Title: Convergency of Telecommunication and Computer Networks
1Convergency of Telecommunication and Computer
Networks
2- In next pages you can see examples of our
researches on the theme Convergency of
telecommunication and computer networks
3GENETIC ALGORITHM AND NEURAL NETWORK
41. Objective
- application of Genetic algorithm (GA) for design
of Neural Network (NN) - control of communication Network Element (NE) by
NN - seek an alternative way of increasing the
performance of NE by parallel data processing
52. Introduction
- classical version of Genetic algorithm uses three
genetic operators reproduction, crossover and
mutation - Radial Basis Function Network (RBFN) is here
used - it is a type of single-direction multilayer
network
63. Design of a General Schematic of the Genetic
Algorithm
- Generating of initial population
- Ageing
- Mutation
- Calculation
- Sorting of upward population
- Crossing
- Finalization
7Generating of initial population
- initialization of all bits of all chromosomes in
initial generation is random - generator of random numbers which is a standard
features of the C Builder 5 - Gray code is used to encode the chromosome bits
8Ageing
- ageing shows up variants with limited length of
life - all the individuals in the population have their
age incremented - if it exceeds a set limit, the element is removed
from the population and a new element is randomly
generated in its place.
9Mutation
- classical method
- back-propagation method
10Calculation
- Genetic algorithm performs the minimization of an
error function
11Sorting of upward population
- we are looking for the function minimum so that a
chromosome with the least object function value
will be in the first place - Quicksort algorithm is used for sorting
- necessity of sorting will not affect the speed of
algorithm negatively
12Crossing
- uniform crossing is used
- every bit of descendants is with the probability
0.5 taken from one of the parents - one half of the population is made by the
crossing
13Finalization
- return to the Ageing if the finalization
condition is not realized - if it is realized then the end of Genetic
algorithm is made - they obtain two finalization variation (or their
combination), they are - 1) maximal number of iterations
- 2) quality of the best solution, smaller then
entered
144. Implementation of GA
- initialization of Genetic algorithm
- calculation of new generation
15Initialization of Genetic algorithm
- we have population in of the number of 100
chromosomes (randomly generated) ) at the
beginning - we preset engaged chromosomes longevity for each
of them and calculate their value of fitness
function - we sort out them according to sizes of fitness
uplink, by this we set the best chromosome on the
first position - classical algorithm QuickSort for sorting is used
16Calculation of new generation
- we expire all chromosomes and recognize its age
for each of them, if it equals to zero - if it is true, we generate new chromosome on its
place and we calculate its fitness - it is necessary again to order the whole
population, the best chromosome to be forever on
the first place. -
17Calculation of new generation
- further it follows crossing
- two parents from whole population are selected at
first, the crossing is made - two new descendants originate and we calculate
for them fitness value - we advance so further, that the number of
chromosomes, from which we select, is decrease
about 8 - on the last step the size of population, from
which parents for crossing will be selected,
equals to 8, it means, that it will be selected
from the first 8 (the best) chromosomes
185. Practical using
- modern possibility, how to change classical
sequential control of Network Elements NE to
control using of neural networks - design a simulation of NE, containing in the
process of control of switching area artificial
neural network with GA - NE switches single data units making provision
for priority
19Fig. 1 Model of the switch with artificial neural
network
20The basic scheme of the element
- We think over the single-stage switching area,
which has three inputs and three outputs, it is
switch on the Fig.2 The switching area is
realized on the cross-bar switch, i.e. in the
described case the switching area with 9
switching points. We can connect arbitrary input
to arbitrary output.
21Fig.2 Switch
22Fig.3 Switching area
23Fig.4 Switching area with addressing
24Fig.5 Frame structure
25Fig.6 Switching area controlled by control matrix
264. Conclusion
- crushing majority to learn neural network for
diagnostic of one object completely on 100 with
GA - time of learning is shorter than for classical
methods - the results and the learning time highly depends
on GA parameters setting - the best results were obtained by GA using the
D-operator and not using sexual reproduction. - it is shown modern possibility, how to change
classical sequential control of network elements
to control using of neural networks -
27Back-Propagation and K-Means Algorithms Comparison
28- The slides describes the application of
algorithms for object classification by using
artificial neural networks. The MLP (Multi Layer
Perceptron) and RBF (Radial Basis Function)
neural networks were used. We compared results
obtained by a using of learning algorithms
Back-Propagation (BP) and K-Means. The real
technological scene for object classification was
simulated with digitization of two-dimensional
pictures.
291 Introduction
- Pattern recognition consists in sorting objects
into classes. Class is a subset of objects whose
elements have common features from the
classification standtpoint. Object has a physical
character, which in computer vision is most
frequently taken to mean a part of segmented
image.
30- Methods for the classification of objects
constitute last and upper-most step in computer
vision theory. - The following methods were mutually compared
- Recognition with the aid of Back-Propagation
algorithm. - Recognition with the aid of K-Means algorithm.
312 Back-Propagation Algorithm
- Back-Propagation algorithm is an iterative method
where the network gets from an initial
non-learned state to the full learned one
32(1)
33(2)
34(3)
35- The following steps can describe the appurtenant
back-propagation algorithm
36- Initialization. All the weights in the network
are randomly set at values in the recommended
range lt0.3, 0.3gt. - Pattern submitting. A chosen pattern from the
training set is put in network inputs. Then
outputs of particular neurons are computed under
relations (2) and (3).
37- Comparison. This step contains the computation of
the neural network energy under relation (1) and
the error for the output layer under the
relation (4) - Back-propagation of an error and weight
modification. The values
38(5)
39(6)
40- Termination of pattern selection from the
training set. Another pattern from the training
set is chosen and the step number 2 follows until
all patterns were submitted.
41- Termination of learning process. The algorithm
ends when the neural network energy in last
computation has been less then the criterion
selected.
42Radial Basis Function Network
- This network belongs to the most recent neural
networks. It is a type of forward multi-layer
network with counter-propagation of signal and
with teacher learning. The network has two
layers, with different types of neurons in each
layer. Its advantage is mainly the speed of
learning.
43- The structure of this two-layer network is
similar to that of the Back-Propagation type of
network but the function of output neurons must
be linear and the transfer functions of hidden
neurons are the so-called Radial Basis Functions
hence the name of the network. The
characteristic feature of these functions is that
they either decrease monotonically or increase in
the direction from their centre point. Excerpt
for the input layer, which only serves the
purpose of handing over values, an RBF network
has an RBF layer (hidden layer) and an output
layer formed by perceptrons.
444 Problem Solution
45(No Transcript)
46(No Transcript)
47(No Transcript)
485 Conclusion
- Back propagation algorithm presented very good
results at classification. The network recognized
all the patterns submitted. The learning using
the BPx method (extended Back propagation
algorithm) was a little slower, but it can be
succesfully used for networks with lower number
of neurons. Radial Basis Function networks can be
designed very quickly.
49- The time necessary for network learning was very
little. The network was able to classify
correctly 100 models and at the same time to
recognize correctly even slightly damaged models.
As the number of radial basis neurons is
comparable the input space size and problem
complexity, RBF networks can be larger than
back-propagation networks. Recognition with the
aid of neural network is suitable where
high-speed classification with randomly rotated
objects is required and where we need to tolerate
some differences between learned etalons and
classified objects.
50Codec G 723.1 by MATLAB simulation
51- The objective of this slides is an introduction
to multimedia signals compressions under ISDN.
Compression and following expansion, in agreement
with standard, will be simulate with the help of
programme MATLAB. Authors contribution is codec
G.723.1 by MATLAB simulation. The ITU T block
structure is observed.
52- Codec realization in MATLAB is applied on test
speech signal and results were indicated in this
research and paper. Graphics results are one part
of research, in the paper is not enough place for
this description. Presentation paper contains
results of simulation in MATLAB programme for
audiocodec by the recommendation G.723.1. This
recommendation is used for ISDN as example and
although ISDN is now replaced by xDSL, software
base used for source encoding is available also
for the future.
531. Introduction
- Simulation is made by MATLAB programme. Resulting
programme is placed on created CD. In MATLAB
command line must be written encoder(name1.wav,
name2.bs), where name1 is name of input wav
file and name2 is encoded sequence, BitStream
by G.723.1, which will be next suitable as input
to decoder.
54- For decoding of this sequence and acquirement
requested wav file must be written
decoder(name2.bs, name1.wav) where name2 is
name of input BitStream file which was obtained
from previous encoding and name1 is resulted
wav type file. To this resulted file was apllied
encoder G.723.1.
55- Input encoder file must be 16-bit, singlechannel
with sampling frequency 8000 Hz, integer type. - Both bit rates are obligatory parts of encoder
and decoder. Bit rates can be overswitched also
in encoder operation, every frame by different
rate. Encoder G.723.1 is optimalized for speech
representation. Music or other audiosignals are
not expressed so truly as speech, but it is
possible to encoded and decoded them.
562. Encoder and Decoder G 723.1
- For encoding is used linear prediction, that is
why codec works on the principle
analysis-synthesis. For higher bit flow is used
Multipulse Maximum Likelihood Quantization (MP -
MLQ) method, for lower bit flow Algebraic Code
Excited Linear Prediction (ACELP) method. The
frame has length 30 ms plus adding 7.5 ms. Whole
algorithm delay is 37.5 ms.
573. Simulation
- Simulation is made by MATLAB programme. Resulting
programme is placed on created CD. In MATLAB
command line must be written encoder(name1.wav,
name2.bs), where name1 is name of input wav
file and name2 is encoded sequence, BitStream
by G.723.1, which will be next suitable as input
to decoder.
58- For decoding of this sequence and acquirement
requested wav file must be written
decoder(name2.bs, name1.wav) where name2 is
name of input BitStream file which was obtained
from previous encoding and name1 is resulted
wav type file. To this resulted file was apllied
encoder G.723.1. - Input encoder file must be 16-bit, singlechannel
with sampling frequency 8000 Hz, integer type. - Some short speech signals was executed by the
simulated Codec G. 723.1.
59- The different between the original signal and
decoded signal is very small. More interesting
results are on the Fig.2. On this figure is
displayed time behaviour of the shorter speech
elements of the e sound. The figure contains
element of the length 300 samples, it is about
40ms if the sampling frequency is 8kHz.
60(No Transcript)
61(No Transcript)
62(No Transcript)
63(No Transcript)
644. Conclusion
- This research is dedicated to introduction to
multimedia compressions signals problems.
Following describe ISDN multimedia compressions,
it is more detailed then before basic describe.
The next problem is desription of recommendation
ITU-T G.723.1, which is used in practical part of
this research.
65- In practical part of this research is codec
realisation in programme MATLAB. This research
gives integrated information about compression
methods, which are used with the most accent to
audio compression, specify to speech. Codec
realisation in MATLAB is applied on test speech
signal and selected results are indicated in this
paper.
66Chaotic signal for transmission of information
67- Authors contribution is design of CPPM modulator
(it is contribution for this paper, for complete
research it is also CPPM demodulator and
transmission channel). The pulse generator, which
makes chaotic sequence of impulses, is the basic
block of CPPM modulator. The impulse is generated
for every front edge of time signal. Mentioned
pulse generator is controlled by chaotic time
signal.
68- The creation of chaotic time signal is next step
of design. It is known, that necessary condition
of chaotic system behaviour is the non-linearity
of this system. The principal results of this
research is CPPM modulator design and simulation
of its behaviour in MATLAB and SIMULINK
environment.
691. Introduction
- Orbit behaviour in the vicinity of balance state
of autonomous system is described by the self
numerous of Jacobins matrix (matrix of
linearization) and by description of trajectories
behaviour in the vicinity of closed trajectory ?
are suitable their multipliers. Trajectory
multipliers ? are self numerous of Jacobins
matrix of Poincaré projection in the fix point .
70- For the description of trajectories of generally
arbitrary trajectory G are used Ljapunov
exponnets (called also characteristic exponents).
Ljapunov exponents (LE) are generalization of
self numerous or multipliers. Ljapunov exponents
are real numerous and they are suitable for
classification of chaotic or non-chaotic
atractors.
71(1)
72- Asymptotic orbits behaviour in vicinity of G(x0)
is given by asymptotic behaviour of Jacobins
matrix (matrix of linearization) of the flow
D?t(x0) of the equation in limits t??.
73(2)
742. Design of CPPM modulator
- Basic element of CPPM system is pulse generator.
On its output is chaotic sequence of impulses.
This generator is controlled by chaotic time
signal. In every front edge of time signal will
be generated one impuls. Every impuls has the
same amplituce and the same duration.
75(No Transcript)
763. Implementation CPPM in the SIMULINK
environment
- Using programme environment MATLAB and SIMULINK
was modelled CPPM signal transmission through
radio environment. Communication system block
diagram is on the figure 2. Signal CPPM is in the
radio channel obstructed by additive white noise
and by interference with obstructing signals.
Results are in our research work.
77(No Transcript)
784. Conclusion
- This paper deals with different communication
techniques which use a chaotic signal for
transmission of information. These informations
are studied from the synchronization and noise
resistance point of view. The main part of this
work is concentrated on CPPM modulation. The
final result is to design a communication system
CPPM by means of MATLAB and SIMULINK
respectively.
79- Several simulations were done for distinct
parameters values of CPPMs modulator and
demodulator. In the end, the multi-user
transmission has been studied and corresponding
BER has been confirmed.
80Research of Advanced Multimedia Transmission
81- One period of research of advanced multimedia
transmission has been finished in our department.
A lot of papers from this problem were published
in the previous conferences and this paper
encloses it. The main problems were mapping the
actual status of audio signal encoding and bit
rate compression, focused to MPEG-1 method,
encoder and decoder MPEG-1 block diagrams and
design of MPEG-1 model simulated by MATLAB.
82- This model was suitable for testing and
optimizing of MPEG-1 principle, analysis and
simulation of the Filter Bank, design and
modification of scale-factor calculating,
modelling of FFT for MPEG-1, determination of the
sound pressure level, study of tonal and
non-tonal components, decimation of tonal and
non-tonal masking components design, individual
masking threshold and global masking threshold,
determining the minimum masking threshold,
signal-to-mask ratio, design of bit allocation
and optimal quantisation and encoding of subbands
samples.
83I.Introduction
- The research has been divided in several chapters
that covers all the possible explenation of the
abstract. Chapter 1 gives a general introduction
about (MPEG) Motion Picture Experts Group and an
overview of the current state of this scientific
issue. Chapter 2 it cover the problem
definition and research outlines. Chapter 3 ties
into these and describes in details the algorithm
used in MPEG standards. Chapter 4 describes the
creation of an MPEG-1 sound coder/decoder with
all steps how to design, as it will be used in
the experiments to be performed.
84- Chapter 5 contains a main results of the
research and Conclusion we can read at Chapter 6.
The first problem was to research encoding and
decoding process of MPEG-1 standard, which gives
the possibility to program Layer-1. First of all,
it design the algorihtm for possibility to
simulate each part of the MPEG-1. MATLAB is used
for simulation of mathematical calculations. It
allows to display each variable with a lot of
format methods. It seems good for displaying
calculated variables or searching mistakes in
algorithm, when program is constructed.
85II.Simulation OF MPEG-I IN MATLAB
- A short description of MATLAB programming was
given in this section. The program consists of a
number of files. The central unit is mpeg1a.m.
The important global variables such as bit-rate,
length of window for analysis (encoding) and
synthesis (decoding), number of samples in one
subband in the frame, number of subbands and more
other parameters are defined in this file. This
central program unit calls table
absolute-threshold and mapping.m.
86- These files are used for encoding and decoding
purposes. For encoder part of the program is much
longer than decoder part. This is the most
important part of MPEG-1 standard. The number of
runs of encoding and decoding are calculated. In
fact, one run of encoding means calculating of
one frame. This number is calculated from length
of input file, which is taken from disk file.
Number of bits are then calculated for one frame,
because user can select arbitrary bitrate of list
taken from norm. This list is included in the
beginning of the main file.
87- The encode part in first run load many variables
into memory. In second, third, fourth runs and so
on, these variables in memory are prepared. It
gives the encoding process slower in first step
than in others. This program consists of subpart
for windowing, partial calculation, matrixing,
big subpart for psychoacoustic model and
quantization.
88A. Test Signal
- In this section, the outputs of the software
calculation in MATLAB are displayed in pictures.
These outputs are the outcomes of the input-test
signal. This signal was connected at the input in
MATLAB programming. With these signals, a lot of
main parts of compression principles are
introduced. In addition, it will be introduced on
real program in future, which is possible to run
in real time, but now will be used step by step
in developing software. The input chosen signal
consists of audio signal which take it from CD.
89B.Filter Bank Analysis
- In the MPEG audio coding, the input signal is
divided into 32 subbands of equal bandwidth. In
each subband, the signal is scaled and quantized
in order to keep the quantizing noise below the
masking curve. The result of the encoding is a
series of scale factors, quantizer information
and coded samples for each subband.
90- According to analysis of subband
filter
flowchart, the filter does a time to frequency
mapping. The 32 subband polyphase filter presents
optimized characteristics with respect to a
performance complexity ratio. The filter, of
order 511, with side lobe rejection better than
96 dB, is a compromise between two competing
specifications of the filter response the
spectral resolution and the transient impulse
response (TIR). The spectral resolution of the
filter bank is important, because it corresponds
to the critical bands found in the human ear.
91- The information yielded by the filter, when
compared with empirical perceptual data,
facilitates the reduction of the bit information
by eliminating masked, unnoticed spectra, and
reduces the number of bits allocated for spectra
carrying low amounts of information. The
time-frequency mapping of the filter allows a
reasonable emulation of the critical bands of the
ear, which correspond to a width of about 100 Hz
in frequencies below 500 Hz, and width of about
20 of the center frequency at higher
frequencies.
92C. Subband Analysis in MATLAB
- In the start of the flow chart, there are two
steps, which are prepared for new input samples.
Firstly, all the samples in the buffer (512
samples) are down shifted and then put new 32
samples at the end of the buffer, which is
required for further calculations. But in this
work different algorithm, which contain both in
only one step is used. One more variable known as
ofs is present, which shows where is the end of
the previous data read, because 32 samples into
512 samples buffer is read, it is different
number every time
93- . Instead of shifting in each run of flowchart
(it is 12 times in one frame) only offset is
shifted. It gives more time for other algorithms.
When 32 new samples are added, it gives 512
together, which is all defined length for buffer
for windowing.
94D. Scale-Factor Calculating
- A scale technique is used in the MPEG-1 coding
scheme, which provides an effective overall
dynamic range of 120 dB/subband with a resolution
of 2dB/scale factor class. The calculation of the
scale factors is made for every 12 subband
samples. The maximum value of these samples is
determined. The lowest value, which is larger
than this maximum value is used as scale-factor.
95- Scale-factor can be calculated after
psychoacoustic model according to standard. In
fact, in MATLAB program, the scale-factor was
chosen before psychoacoustic model. The algorithm
contains to looking for maximum value of samples
and after searching minimum appropriate
scale-factor.
96E. Fast Fourier Transformation (FFT) Analysis
- The masking threshold is derived from an
estimate of the power density spectrum that is
calculated by a 512-point FFT. The FFT is
calculated directly from the input PCM signal,
windowed by a Hann window
97- For a coincidence in time between the bit
allocation and the corresponding subband samples,
the PCM samples entering the FFT have to be
displayed. The delay of the analysis subband
filter in 256 samples, corresponding to 5.3 ms at
the 48 kHz sampling rate 2. A window shift of
256 samples is required to compensate for the
delay in the analysis subband filter. In
addition, the Hann window must coincide with the
subband samples of the frame. For Layer-1, this
amount to an additional window shift of 64
samples
98III.Determination of the sound pressure level
- The more powerful version was chosen for the
determination of the sound pressure level in
program in MATLAB. A maximum is found from all
spectral lines in each subband and one pressure
level gives by scalefactors. In fact, it means
two loops, first is making a searching through
spectral lines in each subband, second is
searching just through subbands comparing to an
appropriate pressure level and maximum from first
loop.
99A.Finding of Tonal and Non-Tonal Components
- The tonality of a masking component has an
influence on the masking threshold. For this
reason, it is worthwhile to discriminate between
tonal and non-tonal components. For calculating
the global masking threshold, it is necessary to
derive the tonal and non-tonal components from
the FFT.
100B.Decimation of Tonal and Non-Tonal Masking
Components
- Decimation is a procedure used to allow
considering fewer maskers when calculating the
global masking threshold. - The decimation was separated into three steps.
Each step is represented by one loop, with a
comparison at the beginning. After the
comparison, a tone or noise is kept or removed,
depending on the comparisons result. These three
parts are decimation for tonal components,
decimation for non-tonal components, and
decimation for tonal components within a distance
of less than 0.5 Bark.
101C.Calculating the Individual Masking Threshold
- The individual masking thresholds of both tonal
and non-tonal components are calculated according
to the MPEG-1 standard 1.
102- The masking function vf of a masker is
characterized by different lower and upper
slopes, which depend on the distance in Bark
dz z (i) z (j) to the masker. In this
expression, i is the index of the spectral line
at which the masking function is calculated and j
that of the masker. The critical band rates z(j)
and z(i) can be found in the tables in the
standard. These are the same tables as before.
Therefore, in the present project, the table for
FS 44.1kHz is included in the appendix with the
MATLAB program.
103D.Quantization and Encoding of Subband Samples
- All computations in the encoder were finished at
the end of the previous chapter. Now the encoder
has to transmit the data. The next paragraphs
describe how the data is transmitted.
104IV.Conclusion
- The main parts of the research was
- Mapping the actual status of audio signal
encoding and bit rate compression, focusing to
MPEG-1 method, design of encoder and decoder
MPEG-1 block diagrams, design of MPEG-1 model
simulated by MATLAB, analysis and simulation of
the Filter Bank,
105- design and modification of scale-factor
calculating, modelling of FFT for MPEG-1,
determination of the sound pressure level,
decimation of tonal and non-tonal masking
components design, individual masking threshold
and global masking threshold, determining the
minimum masking threshold, signal-to mask ratio,
design of bit allocation, optimal quantization
and encoding of subbands samples.
106- The programming for MPEG-1 audio encoder and
decoder for Layer-1 is discussed in detail in
this paper. We use MATLAB for making the proposed
algorithm, which will be applied later on the
Digital Signal Processor (DSP). We did not use
special instructions from MATLAB like
107- In the previous slides you can see possibilities
of the convergency of telecommunication and
computer networks. - Thank you for your interest
- If you have any questions, then write e-mail to
- skorpil_at_feec.vutbr.cz