of amplitude and phase modulation and its development may be attributed to the following logic: In MPSK schemes, signals have the same amplitude but different phases. It may be natural to consider both amplitude and phase modulations (QAM) as the next development step, where the transmitted signals are: i = 1,2, ... ,M (3.8) where Ai is the amplitude and cp1 is the phase of the ith signal in the M-ary signal set. Similarly to the MPSK, a geometric representation called constellation is a very clear way of describing a QAM signal set. A QAM signal is represented by a point or vector, or phasor. The two axes sometimes are simply labeled as 1-axis and Q-axis. Figure 3.6 shows different types ofQAM constellations. 10 ........ ........ : .................. : ....... ":" ......... : ......... : ........... .. .... . . . . :-. . . . . . . . . . ......... ......... : ......... : ......... : ......... -: ........ -. . --e--16QAM : : -a-32QAM ---+-64QAM 10 ..... ----A-128QAM ----+---256QM1 2 4 . . 6 8 10 EbfNo (dB) 12 14 16 Figure 3.6 Uncoded BER Signal Constellations for MQAM 24 18

PAGE 37

4. Channel Coding 4.1 Trellis Coded Modulation (TCM) 4.1.1 Trellis Coded Modulation Background Trellis Coded Modulation (TCM), introduced by Ungerboeck [I], is a very effective method for reducing the required power without any increase in the bandwidth requirement. The innovative aspect of TCM is the concept that encoding and modulation should not be treated as separate entities, but rather, as a unique operation. We usually consider coding and modulation as two separate stages in a communication connection, while in TCM the two stages are united. Trellis Coded Modulation (TCM) is a relatively complex concept, especially due to the nonlinear nature of its operation. TCM belongs in the class of convolutional codes and has been applied for transmissions through telephone, satellite and microwave digital radio channels, where coding gains of the order of 3-6dB may be obtained with no loss in bandwidth or data rate [8] and [9]. Generally, the Hamming distance between binary representations of two signals does not possess a direct translation to their distance in the signal/symbol space (after modulation). It may be concluded, therefore, that the Hamming distance is not the correct distance representation between different symbols. On the other hand, the geometric distance between the signals or their Euclidean distance (ED) may be the appropriate measure. Figure 1.1 is a block diagram of a communication link. Here, we assume that the output of the source is time-discrete. The output of the source is first encoded for error correction after transmission through a distortion-inflicting medium. This encoding induces the addition of some redundant symbols to a group of raw source symbols. The encoded data stream is then modulated and sent over the channel (using Trellis). The objective of modulation is the transformation of encoded (for error correction) symbols into a signal-form that is suitable for transmission through the available channel. For A WGN channels with fading, noise is added to the latter signal-form, while other artifacts are also inflicted upon it. At the receiver, the received noisy signal is first demodulated and the encoded symbols are recovered with some error. Then, the decoder (Viterbi Decoder) attempts to correct the errors using the extra information available due to the redundancy bits added by the channel encoder (Trellis Encoder). 25

PAGE 38

Digilal Infonnation Textual ---------------, infonnation r --------Anllloz infonnation Sample 11-:s -l'-i _; 1 ... t I Quantize Enoode Fonnat 11 Gl l' l .. B 25 a-' II' IIG !i:IO b-IOo C-Olo d-1 ]o r ------: : t I I . I +e> 0 Channel t-+----+1 encode Modulator L.-------: 00 00 00 00 00 10 10 ---.. -.. --"----'-----------------..J Figure 4.1 Source and channel coding As of the results in the information theory of Shannon, the best system performance can be obtained when codes for long message sequences are designed, as long as the transmission rate remains below the channel capacity. The receiver decides then among different long message/symbol sequences rather than making per symbol decisions. The induced probability of error is inversely proportional to the length of the symbol sequences decoded. TCM follows the Shannon principle: At the digital level, it collects a block of bits and encodes them by inserting extra error-correcting bits. The so extended binary sequences are then modulated for conversion to a analog form via the use of a sinusoidal carrier. 4.1.1 Fundamentals and Concept ofTCM The functions of a TCM consist of a Trellis code and a constellation mapper as shown in Figure 4.2. TCM combines the functions of a convolutional coder of rate R = k / k + 1 and a Mary signal mapper that maps M = 2k input points into a larger constellation of M = 2k + 1 constellation points. 26

PAGE 39

Trellis Code K bits -. -. -. -. -j K+ 1 bits I Convolutional Encoder Rate k/(k+ 1) Constellation Mapper r--. MFSKorDPSK ----MPSK t----I MASK j. I I MQAM I --l-Carrier(.) Figure 4.2 Trellis Coded Modulation Symbols Fork= 2, we have a code of rate 2/3 that takes a 4PSK signal (M = 4) and produces a 8-PSK signal (M = 8). Thus, instead of expanding the bandwidth, as the signal transfonns from 4PSK to 8PSK, it doubles the constellation points. The same scenario applies if we select k=4, and we choose a code rate 4/5 that takes a 16-QAM signal (M=l6) and produces a 32-QAM signal (M = 32). Thus, instead of expanding the bandwidth, as the signal transfonns from 16-QAM to 32-QAM, the system is upgraded to one with a larger number of constellation points as shown in Figure 4.3 below. 27

PAGE 40

! :::J ii -<; "' :::J 0 Constellation for QPSK 0.5 : !!! = o. !! ..., "' :::> 0 -0.5 -1 -1 -0.5 0 0.5 11"1-Phase Conslellalion for 160AM 3 .. 3 In-Phase 0.5 ole ore -0.5 -1 .. 1 :; 0 .. 0-1 -1 Constellation for 8PSK -0.5 0 05 11"1-Phase ConsltUalion for 320AM In-Phase Figure 4.3-Constellation doubling in TCM. A (QPSK and8PSK); (16 and 32-QAM) signal transmitted 4.1.3 Mapping and Trellis Diagram The figure above shows Trellis Coded Modulation with different code rates. The coding adds just one extra bit to the symbol bit size. The symbol size increases from k bits to k + l bits which means that the constellation size doubles. Notice that the code rate may not be only l/2, 2/3, and 3/4, but it could be anything: 3/8, for instance. Assuming that the original signal is BFSK, then, a TCM encoder will produce a QFSK, a QFSK will become SFSK, and, if 8PSK was selected, a l6QAM signal will result. 28

PAGE 41

1
PAGE 42

As constellation expands while the signal energy is kept the same and the distance between the symbols decreases, Let us assume that we are given some amount of power and we want to calculate the needed signal BER. Figure 4.5 shows the improvement in BER and power performance if TCM is deployed. ffi 101 CD c c c-.-.. ---.--e--QPSK ---&-16QAM -2560AM Figure 4.5 Uncoded BER vs. Eb/NO in A WGN environment Let us assume the transmission of a BER of 10-4 encoded QPSK signal. If this power level is not available, another option is to add a code of rate 2/3 to reduce the BER and thus the subsequent Eb/N0 requirement. However, another problem arises then. If we keep the same bit rate for the information bits and allow the coded bit rate increase to accommodate the overhead bits, then the bandwidth requirement will increase by an amount inversely proportional to the code rate increase. Thus, addition of coding increases the bandwidth by 3/2. If bandwidth change is not allowed, then the information rate will have to decrease by the same proportion. From Figure 4.6, we observe that due to the use of TCM, 4-6 dB can be saved. For example, let us pick QPSK and compare Figures 4.5 and 4.6 It can be observed that due to the use of TCM, with a BER of 10-2 ,a 5 dB gain could be achieved, meaning that less power is required for the delivery of the same BER. From the same figures, we also notice that, with a BER of 10-2 instead of QPSK, we can instead use 16-QAM with the same power, at the expense of information rate increase. 30

PAGE 43

Modulation scheme ll: w aJ QPSK 16QAM 32QAM 256QAM Table 4.1 Coding Gain Coding gain at 10-2 over Coding gain at 10-2 over Coded Uncoded System System 4dB 1.5dB 8.2dB 2.3dB lldB 5.5dB 16.3dB 11.5dB EbNo Figure 4.6 Coded BER vs. Eb/NO in A WGN environment A rate of a 2/3 convolutional encoder is used as an example to analyze the trellis diagram and mapping method. Figure 4.7 illustrates a convolutional encoder. The input of this TCM encoder is a 4-bit symbol, and the output is a 6-bit symbol. The output of this encoder consists of n bits. These bits are used to choose one of zn partitions in the constellation. This means that the constellation has been 31

PAGE 44

partitioned into 2n subsets. After the constellation is chosen, we need to select the modulation technique. Figure 4.7 shows some the results of this process. 4.1. Viterbi Algorithm The TCM encoder is illustrated using a trellis whose branches are associated with transitions between encoder states and codeword transmitted over the channel. The primary task of the TCM decoder is to estimate the path that the codeword sequence traverses through the trellis. In this manner, TCM decoder is a reverse process of TCM encoder. In addition to the convolutional decoding, the de mapping algorithm is a reverse function of the mapping logic function and the differential decoder performs the reverse function of the differential encoder. The decoder algorithm used in this thesis is based on the Viterbi algorithm. Andrew Viterbi proposed an algorithm in 1967, to decode convolutional codes and this became the Viterbi Algorithm [10]. This algorithm is an application of dynamic programming that finds shortest paths. (maximum likelihood sequences) widely used in solving minimization problems. A critical feature of this algorithm is the complexity of the decoding process grows linearly with the number of symbols being transmitted, rather than exponentially with the number of the transmitted symbols. The Viterbi algorithm uses a metric and tracks this metric for several trellis paths at once. The path with larger metric is dropped when it merges with another. In hard-decision Viterbi decoding, this is done using the Hamming distance as a metric. In TCM the decoding is done with soft-decision algorithm and Euclidean distance is used as the metric. The objective is to track n possible sequences, keep track of cumulative MSEDs. When paths merge at a state, follow only the one with the smallest metric. 32

PAGE 45

4.2 TCM MFSK, MPSK, DPSK, and MQAM Constellation Mapping 4,.; --1------------------i -. Tool I ..... ...;:-----""'I 121!Y..,,.-L-J-,L,-L--J ,.,.. __
PAGE 46

( fd:) Pe Nfed Q (4.1) Where Nred denotes the number of signals sequences with distance Dred that diverge at any state and remerge at the state after one or more transitions. Recall that increasing dmin lowers the BER. In trellis-coded modulation (TCM), this increase is achieved by partitioning a constellation into subsets. This split is done according to Ungerbock's rules Figure, every split of a uniform 16-QAM and 8-PSK constellation increases dmln in the subset by a factor If the lattice at the root has distance I between constellation points, the distance in the subsets at the lowest branch is This distance is the ultimate limit on how well a TCM code can perform. 00 00 oe oo 0000 oooe + e e e e I e e e e . ---.-_J---.-e I eel el e . j. i. ! ... --_,_ .,+-----'-.1---+ .... .... I j 1 I I A-640M1 __:___e ........ 1 e e e ,e e e e ee ee eeee eeee _) dB =dA .-ee ee eeee eeee eo eo oeoe eo eo oe oe 00 00 eo eo oeoe 0000 0000 eo eo oeoe 00 00 0000 eooo oo eo oooe 0000 00 00 0000 oo eo eo oo oeoo 00 00 0000 eeee oeoe 0000 oe oe 00 0 0 oooe 0000 o oo 0000 ... ---._ ee ee_ ........ oeoe eo eo o o eo eo oe oo 0000 oooe 0000 0000 eo eo 0000 eo eo 0000 oo eo 0000 eooo 0000 eooo 00 00 ooeo oooo oooo oooo oooo eooo oooo oooo ooeo ooeo oooo oeoo oooo oooo oooo oooo oooo oooo oeoo oooo oooe oooo oooo oooo oooo oooo oooo oooo oooo oooo ooeo eooo oooo oooo oooo oooo oooo oooo ooeo eooo oooo oooo oeoo oooo oooe oooo oooo oooo oooo oooe oooo oeoo oooo oooo oooo oooo oooo oooo oooo oooo oooo eooo oooo oooo ooo Figure 4.8 16QAM Set-partitioning according to Ungerbock. 34

PAGE 47

Ungerboeck [I] showed that using (Euclidean distance) ED as the metric in systems employing Trellis coding and Viterbi decoding, and designing codes to achieve maximum ED between all possible sequences of channel signals, leads to significant coding gains without any sacrifice in bandwidth or the effective information rate. In fact, it was shown [I] that nearly all the possible gain in performance is achieved by doubling the size of the signal space. This led to the design of rate njn + 1 codes with signal spaces of size M = 2n + 1. Figure 4.8 shows a 64-QAM constellation, which is divided into 16 subsets with 2 signal points per subset. This division stems from successive splits as in Figure 4.8, and the bit-to-constellation point mapping is derived from these splits. Ifthe at the lowest distance is 2a with (Square-Euclidean Distance) SED equal to 4a2 After that the same processes is repeated until the final stage (Last Level), 4a2 -+ 8a2 -+ 16a2 -+ 32a2 and so on, which means we are doubling each time. As shown in figure 4.8. The figure below shows how the 8 points of 8PSK are successively portioned into disjoint cosets such that the SEDs are increasing at each level. There is a total of four partitions, counting the frrst unpartitioned set. At the top-most level, the MSED (Minimum Squared Euclidean Distance) is 0.586. At the next level, where there are only four points in each of the two cosets, the MSED has increased to 2.0 and at the last level, the MSED is 4.0. Each subset is also called a coset and by the lattice terminology, we can show the partition with its coset generators in this way. In general, if M denotes the size of the master constellation, there are log2 M -1 possible splits to finally produce M /2 subsets with 2 signals. Of course, one could split one more time to arrive at one signal per subset. This transforms the set-partition problem back to mapping code symbols efficiently to constellation points. However, it is not considered a set-partition code in the traditional sense, and will not be discussed further. Each Version of TCM as created by Ungerboeck, requires a different rate code. It is assumed that for each input sequence of n bits, the trellis encoder produces a sequence of n + 1 coded bits which are then mapped into a single symbol chosen from an M-ary alphabet where M = zn+t. The modulation 35

PAGE 48

sets from which these M-ary symbols were chosen were typically multiple-frequency-shift-keying (MFSK), multiple-phase-shift-keying (MPSK) or quadrature amplitude modulation (QAM). In particular, they proposed an encoder with b binary input bits and s binary output symbols, which were mapped into k-Mary symbols (k is referred to as the multiplicity) in each transmission interval where s, k, and M obey the relation s = k log M. .. + I I d/ = 4 . I 0 o 1 o .. .. l d0 =0584 I ... . t 0 i 0 0 0 .... .. I I ,-.. i 0 -+ 0 i ' 1 ' 0 0 ' ' 0 0 0 I I I I I -+ i i 0 "io "A" "A" "A" oi" ' ' ' ' Figure 4.9 8PSK Set-partitioning according to Ungerbock. The main concept in multi-dimensionality is increasing the number of symbols created in one processing period. The transmitted symbols are generated together and this co-generation creates dependence and allows better performance. The term multi-dimensionality does not mean anything other than a form of multi-processing. The advantage of multi dimensional TCM are so many for instance we can transmit fractional information rates. Instead of the effective code rate being 2/3 as it is in I x 8PSK, here it can be higher. We can use code rates like 2/5, 2/6, 2/7and 2/8. Perhaps the most significant aspect of MTCM with regard to its application in a fading environment is its ability to provide improved diversity (the rate of descent of error probability with average bit energy-to-noise 36

PAGE 49

ratio) on such a channel. Figures 4.10-4.12 show the whole process of mapping two input bits to 2+n bits using different convolutional encoder rates. The performance of each rate versus different modulation techniques such as Multidimensional frequency shift keying (M-FSK), Multidimensional phase shift keying (M-PSK), and quadrature amplitude modulation M-QAM is shown in figure 4.13. .. i ,' .... -.. :-: .:*""'"-: --.. -!-.. -+L--ii __ ..,..., ...... .. ; >ii= .. = ... '_j_ -i .. ': !;!::...._' ____..__....__ !-----'----\ __________ : .. '. _______ / --i :... ..... /: ;,, nnounwnnn w -----------------------------------nnnnnnnmnn0'="' =='-------Figure 4.10 Conventional versus multiple trellis coded mod with Rate (a) 2/3 (Rate #1) and (b) Rate 2/5 (Rate #2) .. --0:---...... -. --+--.:.-.,.-,: -P'-.. *-. --+-H--i--:---_; ---:------t;--____,. .... mm w wwww .r!Jc;;....n ------+--'-----.IJ==.,=n=n = ---------_: .. ------: :: _.__ ... __ ... __ .. ...__-. . --, -'----i-----......... L..+ .. J : ; _)' n;_;,,_,,wnnnnwnm ... .l ... .O=nn=::::;:::-_.__ __ ' ___ ,i... ---Figure 4.11 Conventional versus multiple trellis coded mod with Rate (a) 2/6 (Rate# I) and (b) Rate 2/6 (Rate #2) 37

PAGE 50

---.---------. .c:j;, :: ----.. -.. --.. --------------..:----- :""'_---_._-r_ ----t---'-----------_, ..:--J... L;... .... ---_____ j__,-----..... --uumuummo-=:-;__' ,,_,,,'----Figure 4.12 Conventional versus multiple trellis coded mod with Rate 2/8 (Rate #5) It can be distinguished from the results that as the rate increase the shape of the bit-error rate BER enhances more, but the disadvantage here is that the bandwidth also increases. Therefore, we should know what the requirements of our system are during the design process to make the best selections of the needed rate. 38 II w ID OPSKWMh OIRares Vs Theorelhocai .... ...... ...... -, .... \ ...... ,._ .... .... ...... ,, 'l! \ '\ \ EbNo '\ \ \ \ \.

PAGE 51

ll w m 16iltiii'NC!I Oil Rilles ,if 5 EbNo ll w m S.QAM Wllh Oil Rilles Vs Theo19111Eal ,, " .... \ .,_ .... Figure 4.13 Multiple trellis coded Modulation with different Rates (a) Frequency Shift Keying (FSK) (b) (QPSK) (c) 16QAM (d) 64QAM 39

PAGE 52

5. Communication Systems Applications 5.1.0il Spill Detection by Satellite Image using Sequential Detection ofCbange Test Oil spills on the sea surface might happen without any previous caution and are seen relatively often. Efficient and effective oil spill monitoring and detection accelerates response time, thus minimizing remediation costs and limiting dangerous impact to the environment. An innovative satellite-based oil pollution detecting framework is demonstrated in this section, including satellite communication link configuration, satellite images transmission, enhancement and segmentation. Finally the sequential detection of change algorithm (SDCA) is proposed to detect oil spills on the ocean surface from the enhanced remote sensing data. MODIS images of the Gulf of Mexico accident from NASA between May and June 20 I 0 are adopted and the results of this research show that the proposed algorithms can effectively distinguish the spills covering vast areas of the marine environment even with severe additive noise, and have good separation properties against complex signatures, e.g. the vicinity to the irregular coast or foggy and cloudy weather conditions. 5.1.1 Introduction Large scale oil pollution is one of most universal problems in the ocean. For instance, the oil spill transportation is increasing every year, and it is hard to detect because the amount of oil spill from these see ships is not recognizable. Therefore, the risk of large-scale oil pollution catastrophe increases substantially during the coming years. Various efforts to manage the spill with controlled burning, dispersal and plugging the leak have so far been ineffective. As a result, the long term influences on fisheries, wildlife, human health, and tourism would occur [II] and [I2]. The idea behind this section started actually after April 20th, when Transocean Ltd. reported an explosion and subsequent fire aboard the semi-submersible drilling rig Deepwater Horizon. This has led to the largest oil spill in American history and that because of three measure breaks in the connected pipe. No one knew for cretin how much oil has been released, but there is no disagreement that the spill was massive. Usually, oil spills occurred at the ocean service, at this time it initiated at the ocean floor and rising up to the service of the water. As the oil reaches the service of the water, it begins to spread and moves duo to the winds in random directions. Therefore, it is very important to detect these oil leaks as soon as possible. As with any catastrophic spill it is difficult to predict the extent of the damage. How much 40

PAGE 53

oil has been spilled, and how far will spread. After a series of failed efforts to plug the leak, government and company officials said oil will likely continue flowing until a relief well cut off the gusher. This event has fmished closing the leaking in middle of July, 2010.The incident has resulted in a massive oil spill and has been announced an incident of national significance by the government. The Deepwater Horizon Unified Command has expected that between 1.5 million and 2.5 million gallons of oil were leaking into the Gulf of Mexico every day (22], (23], and [24]. That would mean between 94 million and 185 million gallons of oil had leaked into the Gulf of Mexico. In this section, the catastrophic explosion that caused an oil spill from a BP offshore drilling rig in the Gulf of Mexico is adopted. The deployment real-time oil spill detection algorithm will facilitate and largely simplify the cleanup process. In the effort to locate oil spills, satellite images are utilized. Such images are valuable, since they cover large areas. A sequential detection of change algorithm (SDCA) is introduced for the detection of oil spills in the Gulf of Mexico. Additive White Gaussian Noise (A WGN) is superimposed, however, for the study of the SDCA 's robustness against noise present in satellite channels. As will be explained in detail later in this section, the proposed SDCA algorithm is robust, allowing for effective oil spills detection even within the near-beach noisy environment. The section is organized as follows: In Section 5.1.2, we explain our approach and present the source of the data and the steps involved in our oil spill detection methodology. In Section 5.1.3, we discuss how the image may be transmitted using a 256-QAM modulation technique, while the transmission is through a noisy and fading satellite channel. In Section 5.1.4, we discuss some image processing components, such as image enhancement and segmentation. In Section 5.1.5, the deployed sequential detection of change algorithm is discussed and experimental simulation results are presented. 5.1.2 Data Source and Methodology 5.1.2.1 Study Area The study area is located aboard the Deepwater Horizon, a drilling rig working on a well for the oil company BP one mile below the surface of the Gulf of Mexico. It lies between longitude -89.0000 and +29.0000 latitude. On April 20th, 2010, Transocean Ltd. reported an explosion and subsequent fire on board the semi-submersible drilling rig Deepwater Horizon as shown in Figure 41

PAGE 54

5.1.1. This has led to the largest oil spill in American history. The incident has resulted in a gigantic oil spill and has been declared an incident of national significance by the government. For more information one may look into references [22], [23] and [24]. 5.1.2.2 Data Source NASA's Aqua satellite captured these images of the Gulf of Mexico between May-June 2010 using its Moderate Resolution Imaging Spectroradiometer (MODIS) instrument. A number of images are handpicked each day and posted on [14] as soon as possible after data acquisition. The satellite may also observe real-time sections, to view the latest MODIS imagery available, within minutes after it is automatically processed by the MODIS Rapid Response System [13] and [14]. Figure 5.1.1 Original image. NASA captured this image ofthe GulfofMexico on May 24,2010 at 16:50 UTC MODIS 5.1.2.3 System Operation In this section, we develop a real-time oil spill detection process depicted by the system in Figure 5 .1.2. The detection process follows the steps below. I) Images of the interested region are captured via any of the scenarios depicted in Figure 5 .1.3. 2) Images are transmitted to the ground receiving station via satellite channels. The data will be corrupted by additive White Gaussian Noise and some fading loss. 42

PAGE 55

3) Depending on the scenario used, different modulation techniques are applicable, before transmission. In all cases, Bit Error Rate (BER) will be calculated to measure the performance of the link. Fig 5.1.2 Block diagram of the implemented system 4) The ground receiving station is a Very Small Aperture Terminal (VSA T), receiving data from all the different satellites that may cover the area of interest. 43

PAGE 56

5) At the receiving end, several image processing techniques, such as image enhancement and segmentation, are used in this study, to remove any noise and artifacts. This step is crucial to the objective of correct oil spill detection. 6) The detection of change sequential algorithm is deployed for the real-time detection of oil spills. 7) The results from the deployment of the algorithm in Step 6 separate the oil from the water. In particular, this step will help the analyst to make the final decision. 5.1.3. Satellite Communication Link Configuration 5.1.3.1 Different scenarios: The significant difference between the three scenarios lies in the levels of superimposed to the transmitted images noise and fading. The applicable scenario should be known to the analyst for the subsequent deployment of the appropriate image processing techniques. I \ / -. __. ... I Figure 5.1.3 Block diagram of the whole system using different scenarios 44

PAGE 57

Scenario #1 :Image tracking and monitoring directly from the satellite. Scenario #2:The boat takes multi-images from the interested area around it and transmits them via a Local Area Network (LAN) to a Wide Area Network (WAN) (such as a VSAT). Then, from the boat's VSA T, the images are transmitted to the ground receive station that processes them and detects the possible presence of oil spills. Scenario #3: A helicopter takes multi-images from the interested area and transmits them to the ground receive station for processing and subsequent oil spill detection. r---r L--&ilti' .. ---l---L-r-1 :I I I Figure 5.1.4 Monitoring System operates for many different scenarios. 5.1.3.2Modulation, Constellation, and the affect of A WGN Figure 5.1.5 shows the constellation of a Gray-coded 256Quadratic Amplitude Modulation (QAM) transmitted and received image signal. Notice that the received signal image constellation plot does not appear exactly like the transmitted signal constellation. The reason from this is that the 45

PAGE 58

received signal constellation will generally have a small cluster of points around the 256 exact points, as a result of superimposed transmission channel noise. As shown in Figure 5.1.4, excessive White Gaussian Noise (A WGN) results in 256 clusters of points whose positions are far away from the exact constellation points. Q) :; iii -o "' "' 15 10 5 0 0 -5 -10 -15 Received and Transmitted Signal I Received Signal I * * + Transmitted Signal ......... *'* '----------'"--' ................................................ .............................................. ..................... ..................... ... ................................................ ........................... ....... ... ................ ........................ ... ... .......... ........................ ... ................................................ ................................................ **************** ........................................ ................................................ **************** -15 -10 -5 0 5 10 15 In-Phase 20 15 10 5 "' iii 0 "' "' 0 -5 -10 -15 -20 -20 Received and Transmitted Signal Received Signal I. Transmitted Signal I ::::.;;;;;;;;;:: ........................................ **************** **************** **************** **************** **************** **************** **************** **************** **************** **************** -10 0 In-Phase 10 20 Figure 5.1.5 The effect of A WGN on transmitted and received Gray-Coded 256-QAM Signal Constellation image 5.1.3.3 Measure tbe performance of tbe link using BER In satellite digital transmission, the Bit Error Rate (BER) is an important performance. It is defmed as the ratio of the number of information bits over the number of bits transmitted, over a fixed time period, where error correction bits are added to information bits before transmission for error protection. The BER is used as a metric of satellite channeVmodulation scheme evaluation [ 19],[20],[21 ]. In this section, the BER of a variety of modulation schemes over an A WGN channel is calculated for use in the oil leak detection case. The x-axis is the ratio of energy to noise power spectral density (EbN0), in dB. They-axis represents the BER for the 256-QAM scheme. In Figure 5.1.6, theoretical results for the 256-QAM scheme are compared with simulation results before and after the addition of AWGN. 46

PAGE 59

Performance of M-OAM for vary1ng M 10 r-----,r------r----,------,---,------r----,----, ffi 101 ID -e-S1!11ulat10n 256QAM Simulat1on 256QAM + More AWGN -+-Theoretical 2560AM 10l L_ ___J __ ____L __ __,_ __ __,_ __ _,_ __ __L_ __ _!.___ ____J 0 6 B EbNo(dB) 10 12 14 15 Figure 5.1.6 BER to measure the perfonnance of the channel for 256 QAM (Simulation Vs Theoretical Vs Simulation with extra A WGN) Figures 5.1.7-5.1.10 are subset images from the Gulf of Mexico, where the incident has happened. From these captured images, it can be recognized that the spilled oils are close to beach and that the weather is also somewhat cloudy, deterring the oil spill identification. Figure 5.1.7 and 5.1.8 Original image after transmitted and received through the satellite. "After modulated and demodulated" 47

PAGE 60

All the pictures were purposely selected to include a mixture of clear water, oil silk, beach, and clouds as well as added A WGN, to evaluate the proposed sequential detection algorithm in unfavorable given conditions. We include additional supported images at the end of this section, to exhibit the efficiency of our proposed algorithm on many different satellite images. Figures 5.1.9 and 5.1.10 are the same as Figures 5.1.7 and 5.1.8; the only difference being that there are many dots on the both images. These dots are generated because of the effect of the A WGN. As mentioned earlier, the amount of A WGN depends on the chosen operative scenario. Figure 5.1.9 and 5.1.10 Original image after transmitted and received through satellite and adding some extra A WGN. "After modulated and demodulated" 5.1.4. Image processing Figures 5.1.11 and 5.1.12 show how the clear water and the oil silk are enhanced. Then, it will be shown how the images can be segmented, via the use of a certain threshold. The darker region represents the oil silk, and the brightness region represents the clear water. It is clear from the Matlab color bar that intensity of clear water is less than 150, while the intensity of oil silk is more than 150. The margin of the separated threshold should be chosen within the range of 140-170. To detect a spilled region of oil from a satellite image, everything was coded in Matlab. Both Figures 5.1.11 and 5.1.12 are converted to binary images. It can be distinguished the effect of adding the A WGN on the binary image where Figure 5.1.11 has more white dots than Figure 5.1.12. 48

PAGE 61

Figure 5.1.11 Received image Figure 5.1.12 Received noisy image "More A WGN" 5.1.5 Sequential Detection of Change Algorithm The sequential test for detection of change was first introduced by Page in 1954 for memoryless processes. Lorden proved its asymptotic optimality for such processes. Then, Bansal and Papantoni-Kazakos extended the test for processes with memory and proved the asymptotic optimality of the extension for general processes possessing mixing conditions. The algorithm for Bernoulli memory less processes was used by Papantoni Kazakos for the detection of faulty links in networks. In this section, it is assumed that the monitoring system will be permitted to receive many images in the base station from the satellite and decide to make an alarm if the oil spill exits. Mainly the system take the image and examine it, by going through the whole pixels in the image and decide when to stop observing as well as what change in the underlying data distribution has occurred "Make an alarm showing that oil spill exists". 5.1.5.1.1. Tests for detecting a change in distribution: Let f0(xn) and f1 (xn) denote then-dimensional density function of two well known, distinct, discrete-time, and mutually independent stochastic processes at the vector point xn. Let it be known that the data sequence is intitally generated by the density function f0 Let it then be possible that, instead, at some point in time the density function f1 may become active and remain so form that point on [ 18]. Then, given an infinite data sequence x = {xi; i 1} the possibilities are: 49

PAGE 62

I) The f0 and f1 change never occurs. Thus, the total sequence x is generated by the density function f0 2) The f0 and f1 change occurs before the sequence x starts being observed; Thus, the total sequence x is then generated by the density function f1 3) The fo and fl change occurs just after the datum Xm; m 1. Thus, the subsequence, xr is then generated by the density function f0 and the remaining sequence x;.+1 = {xk; k m + 1}, is then generated by the density function f1 [8]. In our situation, it is assumed f1 represents the oil spill, while f0 represents the noise, "clear water, clouds and winds, and A WGN". Where x = {xi; i 1} represents the observed data. "scanning all the pixels row by row". Given the finite sequence x ={xi; i 1} and the density functions f 0(x")and f 1 (x"), the objective is to detect a possible f 0 and f 1 change are reliably and quickly as possible. To detect the possible occurrence of an f0 and f1 change. Select some positive threshold o, and scan the given image, and decide that the f0 and f1 change has occurred at the first time n such that T(x") o, where T(O) = 0 (5.1) The last equation operates sequentially via the use of two thresholds, 0 and o where 0 represents a reflecting barrier and o represents an absorbing barrier [ 18]. Oil Spills : Decide H 1 ---------8 0 ... n-data selected sequentially (--. ... .. -: ......... -----0 Figure 5.1.13 Making the final decision in the first crossing to the threshold. 50 Noise: Decide H 0 ..1 n-data selected : sequentially -------. :' .-', . ...... --.. ----..... : ___ ......... ____ 0 Figure 5.1.14 Reflect the final decision when it crosses the 0

PAGE 63

There are going be two thresholds in this case, one detects the change from "not oil spill exists" to "oil spill exists" has occurred, and another one to detect the opposite case. When the two stochastic processes represented by the density functions fo and f1 are memoryless, then Let both the non-composite hypotheses H0 and H1 be described by Bernoulli process that generates binary sequences with elements zero and one, and it can be any kind of distribution. Let the element one occur with probability q under H0 and probability p under H1 Let it be desirable to detect a possible change from the Bernoulli process. We then apply the sequential test in (5.1.1) while we assume q > p. Then we can start the derivation to obtain (5.1.2) (5.1.3) Now, select some positive threshold o. Observe data sequentially and decide that the changes has occurred at the first instant n such that T(x0 ) o, where Let us donate T(O) = 0 n p(1 q)' (1 -p) T(x0 ) = logq(1 xi+ n log(1 -q) 1=1 n n-1 P( 1-q) (1-P) T(x )-T(x )+logq(1-p)xn+ log(1-q) log (1-p) (1-q) y(q,p) = p(1-q) log-q(1-p) Then the test will take the following equivalent form. 51 (5.1.4) (5.1.6)

PAGE 64

(5.1.7) After changing the image to binary, we assume p represents "oil spill", and q represents ''the effects that look like oil spill". The Bernoulli distribution has been chosen in this case, and any kind of distribution can be used for such case. The only difference will be the general formula of the algorithm which means the complexity might increase or decrease. To compute both p and q, we need to train our system using many different images. In this way, the range of each p and q can be known and saved in our data base as the following: /)Select many different images. 2) Compute p by counting the number of oil spill pixels divided by the number of background pixels. 3) Then, compute q using this formula => Number of affects that look like oil q = Total number of pixels-Number of oil spill pixels 4) Results: after repeating the same procedure on many images, it has been found that the range of q is approximately (0.2 to 0.4) while pis between (0.045 to 0.08). The affects of using different pairs of p and q has some affect on the results of detecting the oil spill. A threshold on the other hand is also a very important factor; therefore, the choice of the threshold will be determined from specific false alann and power requirements and will depended on where the emphasis of power Vs false alann is placed. The false alann and power curves will be further discussed in section 5.1.2. Figures 5.1.15 and 5.1.16 show the steps of how the oil spill can be extracted. It is obvious that there are a lot of small dot spots because of adding up A WGN in the satellite communication link between the interested region and the ground receiving station. In fact, the whole satellite communication system model was built in this section to show the effect of increasing A WGN levels on the final decision. We thus did not apply processing operations on these dots, to allow the evaluation ofthe robustness ofthe sequential algorithm against the AWGN. 52

PAGE 65

If the process Hn ,(Ho), causes the stopping of the algorithm, it is decided that a q -+ p shift has occurred. If, on the other hand, the Tn ,(H1), process causes the stopping, it is decided that a p-+ q shift has occurred instead. For any given (q, p) pair, every threshold t and h selection is characterized by a pair of statistical curves: the probability of false alarm and power curves. The flowing steps below show how the false alarm and power curves can be computed. Figure 5.1.17 and 5 .1.18 show how the image can be scanned row by row. CoUnns Rows Figure 5.1.17 Method of scanning images Figure 5.1.18 one row to calculate false alarm and power In our case, one row has been used that contains about 600 columns using some certain thresholds t and h to compute the probability of false alarm and power curves. The derivation of the Markov Chain expression in [ 18] and [25] is used in order to generate a matrix. This matrix actually is function of ( q, p, t, s, I, v) for q -> p case, and function of ( q, p, h, s,l, v) for p -> q case and the results are given below for the q -> p case, and it is similar for p -> q with different symbols. Case #I: Detect from q -> p Given some threshold t > 0, the mode not-oil-spill (q) to mode oil-spill (p) change monitoring algorithm is basically characterized by two time curves: the power and false alarm curves, denoted and Ut(n); respectively. Where n denotes the number of samples and t represents the case q -> p where, 54

PAGE 66

pt(n): The probability that the q -> p mode change monitoring algorithm crosses its threshold t before or at time nt. given that the operation mode is p throughout .. Ut(n): The probability that the q -> p mode change monitoring algorithm crosses its threshold t before or at time nt, given that the operation mode is q throughout. Case #2: Detect from p -> q Given some threshold h > 0, the oil-spill (p) mode to not-oil-spill (q) mode change monitoring algorithm is basically characterized by two curves: the power and false alarm curves, denoted respectively ph(n) and ah(n). Where n denotes the number of samples and h represents the case p -> q where, ph(n): The probability that the p -> q mode change monitoring algorithm crosses its threshold t before or at time nh, given that the operation mode is q throughout. ah(n): The probability that the p -> q mode change monitoring algorithm crosses its threshold t before or at time nh, given that the operation mode is p throughout. In Figure 5.l.l9 and 5.1.20 below, we depict the Pt(n), Ut(n), Ph(n) and ah(n)curves, to discuss qualitative behavior. We plot these curves for different threshold values and (q,p) pairs. The (q, p) pairs (0.3, 0.075) has been considered. These two pairs are most representative of the quality shifts that are of practical interest. For the above pairs the false alarm and test power curves have been drawn as functions of the sample size and also using some threshold t and h. The Markov chain model has been used for calculation the pt(n), Ut(n), ph(n) and ah(n) probabilities [25]. The ftrst two curves in the Figure.5.l.l9 are the Pt(n) and Ut(n); "detect from q -> p" where the threshold that has been used t = 25,50,100, and 200; y = 0.1674, and the stopping time for monitoring a change algorithm when it crosses its threshold t before or at time nt equal to 262. On the other hand, the second two curves in the Figure 5.1.20 below are the Pt(n) and Ut(n); "detect from p> q" where the threshold that has been used h = 30,60,90,and 150; y = 0.1674, and the stopping 55

PAGE 67

time for monitoring a change algorithm when it crosses its threshold h before or at time nh equal to 413. Figure 5.1.21 shows the affect of changing the threshold on the oil spill's image. > 1-:::::i ffi <{ til 0 c: (l. Perfomance Charach1ens1Jcs of Sequen!Jal Algonthms '11=200 --a.1 -1!2.12=100 u.2 12=100 """'"ID. 13=50 ""'""a3' 13=50 --J'l4,14=25 --aA. 14=25 10"1 L_ __ _,__ ___ .__ __ ___JL_ __ _. ___ ___,_ ___ ___l,. ___ __,_ ___ _, m ro ffi ro oo NUMBER OF SAMPLES ( USING ONLY ONE ROW FROM THE PICTURE "600 COL") Figure 5 .1.19 Performance Characteristics of Sequential Algorithms using two pairs ( q=0.3, p=0.075) and four different thresholds (t =200, I 00,50,and 25). Detect from q to p. ,_ 1-:::::i 0 c: (l. Perfomance Charach1enstics of Sequenllal Algonthms --"1!1 'h1=150 --"a.1 'h1=150 -u.2; tQ=g) ..... '1!3 ti3=60 """"'a3' h3=60 --tM' h4=lJ m' .._ __ ._ __ __. ___ __. ___ _,_ ___ _._ ___ _._ ___ ....._ __ __, 10 50 ffi 70 Ill NUMBER OF SAMPLES (USING ONLY ONE ROW FROM THE PICTURE "600 COL") Figure 5.1.20 Performance Characteristics of Sequential Algorithms using two pairs (q=0.3, p=0.075) and four different thresholds (t = 150,90,60,and 30). Detect from p to q. 56

PAGE 68

(a) Original Image (c) (e) Binary Noisy Image (b) Gray Image (d) Detecting Oil Leak (I) Detecting Oil Leak Figure 5.1.21 results of choosing wrong thresholds The results from chosen different pairs (q=0.2, p=0.055) is shown in Figure 5.1.22 and 5.1.23. The stopping time in each case is shown where nt equal to 262 and nh equal to 413. Different thresholds have been tested for the pair (q = 0.2, p = 0.055) .The thresholds have been chosen to achieve certain false alarm probability a within different sample sizes n. In particular, in this case, y2 = 0.1143 was used. It is clear from Figure 5.1.22 and 5.1.23 that the a curves are sensitive to the variation of the threshold t and h. we note also that as the value of the decision threshold increases, the false alarm and power curves decrease. Different thresholds have been tested for the pairs (p = .075, q = 0.3) and (p = .055, q = 0.2 ). The thresholds have been chosen to achieve a certain false alarm probability a within different sample sizes n., in this case, y1 = 0.1674 and y2 = 0.1143 were 57

PAGE 69

used. It is clear from figure 5 .1.19 and 5 .1.20 that the a curves are sensitive to the variation of the threshold, while the curves are stable. >.... ::; 0 a: 0.. Perfomanco CharachlerisliCS of Sequenl1al Algonlhms 10" --Jll ,11=200' q=O 3. p=0075 --a.1; 11=200. q=03. p=0075 --p2 12=200. q=0.2 p=O 055 -a.2 12=200 ; q=0.2 p=O 055 10., 5 10 15 20 25 XI 35 40 45 50 NUMBER Of SAMPLES (USING ONLY ONE ROW FROM THE PICTURE "EUJ COL'i Figure 5.1.22 Performance Characteristics of Sequential Algorithms using two pairs (q=0.3, p=0.075) and (q=0.2, p=0.055); Detect from q top. >.... ::; 0 a: a. Perfumance Charachlenstics of Soquenllal Algonthms --"Ill h1=150. p=0075. q=O 3 --"a.l. h1=150. p=0075' q=03 --jl2' h2=150. p=0.055. q=02 --a.2. h2=150' p=O 055. q=02 10' L._ __ __,_ ____ ____ .__ ___ _._ ____ ..J..... ___ --::'':-----=' 10 20 XI 40 50 60 70 NUMBER OF SAMPLES (USING ONLY ONE ROW FROM THE PICTURE "EUJ COL') Figure 5.1.23 Performance Characteristics of Sequential Algorithms using two pairs ( q=0.3, p=0.075) and (q=0.2, p=0.055); Detect from p to q. 58

PAGE 70

5.2 Conclusion In this study, many areas of study are combined together and applied for one chosen application which is oil spills detection. These areas are communication systems, image processing, and detection and estimation area. It was first shown the data source of the images which were taken from the Gulf of Mexico between May and June 2010. Then, they have been sent through a satellite communication channel; then, the effect of the A WGN on various images was applied. After that it was shown how the image signal process contributed significantly to enhance the received noisy image. Different results have illustrated how the image enhancement assisted the analyst to make the final decision. Finally, the implanted robust sequential detecting of change test was derived for Bernoulli case, and the results have shown the efficiency of this algorithm against different environment. The obtained results took approximately two minutes to obtain instead of hours. Therefore, if all the points that have been covered in this section are considered, a gigantic time and money can be saved to decrease the pollution and the harm on our planet. 59

PAGE 71

5.2. Sequential Tests for the Detection of Voice Activity and the Recognition of Cyber Exploits We consider the problem of automated voice activity detection (V AD), in the presence of noise. To attain this objective, we introduce a Sequential Detection of Change Test (SDCT), designed at the independent mixture of Laplacian and Gaussian distributions. We analyse and numerically evaluate the proposed test for various noisy environments. In addition, we address the problem of effectively recognizing the possible presence of cyber exploits in the voice transmission channel. We then introduce another sequential test, designed to detect rapidly and accurately the presence of such exploits, named Cyber Attacks Sequential Detection of Change Test (CA-SDCT). We analyse and numerically evaluate the latter test. Experimental results and comparisons with other proposed methods are also presented. 5.2.1 Introduction Voice Activity Detection (VAD) is deployed extensively, including the Global System for Mobile Communications (GSM), as well as several satellite and radar military and civilian applications, (see in Figure 5.2.1). Thus, VAD is an important component of most systems that incorporate digital voice transmissions. During real time voice transmission, periods of voice activity are followed by silence, where both voice and silence periods are imbedded in background noise. Since voice is generally transmitted through fixed bandwidth links, the transmission of the silence periods induces severe bandwidth waste. Voice Activity Detection (VAD) allows for the compression of the silence periods and may result in up to 30 to 40 percent of bandwidth savings. To detect voice activity versus silence periods, the starting and ending points of continuous speech activity must be detected. Several research efforts have been invested in this area, [28] [30]. In this section, we propose a novel V AD algorithm, named Voice Activity Detection using a Sequential Detection of Change Test (SDCT-VAD). The algorithm is designed at an independent mixture of Laplacian and Gaussian distributions; it is tracking effectively the boundary points between continuous voice activity and silence time periods, where during silence there is only noise, while during speech there is speech plus noise. The noise and noisy speech are modelled by a "Gaussian" versus "Laplacian plus independent Gaussian" distributions. Results are included for the cases where the SDCTV AD is applied to detect voice activity, in both the presence and the absence of noise. The 60

PAGE 72

algorithm is also tested within a real-time scenario, to exhibit its robustness and low complexity properties. ........ \, -_,.. .;; _./ I ,/ civiiWuppiQbons ------Figure 5.2.1 V AD integrated in a telecommunication system Considering the possibility of cyber exploits during voice transmission, we also present a novel cyber attack-sequential detection of change Test (CA-SDCT). The CA-SDCT algorithm is deployed during voice activity periods, as detected by the SDCT-VAD algorithm. The proposed CA SDCA algorithm is designed at the Additive White Gaussian Noise (A WGN) cyber attack model and is fully analysed and numerically evaluated in various environments. This section is organized as follows: In Section 5.2.2, the SDCT-V AD algorithm is presented. In Section 5.2.3, the CA-SDCT algorithm is developed. In Section 5.2.4, experimental results are included. In Section 5.2.5, we draw some conclusions. 5.2.2 Voice Activity Detection Algorithm The general operation flowchart of the VAD algorithm is depicted by Figure 5.2.2. The problem to be solved here is the effective distinction between active and inactive voice periods. However, the variety of both the active voice and the ambient noise make this problem quite complicated in real life. As shown in Figure 5.2.2, first the speech signal is generated and is then corrupted by Additive White Gaussian Noise (A WGN). The A WGN affects the shapes of both the active voice and silence periods. The SDCTV AD is then applied to detect voice activity periods. As 61

PAGE 73

will be further discussed in Section 5.2.3, the SDCT-V AD operates on various Signal-to-Noise Ratios (SNRs). Abooem:e oigoal or Noise: its Pdf j .... the G.ussiao distribulion. Noisy voic:e sianal: use the convoluaion property ID fiDd its Pdf (Gaussia+ l.oiplcian). I No :_ .. -.. -.. . -.. -.. -.. J Figure 5.2.2 The general operation flowchart ofthe V AD algorithm 5.2.2.1 Speecb and Noise Probability Distributions Throughout this section, we consider two distinct probability density functions (pdfs) which represent the voice and noise amplitude distributions of the proposed model. The two distributions for speech and noise are assumed to be "Laplacian" and "Gaussian", respectively, as in [26], where different speech distribution models are shown in Figure 5.2.3. Figure 5.2.4 and 5.2.5 show noiseless and noisy actual speech signals, respectively. 62

PAGE 74

0.9 0.8 0.7 -0.6 -0 . -0.2 GlltU .. ian + Gmm L.abplclan 0.6 Figure 5.2.3 Distributions Voice Signals To decrease algorithmic design complexity, we assume statistical independence between successive voice periods, as well as between signal and noise. We then derive the noisy speech distribution via the convolution of the Laplacian and Gaussian densities. Assuming that the corrupting noise is A WGN, the noisy speech signal is represented below, Y = Xsignal + NAWGN (5.2.1) where Y denotes the noisy speech, X signal stands for the clean speech and NAWGN represents the noise, and where X signal and NAWGN are statistically mutually independent. J '"' :: 1-c.. ..... u.:! Figure 5.2.4 Actual noiseless voice signal "silence+ active voice" 63 Figure 5.2.5 Actual Voice Noisy Voice Signal

PAGE 75

5.2.2.2 The Sequential Test for the Detection of Change The sequential test for the detection of change, in its general form, was introduced and analysed in [16) and [18]. It is assumed that an automated system will be monitoring the signal activity to decide when the voice is active versus not, whose design is based on the detection of change in the data generating stochastic process. The automated system will be implemented via the deployment of the SDCTV AD algorithm which will be tracking voice to silence and silence to voice shifts, where voice and silence are modelled by two distinct stochastic processes. Below, we first present the general model considered in references [ 16) and [ 18). Let f0(xn) and f1 (xn) denote then-dimensional density functions of two well known, distinct, mutually independent discrete-time stochastic processes at the vector point xn = { x11 x2 Xn} Let it be known that the active process is initially generated by the density function f0 [16). For the problem addressed, it is assumed that f1 represents the noisy voice process, while f0 represents the AWGN noisy silence. Given the fmite sequence x = (xi; i 1} and the density functions f0(xn) and f1 (xn), the objective is to detect a possible f0 to f1 change as reliably and as quickly as possible. To detect a possible such change, select some positive threshold 6. Then, observe data points sequentially and decide that the f0 to f1 change has occurred the first time n such that T(xn) 6, where (5.2.2) The above algorithm operates sequentially via the use of two thresholds, 0 and o, where 0 represents a reflecting barrier and o represents an absorbing or decision barrier [ 16). When the two stochastic processes represented by the density functions f0 and f1 are memoryless, then the conditioning in the log likelihood in (5.2.2) drops and the algorithmic operations are memoryless as well. A symmetric algorithm that detects a shift from f1 to f0 instead, can be easily derived. In Figure 5.2.6, the time-evolution of both algorithms is depicted. 64

PAGE 76

Active Voice: Decide HI 0 N-Data sequentially ..,, Observed r ., : I J Threshold Reflector Silence : Decide HO ---------------o-81 I 0 N -Data sequentially_-t ... Observed r. /' :J I Reflector 1. Threshold Figure 5.2.6 Making the final decision in the first crossing to the threshold; detecting a change (a) from f0(x) to f1(x) (b) from f1(x) to f0(x). 5.2.2.3 The SDCT-VAD using Laplacian and Gaussian Distributions Let f1 (x) and f0(x) represent the density functions of a single datum x from noisy voice versus just noise, respectively. Let it be desirable to detect a possible change from f0(x) to f1 (x), where f0 (x) is Gaussian and f1 (x) is Laplacian plus Gaussian. The assumption here is that the observed voice process is stationary, memoryless and Laplacian, while the noise process is independent from the voice process and A WGN. Then, (5.2.3) where fs(x) is the Laplacian distribution that represents the voice speech, and fN(x) is the Gaussian distribution representing the A WGN. a f5(x) = z-exp(-alxl) (5.2.4) fN(x) = (5.2.5) 65

PAGE 77

where a denotes the standard deviation of the Gaussian distribution. We now derive the expression f1 (x): fx a [1 y ] / 1(x) = _,, (zexp(-alx-yD] .dy ft (x) = foo [i exp( -alx-yl)] [ ({2 2)]. dy av2rr a a (a2a2 ) x x ft (x) = 2exp -2 -{ exp( -ax) aa) + exp(ax) (aa)} Alternatively, the distribution /1 (x) can be expressed as follows, a (aza2) ft(x) = 2exp -2 -.{h(x) + h(-x)} (5.2.6) where h(x) = exp( -ax) aa) h( -x) = exp(ax) (aa) We then compute the log likelihood ratio updating step in the sequential test: f1(x) 2exp -2 -.{h(x) + h(-x)} [ a (uZaZ) l log fo(x) = log ;


PAGE 78

p{Zii p2 e g(() = ln-2 -+ 2 + 2 + ln{h(() + h( Where h(() exp(-p() p) The algorithmic step may be subsequently modified as follows [ p{Zii p2J e ln-2-+2 +z-+ln{h(()+h(-f)} where X P a h(() exp( -p() p) 1 1 ll(fx) = _",O:
PAGE 79

a: z (fJ .., .,._ .......... .. -...... 10 5 0 -+-o1"'()1 -02=0.6 c3=1.1 -0'1=1.7 .... .... "0 ...... .Q ....... 02 03 04 05 os 01 oe n9 Alpah Figure 5.2.7 Signal-to-Noise Ratio SNR Let us now select some positive threshold & and defme: We may then modify the algorithm in (5.2.2), as manifested by the distributions derived in this section, via scaling, resulting in the following operation: Observe data sequentially and decide that the change from noise to voice activity has occurred the first time instant n such that T(xn) where T(O) = 0 [ [ p.JZi pz]-1 [(2 ]j T(xn) =max 0, T(xn-1 ) + 1 + ln-2 -+ 2 2 + ln{h(() + h( -()} (5.2.10) We note that the algorithm in (5.2.1 0) detects change from silence to active voice. The algorithm that detects change from active voice to silence, instead, is similarly derived, where its recursively derived algorithmic values are given by the expression in (5.2.12) below and where its decision threshold is generally different than that of the algorithm in (5.2.10). 68

PAGE 80

I fo(x) I [ ;'P(;) I og--= og a u2az f1 (x) 2 exp (-2-) (h(x) + h( -x)} (5.2.11) [ [ p.JZrr pz]-l J] T(x") =max 0, T(x"-1)-1-ln-2 -+ 2 2 + ln{h(O + h( (5.2.12) 5.2.2.4 Power and False Alarm Curves for Tbresbold Values Selections In this section, we present algorithmic performance criteria and their use in the selection of the decision thresholds. We specifically evaluate power and false alarm curves induced by the two algorithms in Section 5.2.2.3 for several given decision thresholds. We then compare such curves for different threshold values, to subsequently decide on the values of the operational algorithmic thresholds. Let us define, The probability that at time n the algorithm has not crossed the threshold, o, and its value lies in + given that the acting pdf is f;. where, the recursive expression below can be derived, fn,l (I;) = J fn-l,i (x). fsn -x). dx (5.2.13) x=O
PAGE 81

updating step shown in equation (5.2.8). This process is explained in the Appendix, where the expressions for the computation of the sequences n ;;::: 1} and {nn; n ;;::: 1} are also derived. Given threshold o, the silence mode to active voice mode change detecting algorithm is basically characterized by two time curves: the power and false alarm curves, denoted respectively and an, respectively, where n denotes time instant and where, The probability that the silence to active voice mode change detecting algorithm crosses its threshold before or at time n, given that the operation mode is active voice mode throughout [32]. nn: The probability that the silence to active voice mode detecting algorithm crosses its threshold before or at time n, given that the operational mode is silence mode throughout [32]. When the algorithm that monitors change from mode silence to mode voice is considered, the threshold o may be selected based on the following principle: At given time n have the powers induced by the parallel algorithms be above a predetermined lower bound, while the false alarm induced by each algorithm remains below a predetermined upper bound. The threshold for the algorithm that monitors ch!lJ1ge from voice to silence, instead, is selected similarly. I
PAGE 82

5.2.8, we note that as the value of the decision threshold increases, the false alann curve decreases, but so does the power curve. The threshold selection for the silence to active voice change monitoring algorithm may be based on a required lower bound for the power and a required upper bound for the false alann, at a given time instant n. A similar criterion may be adopted in the threshold selection for the active voice to silence monitoring algorithm. 5.2.3 An Algorithm for Detecting Cyber Attacks during Speech Activity In this section, we consider the case where the voice transmission channel may be vulnerable to cyber exploits. We then focus on developing an automated system that, in concurrence with voice activity detection, also detects cyber exploit activities. We thus develop a Cyber AttackSequential Detection of Change Test (CA-SDCT), designed to detect cyber attacks during voice activity periods, as the latter are detected by the SDCT-VAD algorithm in Section 5.2.2. The block diagram ofthe overall system is depicted in Figure 5.2.2, Section 5.2.2. As shown in Figure 5.2.2, first the speech signal is generated and is then corrupted by additive white Gaussian noise (A WGN). The SDCT-VAD is then deployed to distinguish between voice activity and silence periods. We fmally wish to detect possible cyber attacks during voice activity periods. As in Section 5.2.2.2, let f0{x") and f1 (x") denote then-dimensional density functions of two well known, distinct, mutually independent discrete-time stochastic processes at the vector point x" = { x1 x2 Xn}. For the problem addressed here, f1 represents the process of cyber exploits superimposed on noisy voice activity, while f0 represents the noisy voice activity process in the absence of cyber attacks. Given the infinite sequence x ={xi; i 1}, let the n-dimensional density functions be denoted f0(x") and f1{x"). The objective is to detect a possible f0 to f1 change as reliably and as quickly as possible, utilizing the observed data sequences. As in Section 5.2.2.2, we first select some positive threshold o0 Subsequently, we observe noisy voice data sequentially, during voice activity periods detected by the SDCTV AD, and decide that the f0 to f1 change has occurred, the first time n such that T{x") o0 where (5.2.14) 71

PAGE 83

A similar algorithm may be devised for the detection of shifts from f1 to f0 instead. The sequential operation of the two algorithms is depicted in Figure 5.2.9. Cyber Exploits Present Decide HI ----------. {)--Oo 0 N -Data sequentially .., 1 Observed r ., I J Threshold Reflector Cyber Exploits Absent : Decide HO s: 0 N-Data Observed r. /". :Jt Reflector / Threshold Figure 5.2.9 Making the final decision in the frrst crossing to the threshold, (a) Cyber Exploits Present, denoted H1 (b) Cyber Exploits Absent, denoted H0 We model the presence of cyber exploits by Additive White Gaussian Noise (A WGN) that is superimposed on the transmission channel A WGN, resulting in relatively excessive cumulative white noise. When the two stochastic processes represented by the density functions f0 and f1 are memoryless, the conditioning in the log likelihood in (5.2.14) drops and the algorithmic operations are memoryless as well. As directly deduced from Section 5.2.2.3, in the present case we have: (5.2.15) (5.2.16) where, hu(x) = exp( -ax) ci> a0 : Standard deviation of the transmission noise, in the absence of cyber exploits. a1 : Standard deviation ofthe cumulative noise when cyber exploits are added to the transmission noise. Then, 72

PAGE 84

Or, (5.2.17) The implementation of the cyber exploits detection algorithm is then as follows: During voice activity periods, as detected by the SDCTV AD algorithm, observe data sequentially and decide that the change from absence to presence of cyber exploits has occurred, the first time instant n such that T(x") 60 where (5.2.18) We note that the algorithm in (5.2.18) detects a f0 to f1 change. The algorithm that detects a f1 to f0 change, instead (from presence to absence of cyber exploits), is similarly derived, where its recursively derived algorithmic values are given by the expression in (5.2.19) below and where its decision threshold, 61 is generally different than that of the algorithm in (5.2.18), as shown in Figure 5.2.9. (5.2.19) 5.2.4 Experimental Results 5.2.4.1 Testing tbe SDCT-VAD In this section, we state the steps involved in the numerical evaluation of the SDCTV AD algorithm. First, we select the pertinent involved parameters and deploy the resulting SDCTV AD algorithm, to detect any voice activity in the communication link. Then, the SDCTV AD is evaluated 73

PAGE 85

in various noisy environments. In our simplified model, the silence plus noise mode of operation is assumed to be represented by a Gaussian distribution, while the noisy voice signal is represented by a mixture of Laplacian and Gaussian distributions, as shown in Figure 5.2.1 0. The pertinent parameters to be chosen in the SDCTV AD design are the Laplacian parameter, the standard deviation of the Gaussian noise and the two algorithmic thresholds: A threshold 50 used by the algorithm in (5.2.10); for the detection of change from noise to noisy voice activity, and a threshold 51 used by the algorithm in (12); for the detection of change from noisy active voice to just noise. We used the power and false alarm curves discussed in Section 5.2.2.4, to decide on the values of these two thresholds. In particular, we selected the {50 51 a, a) values (0.3,0.05,0.98,0.0523). We used the design parameter values stated above and tested the robusmess of the resulting SDCTV AD algorithm in the presence of various noisy environments. Various noises were mixed with the clean speech signals. Six different noises were used in our evaluations, including white noise, wind, computer fan, babble, flowing traffic and train passing, as shown in Figure 5.2.11, with different SNRs (25,20, 15, I 0 and 5), as shown in Figure 5.2.12 and 5.2.13. Original Signal Time(Sec) Figure 5.2.10 Original Voice Signal 74

PAGE 86

No1se Babbl8 0 JJ 4() ffi Ell tJEIJ T1me (sec) Ttme (sec) Oog111l Sognl SNR=JJ Ttme (sec) Onqtnal Stgnal SNR=10 Ttme (sec) Ttme (sec) Ougtnil Stgnil SNR=15 Ttme (sec) Ongmal Stgr1al -1 0 Figure 5.2.11 Various noisy environments Figure 5.2.12 Original signal corrupted by AWGN (a) Original Signal (g) Noisy Signal "AWGN" for SNR= 5 dB E-0 t I : -1 -2 0 20 30 40 50 0 10 20 30 40 50 (b) Noise: Wind (h) Noisy Signal "\lllind" for SNR= 5 dB -.. -0 5 -1 0 10 20 30 40 50 0 10 20 30 40 50 (c) Noise: Babble (1) Noisy Signal "Babble" for SNR= 5 dB : Ill: I t -5 -5 0 10 20 30 40 50 0 10 20 30 40 50 (d) Noise Computer Fan (J) Noisy Signal "Compuler Fan" for SNR= 5 dB 1: L L (1 :I l t -5 -10 0 10 20 30 40 50 0 10 20 30 40 50 (e) Noise: Flowing Traffic (k) Noosy Sognal "Flowing Traffic" for SNR= 5 dB I 5 0 -5 -5 0 10 20 30 40 50 0 10 20 30 40 50 (f) Noise: Train Passing (I) Noosy Signal "Train Passing"for SNR= 5 dB E: t 1111!I <( -2 -2 0 10 20 30 40 50 0 10 20 30 40 50 Time(Sec) Time(Sec) Figure 5.2.13 Results of adding noise to the original speech signal (5d8 SNR). (a) Clean speech. (b) Wind Noise. (c) Babble Noise. (d) Computer Fan Noise. (e) Flowing Traffic. (f) Train Passing Noise. (g) Noisy Signal "Noise: A WGN". (h) Noisy Signal "Noise: Wind". (j) Noisy Signal "Noise: Computer Fan". (k) Noisy Signal "Noise: Flowing Traffic". (I) Noisy Signal "Noise: Train Passing Cut". 75

PAGE 87

For comparison, the same voice and noise environments are also tested by the approach presented in [29] and the G.729 VAD algorithm in [31 ]. The results are summarized in Tables 5.2.1 and 5.2.3. The noise data are obtained from http://www.freesound.org/index.php and are added to the clean speech signal at SNRs varying from 5d8 to 25d8. To empirically evaluate the SDCT-VAD algorithm, many audio messages were used, with different lengths, (3 sec and 50 sec), with both male and female speakers and with different SNRs, (5 dB -25d8). The effect of these SNRs on the audio messages is exhibited in Figure 5.2.12, where Figure 5.2.10 exhibits the original speech signal. Figures 5.2.14 and 5.2.15 below show results for the SDCTV AD algorithm with operating parameters as those stated in this Section, when the SNR is 25d8 and 5d8, respectively. The accuracy of the results depends on the level of the SNRs and the type of the noise environment, as shown in Tables 5.2.1 and 5.2.2. Sequentral Delectmg a Change Test 0 3 025 02 f 015 0 I 005 y 0 Figure 5.2.14 SDCT-VAD results "Original Signal corrupted by A WGN (SNR=25d8)" ,.,.. y 0 Sequentral Deleclmg a Change Test ',,.., Trme(Sec) yo Figure 5.2.15 SDCT-VAD results "Original Signal corrupted by A WGN (SNR=5d8)" The efficiency of the SDCTV AD algorithm was evaluated for various noisy voice signals. In the first experiment, we tested the efficiency of the proposed method using the same audio recording discussed above, tracing the speed and the accuracy of the algorithm in detecting the silence mode to active mode change and vice verse. To comparatively evaluate the performance of the proposed SDCT-VAD algorithm, we compared its induced results with those of the manual segmentation. Figure 5.2.1 0 exhibits the "hand marked" results of manual segmentation. Figures 5.2.14 and 5.2.15 exhibit the automated 76

PAGE 88

segmentation induced by the SDCT-VAD proposed algorithm. We evaluated the algorithmic probability of error (Pe), using the formula below: N = [ IAVSP(manual)-AVSP(RSDCVAD)I2 ] L AVSP(manual) n=l n AV: Active Voice; AVSP: Active Voice Starting Point; AVEP: Active Voice Ending Point; N: Number of Active Voice regions. + [ IAVEP(manual) AVEP(RSDCVAD)I2] AVEP(manual) (5.2.20) n The performance of the SDCAV AD is evaluated in terms of probability of false and correct decisions, where Pc is the probability of correct speech classification and where Pe is the probability of false speech classification, computed as in (5.2.20). To compute Pe, we start with known voice activity and the starting and ending points of voice activity marked. We then superimpose AWGN voice contamination with various SNRs (25,20,15,10 and 5dB) and deploy the corresponding SDCT-VAD for various noisy environments, as shown in Figures 5.2. 10, 5.2.11 and 5.2.12. Results comparing the SDCT-V AD with the manual approach are shown in Table 5.2.1. Table 5.2.1 Comparing the Starting and Ending detection time instances ofthe Noisy Active Voice Messages Using the Manual and Proposed method. Message 1 Message3 Message3 SNR (dB ) AVSP AVEP AVSP AVEP AVSP AVEP Manual 0 0.5294 0.8904 1.472 1.769 2.19 2.75 25 0 5295 0.8711 1.472 1.769 2 .19 2 .75 ::t' 20 0.5298 0.8710 1.472 1.769 2 19 2 .75 riJ > t:::' c ('l -15 0 5356 0 870 1.472 1.769 2 .19 2.75 <:> > a I <: = 10 0.5414 0.8695 1.472 1.769 2 .19 2 .75 ;:; 5 0.5366 0 8654 1.472 1.769 2.184 2.75 77

PAGE 89

After finding the starting and ending points of voice activity, as shown in Table 5.2.1, we used the formula Pe(Av) in (5.2.20) to evaluate the accuracy of the algorithm. The results are summarized in Table 5.2.2. From the experimental results, it is clear that the proposed SDCT-VAD algorithm outperforms the manual method. Table 5.2.2 Pc's and Pe's OF the Proposed RSDCA-V AD for Various Environmental Conditions. SNR SNR SNR SNR SNR Noise I SNR(dB) 25 20 15 10 5 White Pc(%) 99.958 99.957 99 946 99.92 99.91 Pe(%) 0 0418 0 0423 0 0540 0 076 0 .081 Pc(%) 99.976 99.972 99. 945 99.89 99.88 Wind Pe(%) 0 0238 0 0279 0 0545 0 107 0.117 Computer Pc(%) 99 976 99.979 99.977 99.97 99.97 Fan Pe(%) 0.0234 0 0210 0.0227 0.025 0 026 Pc(%) 99.979 99.971 99.969 99 96 99.95 Babble Pe(%) 0 0208 0 0285 0.0303 0 030 0 037 Flowing Pc(%) 99.996 99.982 99 953 99.94 99.83 Traffic Pe(%) 0 0037 0.0178 0 0468 0.058 0.162 Train Pc(%) 99 979 99 977 99.975 99.96 99 95 Passing Pe( % ) 0 020 0.0224 0.0243 0.038 0.048 Pc(%) 99.977 99.973 99.96 99 94 99 .91 Average Pe(%) 0.0225 0 0266 0.0387 0 0556 O.D78 In Figures 5.2.16 and 5.2.17 we plot thePc'sandPe's results included in Table 5.2.2. To further validate the effectiveness of the proposed SDCTV AD, we compared its probabilities of correct speech activity detection with those of other approaches. Table 5.2.3 shows a comparison between the SDCT-VAD and two different VAD approaches: the G.729 in [31] and the proposed method in [29]. 78

PAGE 90

The Propsed RSDCA-VAD For Various Environmental Conditions 0.18 .-------.-------.--------.-----------, \ \ \ ____,..__ AWGN --o--Wind \. .......... ----GComputer Fan Babble -+-Flowing Traffic \':"''--q \<"-: --+-FontSize ' \ .......... ______ L_ ______ L_ ______ L_ _____ __y 10 15 20 25 SNR (dB) Figure 5.2.16 Pe's ofthe proposed RSDCA-VAD various environmental conditions. The Propsed RSDCA-VAD For Various Environmental Conditions 100 I ./ .;:. ; .. 0 :./ .... / """/" .,/ .......... . 10 15 SNR (dB) --o--Wind --o-Computer Fan Babble -+-Flowing Traffic --+-FontSize 20 25 Figure 5.2.17 Pc's ofthe proposed RSDCA-VAD various environmental conditions. 79

PAGE 91

The clean speech signal that has length of 50 sec, 60.05% speech and 39.95% silence, was used to evaluate the proposed algorithm against various environmental conditions and to compare it with the ITU standard G. 729Annex B (31] and the proposed in (29] approaches. From Table 5.2.3, it can be recognized that even with environmental challenging conditions, the proposed SDCT-VAD outperformed the G729B V AD and the method in (29]. Table 5.2.3 Pc's OfThe Proposed RSDCT-VAD, and Different VAD Approaches for Various Environmental Conditions. G .729VAD Proposed Proposed Environment [31] and [29] Method in [29] RSDCA-VAD Noise SNR Pc(%) Pc(%) Pc(%) 5 87.46 75.62 97.563 White 15 97 .83 93.42 99 136 25 99 69 99 07 99 328 5 97 29 97 .83 98 .131 Vehicle 15 99.77 99 46 99 .711 25 100.00 100 00 99 .781 5 92 96 86.38 97.201 Babbl e 15 98.45 94.89 99.504 25 99 77 99.38 99 664 4.2 Testing the CA-SOCT As stated in Section 5.2.3, to detect shifts from absence to presence of cyber exploits and vice versa using the CA-SDCT algorithm, two algorithmic decision thresholds are needed: A threshold cS0 used by the algorithm in (5.2.18); for the detection of change from Cyber exploits absent to Cyber exploits present, and a threshold cS1 used by the algorithm in (5.2.19); for the detection of change from Cyber exploits present to Cyber exploits absent. It is assumed that the cyber attack is represented by AWGN. Using the power and false alarm curves, as with the SDCT-VAD algorithm, we selected 80

PAGE 92

thresholds. In particular, we selected the a, 0"1 ) design values (0.07,0.01,0.98,0.0773) and tested the 0"2 values (0.0798,0.0849, 0.1 034), to evaluate the robustness of the resulting CA-SDCT algorithm. (a) (b) (c) (d) Figure 5.2.18 (a) Original Signal (b) Noisy signal with SNR=I5dB (c) Noisy signal with SNR=10dB (d) Noisy signal with SNR=5dB From Figure 5.2.19, parts (a), (b) and (c), we may observe the evolution of the deployed CA SDCT algorithm for different SNR values, where the latter values reflect the cumulative effect of normal channel noise and cyber noise. Each time the threshold is crossed, an alarm is activated. 81

PAGE 93

0 Tlnshold Tesi 0 I Dill O(li 0().1 (a) 0 Cll -01 02 0 .., i 01 03 04 T1me(Sec) (c) 01 05 Threshold Sequential Test 02 03 04 05 06 T1me(Sec) (b) 06 Figure 5.2.19 Cyber Detection during speech activity detected periods using the CA-SDCT (a) Sequential Test: SNR=l5. (b) Sequential Test: SNR=IO. (c) Sequential Test: SNR=5. Figure 5.2.20 shows alann activation scenarios regarding cyber attacks, where in (a) no alann is activated, where in (b) one alann is activated and where in (c) several alanns are activated. 82

PAGE 94

1 2 .1---<> SNR=15 1 2 1 ----<> SNR= 10 06 i 06 o 02 U U..! U .j U.4 U U til (b) 0 e ii 0 0 0 2 1 B 6 4 2 u u OB 04 02 U U. U .:l UA U'.:l U.tl TUne( sec) (b) Cyber .AJ:tack alarms SNR=SdB SNR=5 L U4 u (c) Figure 5.2.20 Cyber Detection during speech activity detected periods using RSDA-CA (a) Alanns SNR=I5. (b) Alanns SNR=IO. (c) Alanns SNR=5. S.l.S Conclusions A novel voice activity detection (V AD) approach was presented. The approach uses the Sequential Detection of Change Algorithm (SDCTV AD), designed at the Laplacian-Gaussian distributions additive mixture. We analysed and evaluated the robust sequential algorithm in the presence of Additive White Gaussian Noise. Several different speech messages were chosen for the effectiveness evaluation of the SDCTV AD, regarding its accurate detection of changes from voice 83

PAGE 95

activity to silence and vice versa. The experimental results have shown that the algorithm is effective in various noisy environments and outperforms other existing voice activity detection methods. A novel cyber attacksequential detection of change algorithm (CA-SDCA) was also presented, deployed to detect cyber attacks during speech activity periods. The proposed algorithm is preceded by the Voice Activity Detection algorithm Sequential Detection of Change Test (SDCT VAD). We considered the case where voice messages are transmitted through the communications system, while a cyber attack may occur at any point in time. The proposed algorithm was analysed and evaluated. The latter algorithm detects cyber attacks effectively; during speech activity periods detected-by the SDCT-VAD algorithm. We modelled the cyber attacks by Additive White Gaussian Noise. The experimental results have shown how the algorithm can be implemented with effective detection results, in a variety of different environments. 84

PAGE 96

5.3 Detect Cyber Attacks Using BERThreshold Tool The risk of cyber attacks has grown tremendously in the recent years. Many of these cyber attacks can be linked to current and historic events in the social, political, economic, and cultural dimensions in the human world. The goal of this section is to detect cyber attacks using a Bit Error Rate Threshold (BERT) tool. The system administrator may get an early warning, the, regarding system hacking, allowing time for possible recovers and future protection. A study case related to the previous section is presented in this chapter. It is a combination of the Robust Sequential Detection of Change-Voice Activity Detection Algorithm (RSDC-VAD) and detection of Cyber Attacks using Bit Error Rate Threshold (CA-BERT). The model is analyzed and some experimental results show how the proposed model may conserve bandwidth and energy, while increasing the level of security against cyber attacks. 5.3.1 Introduction Cyber attacks impose a major threat to our communication networked systems. The motivations vary, while, the victims of cyber attacks may be commercial organizations, governments or military infrastructures. The attack agent may be a hacker or a terrorist. The opportunity of these attacks to hack the system increases as the weaknesses of the system decreases. To detect cyber attacks, we monitor the traffic and the performance of the network. After detecting that the network has been hacked, the system administrator starts investigating and troubleshooting. In this section, we only monitor the traffic performance of the network, to detect if any change in the communication link has occurred. The proposed method for cyber attack detection is called "the Cyber Attacks using Bit-Error Rate -threshold" (CA-BERT) tool, which may be low cost, rapid and efficient, for different cyber attack scenarios. The proposed simulation methodology is designed to test information fusion systems for cyber security that are under development or already built. The HER Threshold method may be used for conducting experiments to analyze computer network vulnerabilities and evaluate efficiency and effectiveness of security policies. Systems administrators normally deal with countless daily warning that is hard to track each one of them. Alternatively, an 85

PAGE 97

automated system may transmit warning signals, when unusual anomalies occur, as detected by CA BERT traffic monitor. The complete model is shown in figure 5.3.1. Cyber '. 1------I --, : 1 CAForm#J 1 CAFormll2 ""' CAFormlll 1 = = 11 1 _: --r -1 : 1 -f-1 1 AID Converter i 1 I Random Noise 1 I I. -----. I 1 I -. _f --.. -I I (Distn"bub"oo Not I Noise(AWGN) I VOiceSigoal#(l) I r---------------- I I I I "Jnterfen:nce" : 1 __ __ 1 1 --.--: ---,f --1 Mdulldor Danodulidor , : 1 No -l--I No ..;_ If ,' If ,'----. I Thn:sbold I , ., 1 ' Yes ._., + r c5 } -;Yes -------: o Dcmodulator I 0 I : 1 --T-: 0/A Cmverter I -;-:t.:: I I r BER:--I r Akk-I I ---" r --i I_RS ___ oc_}-_v __ AD ____ r---: O: Detected 0/ACO!Iverter .u2 &Noise 1 1 __ t 101.. -r-____ 11Sav;.:w r---Sepande StanPr I I Founer the 1 Fomier 1 occssmg 1 Voice Signal #(2) '41 RSDCV AD ,.;. T r. 1o4, Transti to recover the 1 _ _ _ i 1 nos orm Messages 01111 1 011.ginal 51-goal 1 _:] _-. I Only for Voice I I lllld noise I 1 ---. I .. _ _ 1 I SaveBWmd I I_ I using filters I I : Figure 5.3.1 Model: Detect Cyber Attack using HER-Threshold The present work motivated by the need for testing situational awareness tools, developed to detect and analyze attacks on computer networks. The simulation approach requires knowledge of the network operation, which must be captured by the simulation model. However, as discussed briefly in the introduction, the level of detail included in the model will depend on the objective of the simulation. In this case, the packet level information and computer network traffic details are not 86

PAGE 98

needed, so the simulation may be constructed at a higher level, to produce alerts caused by cyber attacks on information carrying network traffic. Let us assume that an organization uses the model shown in figure 5.3.1, utilizing an encoded communications link to relay secure messages among its members. Our objective here is to keep the encoded link secure and use it to detect possible cyber attacks. It is then implied that the transmitted messages through the secure link are confidential and it can be recovered regardless of possible cyber attacks. Let the transmitted voice message to one of the members be "/ am completely operational and all my circuits are functioning petfectly ", where the latter is the signal that we wish to recover. Let the signal generated by the attackers be "I cannot take any of this seriously unless I know who I am talking to", which is the signal that causing interference. 5.3.2 Modeling Cyber Attacks The term "Cyber Attacks = hacker" is refers to anyone who is authentically interested in pushing the limits of software and hardware and is eager to do so in a fully supportive, low cost and sharing network environment. The progress that a hacker can make in an attack depends upon the hacker's capabilities and the vulnerabilities of the network. The methods for modeling and simulating the initiation and progression of cyber attacks through a computer network included in this model are based on A WGN, random noise or interference. 5.3.3 Bit-Error RateThreshold (BERT) Over the last two decades, communication networks have evolved from all analog telephone and radio transmissions to the modem digital communications systems. These digital networks are required to carry large amounts of data at high rates. The need for increased bandwidth has caused an evolution in the media from wires and air. Service providers must guarantee data integrity to their clients, where the bit-error rate measurements are used to measure error resistance performance. The pertinent criterion is called bit-error rate (BER) and its definition is given below, 87

PAGE 99

Number of changed bits BER = -=---=----=---:-:--=--Total Number of bits (5.3.1) Thus, bit-error rate monitoring plays an integral role in detecting changed error performance and quality degradation of communication links. An important aspect of the bit-error rate monitoring that we have designed is that it may operate at a variety of modulation techniques and rates. The BER performance is affected by many factors such as power, noise, or modulation method. If we assume that our system has a specific power, known noise and given modulation technique, the BER performance can be predicated, thus, its change from predicted value may point to the presence of cyber attacks. The BER measurement is not so complicated; it is affmed by sending a stream of bits through our communication systems and compares the output to the inputs using the formula in (5.3.1), where the input stream is known to the receiver via its transmission through a secure link. Our proposed monitoring method is named Bit-Error Rate -Threshold (BERn; we will present and analyze it, in detail. The complete model is shown in figure 5.3.1. To start applying the BERT, let us make some assumptions: 1-The system is already installed, while we wish to increase the level of the security by detecting effectively cyber attacks. 2Generate a well known message at the receiver with a specific length. 3The cyber attack is modeled as Additive White Gaussian Noise (A WGN), as well as additional distortion and interferences. 4The optimum BER system performance is assumed well known. The BERT tool starts operating as an investigator. If, for example the BER at some EbNo is normally equal to BER = 0.000 I, then we assume that a minimum and a maximum threshold define derivation tolerance; that is O.OOI>BER>O.OOOOI, represents normal performance, where an alarm is initiated if either threshold is crossed. To demonstrate this method in more detail, let us assume that the communication link uses some modulation method such as, QPSK shown in figure 5.3.2, and normal optimal BER performance is known (Assumption #4) and is represented by the "Blue curve". Then two thresholds are created by "black curves". 88

PAGE 100

This operation presents a simulation model for the intrusion detection system, depicted by simulated alerts (Red, Yellow or Green), where these colors level represent the cyber attacks, and the level should be assigned by the user's specifications. The user's specifications represent design requirements for the system, operation, such as bandwidth, speed, type of information (voice or data), security level, distance between each Local or Wide Area Network (LAN or WAN), quality of the communication link, type of modulation techniques, etc. There are 4 subplots in figure 5.3.2, where each one represents a specific behavior for the communications link; i.e.; 1) Plot (a) demonstrates that no attack has been detected because the evaluated BER is still between the two thresholds; so, the alarm in this case will be green. 2) Plot (b) and (c) illustrate that there is something wrong in the communication link, but it is not critical, since the evaluated BER is still between the two thresholds; so, the alarm in this case is yellow. 4) Plot (d) demonstrates that an attack has been detected, since the evaluated BER is not between the two thresholds; so, the alarm in this case is red. The same previous steps can be repeated again if we use another modulation technique such as 64QAM. Figure.5.3.3 shows the some results .. 10' -Oplmum CUM! M'" Th
PAGE 101

10' CUM' Mm ThriPs hold Ma. Threshold .Alter Cyber Anack ,_ EbNo 10' -Ogtln'U'YICUM -W. Threshold -AierCyber.A11ack EbNo Figure .5.3.2 QPSK Modulation Vs Cyber Attack using HER-Threshold (a) No cyber attack detected (Green alarm) (b) and (c) minor warning (Yellow Alarm) (d) Cyber attack detected (Red Alarm) The proposed technique is not a good choice if the bandwidth is limited since a well know message is sent periodically to measure the BER, utilizing excessive bandwidth. On other hand, the BERT is an efficient tool, where excessive bandwidth is available simultaneously increasing securely. The cyber attacks have been modeled as a combination of A WGN "generated from the circuits such as semiconductor junction noise" and some interference "generated from the fading multipath transmission channel". The objective is to "first detect, and then possibly recover using some troubleshooting process" EbNo 90 EbNo

PAGE 102

10. ,__---=-------:':---:':-----!::------'"':--------:' 0 10 15 !) 25 ]] EbNo Figure 5.3.3 64QAM Modulation Vs Cyber Attack using BER-Threshold (a) No cyber attack detected (Green alarm) (b) and (c) minor warning (Yellow Alarm) (d) Cyber attack detected (Red Alarm) 5.3.4 Fourier and Inverse Fourier Transform The frequency domain usually offers attractive advantages for signal processing. It induces filtering operations that are faster, and it may collects and separate signals from each other or allow measurements that would be very complicated to analyze in spatial domain. The Fourier transform is needed here to determine bandwidth and frequency location of each transmission channel. First, we briefly review the continuous Fourier transform from the purely mathematical point of view. Let x(t) be "the noisy speech signal" represented by a continuous time function, where t is time. The Fourier transform of X( ro) and its inverse x(t) are defined by the equations below X(ro) = L: x(t) e-jw'dt x(t) = L: X(ro) ei011dt (5.3.2) In our case, we have a speech signal that has been transmitted and conupted by cyber attacks (combination of A WGN and some other voice signal "Inference"). The objective is to eliminate cyber 91

PAGE 103

attack affects, an objective which may be hard to obtain in the time domain. Conversion to frequency domain may facilitate the objective. See figure 5.3.4 for the demonstration. Voose Sognal 1 & 2 1.5 05 "' "0 0 Q_ E <( -0.5 -1 -1.5 -2 0 2 4 6 8 10 12 14 16 18 20 Time Figure 5.3.4 The received noisy speech signal Figure 5.3.5 Two-Dimensional (2-D) Time-vs-Frequency Spectrogram For each observed message channel, the lowest and highest frequency extent associated with each channel is shown in table 5.3.1, as well as its bandwidth. 92

PAGE 104

Channel #I Channel #2 5.3.5 Conclusion Table.5.3.1 Frequency Bands Lowest Frequency 2KHz lOKHz Highest Frequency lOKHz 15KHz Bandwidth 8KHz 5KHz In this section, we considered two approaches towards the detection of cyber attacks. In the first approach, the robust sequential detection of changevoice activity detection (RSDCV AD) monitoring combined cyber attack and voice activity detection in one model. In the second approach, the Bit-Error rate measurement was used to detect rate changes that may point to cyber attacks. In the latter case, the bandwidth clearly increases duo to the periodically injected messages used for BER detection. The RSDCV AD was used to conserve about 30-40 percent of bandwidth, and it was explained in detail in section 5.2 of this chapter. 93

PAGE 105

6. Conclusions In this thesis, we focused on the performance metrics of accuracy, bandwidth/energy efficiency, security and cost effectiveness, for a variety of communication and signal processing applications that are socially, politically and economically current. In the first four chapters, some background on communication and signal processing such as modulation and coding was presented. In chapter five, we proposed solutions to some current communication system applications. We firSt selected the problem of oil-spill detection, focusing on accuracy for the minimization of remediation costs and the limitation of dangerous impacts on the environment. The data source studied in this case were the images taken from the Gulf of Mexico between May and June 20 I 0. These images were then considered as transmitted via a satellite communication channel, whose effect was modeled as A WGN. Different results have illustrated how image enhancement may assist the analyst in his final decision. Finally, the implanted robust sequential detection of change test, as induced by the Bernoulli model, was deployed for the detection of oil spills; the produced results have shown the efficiency of the latter test in the presence of various different environments. These results were obtained in approximately two minutes, rather than hours, with subsequent significant savings in costs and environmental negative effects. To address bandwidth/ energy efficiency, we selected a voice activity detection problem. A novel voice activity detection (V AD) approach was presented. The approach uses the Sequential Detection of Change Algorithm (SDCT-VAD), designed at the Laplacian-Gaussian distributions additive mixture. We analysed and evaluated the robust sequential algorithm in the presence of Additive White Gaussian Noise. Several different speech messages were chosen for the effectiveness evaluation of the SDCTV AD, regarding its accurate detection of changes from voice activity to silence and vice versa. The experimental results have shown that the algorithm is effective in various noisy environments and outperforms other existing voice activity detection methods. To address security, we selected the problem of timely and accurately detecting the presence of cyber exploits in the communication transmission, where a new detection method is introduced and analyzed. We considered two approaches towards the detection of cyber exploits. In the first approach, we applied cyber exploit detection on detected by the RSDCT-VAD voice activity periods. In the second approach, the Bit-Error Rate (BER) measurement was used to detect rate changes that may point to cyber exploits. In the latter case, the bandwidth clearly increases due to the periodically 94

PAGE 106

injected messages needed for the BER detection. The first approach that involves detecting cyber exploits during speech activity periods was named the Cyber AttackRobust Sequential Detection of Change Algorithm (CA-RSDCA). The proposed algorithm is preceded by the Voice Activity Detection algorithm Sequential Detection of Change Test (SDCT-VAD). We considered the case where voice messages are transmitted through the communications system, while a cyber attack may occur at any point in time. The proposed algorithm was analyzed and evaluated. The latter algorithm detects cyber attacks effectively; during speech activity periods detected-by the SDCT-VAD algorithm. We modeled the cyber attacks by Additive White Gaussian Noise. The experimental results have shown how the algorithm can be implemented with effective detection results, in a variety of different environments. 95

PAGE 107

APPENDIX A From expression (5.2.8), in Section 5.2.2.3, we have: [ p..fi.rr p2] (2 g(()= ln-2-+-z +2+ln{h(0+h(-0} In Figure A below, we plot as a function Let g-1(y) is such that, = y. Then, 16 ., f6 .175 B5 95 '4 . 1 1 . .(16 .(II. .{)2 0 02 o.. 06 w .:j I l . 1 j 09 1 Figure. A Evaluate the whole updating step algorithm 96

PAGE 108

From Figure A.l, it can be concluded that [ 132] In 2 -+ 2 + 2 + h( and then, fn.tCO = J fn-t,i(x). [Ji((x) + Ji( -((x))]. x=O
PAGE 109

fn.1Cf) = J fn-uCx)[h({x) + h(x{)]2 C1 aexp -2 -.dx [ (u2a2) l The following expressions are used for the computation of the false alarm curves: fn.oC{) = J fn-l,o(x). [h({x) + h(x{)] OIa Pn = Jo fn,l (S). ds {0/CJ Cln = Jo fn,oCO. ds 98

PAGE 110

BIBLIOGRAPHY [I] G. Ungerboeck, "Trellis-coded modulation with redundant signal sets part 1: introduction," IEEE Communications Magazine, vol. 25-2, pp. 5-11, 1987. [2] Dennis Roddy "Satellite Communications", Third Edition Publisher: McGraw-Hill Pub Date: 2001. [3] Tri T.Ha, "Digital Satellite Communications", Second Edition Publisher: McGraw-Hill Pub Date: 1990. [4] International Telecommunications Union, "Handbook on Satellite Communications", Third Edition Publisher: Wiley Pub Date: 1995. [5] Louis J. Ippolito, Jr., "Satellite Communications Systems Engineering Atmospheric Effects, Satellite Link Design and System Performance", Publisher: John Wiley & Sons Ltd Pub Date: 2008. [6] Stojce Dimov Ilcev, "Global Mobile Satellite Communications: For Maritime, Land and Aeronautical Applications", First Edition Publisher: Springer Pub Date: 2005. [7] Claude E. Shannon, "A Mathematical Theory ofCommunication" is an influential'' 1948. [8] Bernard Sklar ,"Digital Communications: Fundamentals and Applications", Second Edition Publisher: Prentice Hall Pub Date: 2001. [9] Bernard Sklar,"Digital Communications", Second Edition Publisher: Prentice Hill Science/Engineering/Math: 2000. [10] A. J. Viterbi and J.K. Omura. "Principles of Digital Communication and Coding'. McGraw-Hill, 1979. [II] The New York Times, GulfofMexico Oil Spill (news), July, 2010. [12] CNN, Gulf of Mexico Oil Spill (news), July, 2010. [ 13] NASA Database Gulf of Mexico Oil Spill, http://www .nasa.gov/topics/earth/featuresloilspill /oil_spill_gallery.html. [14] MODIS Rapid Response System, Gulf of Mexico Oil Spill, http://rapidfire.sci.gsfc.nasa.gov/ gallery /?search=oi I. [15] A. Wald, Sequential Analysis. New York: Wiley, 1947. [16] P. Papantoni-Kazakos, "Algorithms for-monitoring changes in qualityof communication links," IEEE Trans. Commun., vol. COM-27,vv. 682-692. 99

PAGE 111

[ 17] Rakesh K. Bansal and P. Papantoni-Kazakos, "An Algorithm for Detecting a Change in aStochastic Process," IEEE TRANSACTIONS ON INFORMATION THEORY; VOL. IT-32, N0.2, MARCH 1986. [18] D. Kazakos and P. Papantoni-Kazakos, "Detection and Estimation", New York Computer Science Press, 1990. [19] Louis J. Ippolito, Jr. ITT Advanced Engineering & Sciences, USA, and The George Washington University, Washington, DC, USA, "Satellite Communications Systems Engineering Atmospheric Effects, Satellite Link Design and System Performance". [20]DennisRoddy, "Sattelite Communications", Third Edition, McGraw.Hill TELECOM Engineering. [21] Robert W. Jones, Radio communication Bureau, "Handbool on Satellite," Third Edition. WIEL Y. [22] "BP chief to testify to uncertainty of efforts to stop oil leak". CNN.2010-06-17. http:/ /edition .cnn.com/20 l 0/US/06/ 16/gulf.oil.disaster/index.html. Retrieved 25 June 20 l 0. [23] "Press Briefing by National Incident Commander June 21, 2010". Deepwater Horizon Unified Response Command. 21 June 2010. http://www.deepwaterhorizonresponse.com/go/doc/2931!683223. Retrieved 25 June 20 l 0. [24] "The Ongoing Administration-Wide Response to the Deepwater BP Oil Spiff'. Deepwater Horizon Incident Joint Information Center. 15 July 20 I 0. http://app.restorethegulf.gov/go/doc/2931 1786995/. Retrieved 16 July 2010. [25] P. Papantoni-Kazakos, "Algorithms for Monitoring Changes in Quality of Communication Links," IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. COM-27, NO.4, APRIL 1979. [26] S. Gazor and W. Zhang, "Speech probability distribution," IEEE Signal Processing, vol. 10, pp. 204-207, July 2003. [27]Rakesh K. Bansal and P. Papantoni-Kazakos, "An Algorithm for Detecting a Change in a Stochastic Process," IEEE Trans. on lnfomration Theory; vol. IT-32, No.2, March 1986. [28] Jong Won Shin, Hyuk Jin Kwon, Suk Ho Jin, and Nam Soo Kim," Voice Activity Detection Based on Conditional MAP Criterion" IEEE SIGNAL PROCESSING LEITERS, VOL. 15, 2008. [29] Jongseo Sohn, Nam Soo Kim, and Wonyong Sung "A Statistical Model-Based Voice Activity Detection", IEEE SIGNAL PROCESSING LEITERS, VOL. 6, NO. 1, JANUARY 1999. [30] Jong Won Shin, Joon-Hyuk Chang, Hwan Sik Yun, and Nam Soo Kim," VOICE ACTIVITY DETECTION BASED ON GENERALIZED GAMMA DISTRIBUTION". 100

PAGE 112

[31] ITUT, "A silence compression scheme for G. 729 optimised for terminals conforming to ITUT V.70," ITU-T Rec. G.729 Annex B, Nov. 1996. [32] A.T. Burrell and P. Papantoni-Kazakos "Detecting Software Faults in Distributed Systems" IEEE 2009 World Congress on Computer Science and Information Engineering. [33] Charles W. Steams "A Generalized Hann Window for Apodization of Filtered Backprojection PET Images". 2005IEEE Nuclear Science Symposium Conference Record 101