Site Loader
Rock Street, San Francisco

EEG signal processing is one of the hottest countries of research in digital signal processing applications and biomedical research. Analysis of EEG signals provides a important tool for diagnosing of neurobiological diseases. The job of EEG signal categorization into healthy and pathological instances is chiefly a form acknowledgment job utilizing extracted characteristics. Many methods of characteristic extraction have been applied to pull out the relevant features from a given EEG informations. The EEG information was collected from a publically available beginning. Three types of instances were classified viz. signals recorded from healthy voluntaries holding their eyes open, epilepsy patients in the epileptogenic zone during a seizure-free interval, and epilepsy patients during epileptic ictuss. The characteristic extraction was done by calculating the distinct ripple transform and spectral analysis utilizing AR theoretical account. The ripple transform coefficients compress the figure of information points into few characteristics. Assorted statistics were used to further cut down the dimensionality. The AR coefficients obtained from burg auto-regressive method provide of import characteristics of the EEG signals. Categorization of the EEG informations utilizing commission nervous web provides robust and improved public presentation over single members of the commission. F-ratio based dimension decrease technique was used to cut down the figure of characteristics without impacting the truth much.

I

List OF FIGURES

Figure 1.1: Conventional position of the scalp……………………………..…………….… … … 2

Figure 1.2: Structure of a nerve cell… … … … … … … … … … … … … … … … … … … … … … … … … … … … … .3

Figure 1.3: Conventional 10-20 electrode arrangement strategy… … … … … … … … … … … … … … ..4

Figure 1.4: Block diagram of EEG signal processing… … … … … … … … … … … … … … … … … … ..7

Figure 1.5 ( a ) EEG section of category A…………………………………………….……..8

( B ) EEG section of category D… … … … … … … … … … … … … … … … … … … … … … … … … … ..9

( degree Celsius ) EEG section of category Tocopherol… … … … … … … … … … … … … … … … … … … … … … … … … … ..9

Figure 2.1: Representation of a moving ridge and ripple… … … … … … … … … … … … … … … … … … … ..11

Figure 2.2: DWT calculation utilizing filter Bankss… … … … … … … … … … … … … … … … … … … … .13

Figure 2.3: Ripple families…………………………………………..…………… … … .15

Figure 2.4: ( a ) Wavelet coefficients of category A……………………………………….…17

( B ) Wavelet coefficients of category D… … … … … … … … … … … … … … … … … … … … … ..17

( degree Celsius ) Wavelet coefficients of category E……………………………………….…18

Figure 2.5: ( a ) Power spectral denseness of category A utilizing AR theoretical account… … … … … … … … … … … 20

( B ) Power spectral denseness of category D utilizing AR theoretical account… … … … … … … … … … ..20

( degree Celsius ) Power spectral denseness of category E utilizing AR theoretical account… … … … … … … … … … … 21

two

Figure 2.6: ( a ) Feature vector of category Angstrom… … … … … … … … … … … … … … … … … … … … … … … … ..21

( B ) Feature vector of category D… … … … … … … … … … … … … … … … … … … … … … … … ..22

( degree Celsius ) Feature vector of category Tocopherol… … … … … … … … … … … … … … … … … … … … … … … … … 22

Figure 2.7: Model of a nerve cell… … … … … … … … … … … … … … … … … … … … … … … … … … … … … … 23

Figure 2.8: Structure of a nervous web… … … … … … … … … … … … … … … … … … … … … … … … .24

Figure 2.9: Committee nervous web… … … … … … … … … … … … … … … … … … … … … … … … … .28

Figure 3.1: Diagram for multi-cluster informations… … … … … … … … … … … … … … … … … … … … … … … ..33

three

List OF TABLES

Table 2.1: Statisticss of ripple coefficients… … … … … … … … … … … … … … … … … … … … … … … .16

Table 2.2: Confusion matrix of nervous web 1… … … … … … … … … … … … … … … … … … … … .29

Table 2.3: Confusion matrix of nervous web 2… … … … … … … … … … … … … … … … … … … … .29

Table 2.4: Confusion matrix of nervous web 3… … … … … … … … … … … … … … … … … … … … .29

Table 2.5: Statistical parametric quantities of nervous web 1… … … … … … … … … … … … … … … … … … .30

Table 2.6: Statistical parametric quantities of nervous web 2… … … … … … … … … … … … … … … … … … .30

Table 2.7: Statistical parametric quantities of nervous web 3… … … … … … … … … … … … … … … … … … .30

Table 2.8: Statistical parametric quantities of Committee nervous web… … … … … … … … … … … … … 31

Table 2.9: Accuracy of Committee nervous network………………………..……………31

Table 3.1: F-ratio of the extracted characteristics… … … … … … … … … … … … … … … … … … … … … … … ..35

Table 3.2: Decrease of characteristics… … … … … … … … … … … … … … … … … … … … … … … … … … … … … 36

four

Chapter 1

Introduction

1.1 What are EEG Signals?

Electroencephalography ( EEG ) is the recording of self-generated electrical activity of the encephalon which is obtained by firing of nerve cells within the encephalon. EEG signals are recorded in a short clip, usually for 20-40 proceedingss. We get the recordings by puting the electrodes at assorted places on the scalp. Figure.1 shows the conventional position of the scalp and points represent the placing of the multiple electrodes on the scalp. It is believed that the EEG signals non merely represents the encephalon signal but represents the position of the whole organic structure. The diagnostic application in instance of epilepsy gives us the premier motive to use the Digital Signal Processing techniques to the EEG signals.

Figure 1.1: Conventional position of the scalp [ 20 ]

EEG coevals

An EEG signal is generated due to the currents that flow between the encephalon cells in the intellectual cortex part of the encephalon. When the nerve cells are activated, current flows between dendrites due to their synaptic excitements. This current generates a magnetic field and a secondary electric field. The magnetic field is mensurable by EMG ( EMG ) machines and the electric field is measured by EEG systems over the scalp [ 1 ] .

The human caput consists of assorted beds including the encephalon, skull, scalp and other thin beds in between. The degree of fading due to skull is about 100 times greater than that of the soft tissues. While entering EEG signals noise can be internal ( generated within the encephalon ) or external ( over the scalp ) . Hence merely a big figure of activated nerve cells can bring forth adequate potency to hold a recordable signal. These signals have to be amplified for farther processing [ 2 ] .

Degree centigrades: UsersmanguDesktopTHESIS WORK
euron.png

Figure 1.2: Structure of a nerve cell [ 2 ]

EEG recordings

EEG systems consist of a figure of electrodes, differential amplifiers, filters and needle ( pen ) -type registries [ 3 ] . The EEG signals can be easy plotted on paper. Recent systems use computing machines for digitisation and storing intents. For digitisation sampling, quantisation and encryption is done. The effectual bandwidth of the EEG signals is about 100 Hz. Thus a lower limit of 200 samples per second is necessary for trying ( Nyquist standard ) . For quantisation representation utilizing 16 spots is largely used. Fig 1.3 shows the conventional electrode agreement recommended by the International Federation of Societies for Electroencephalography and Clinical Neurophysiology for 21 electrodes ( called 10-20 electrode place ) [ 4 ] .

Degree centigrades: UsersmanguDesktopTHESIS WORKelectrode.png

Figure 1.3: Conventional 10-20 EEG electrode places for the arrangement of 21 electrodes

Brain beat

Brain beat can be easy recognized by ocular review of the EEG signal therefore many neurological upsets can be easy identified. The amplitude and frequence of these signals vary with human province ( asleep or awake ) , age, wellness etc. There are five major encephalon moving ridges distinguished by their frequence ranges. These are alpha ( ? ) , theta ( ? ) , beta ( ? ) , delta ( ? ) , and gamma ( ? ) .

Alpha waves frequences lie within the scope of 8-13 Hz. They can be detected in the posterior lobes of the encephalon. In the whole kingdom of encephalon activity alpha moving ridges are the most outstanding beat. They are detected in a normal individual when he is in a relaxed province without any attending or concentration. Closed oculus province besides produces some alpha moving ridges [ 5 ] .

A beta moving ridge lies in the scope 14-26 Hz. It is chiefly encountered in frontal and cardinal parts. It is the usual waking beat of the encephalon associated with active concentration, active

thought, job resolution, concentrating on things. When a individual is in a panic province a high degree beta moving ridge is generated [ 6 ] .

Theta waves lie within the scope 4-7.5 Hz. It is assumed that it has origins in the thalamic part. When a individual is stealing into a drowsing province from witting province theta moving ridges are observed. They play a important function in babies and immature kids. Creative thought, deep speculation, entree to unconscious stuff is associated with theta moving ridges [ 7 ] .

Delta moving ridges are within the scope 0.5-4 Hz. They are found frontally in grownups and posteriorily in kids. They are associated with deep slumber and may be present in waking province.

Gamma moving ridges are besides called fast beta moving ridges and they have frequences above 30 Hz.

The amplitude of these moving ridges is really low and they have rare happening. They are associated with certain cognitive and motor maps. Detection of these beats can be used to corroborate certain neurological diseases. It is besides a good index of event related synchronism ( ERS ) of the encephalon [ 8 ] .

1.5 Why do we utilize EEG signals?

There are assorted advantages of EEG signals some of them can be provinces as follows:

Temporal declaration of the EEG signal is high.

EEG measures the electrical activity straight.

EEG is a non-invasive process.

It has the ability to analyse the encephalon activity ; it unfolds in existent clip at degree of msecs, i.e. 1000s of a 2nd.

It is really difficult to happen the beginning of electrical activity where the electrical activity is coming from. This is the major disadvantages of EEG signals. By puting the multiple electrodes on the scalp we can acquire some information where the ERP is strongest.

Electroencephalogram signals are used for assorted undertakings. We can split the utilizations of EEG as clinical utilizations and research uses [ 21 ] :

Clinical utilizations:

Electroencephalogram signals are used in the diagnosing of several neurological diseases.

Electroencephalogram signals are used to qualify the ictuss for the intent of intervention.

Electroencephalogram signals are used to supervise the deepness of anaesthesia.

Electroencephalogram signals are used to find the wean-epileptic medicine.

Research uses:

Electroencephalogram signals are used in neuroscience.

Electroencephalogram signals are used in cognitive scientific discipline.

EEG signals can be used for the psychophysiological research.

EEG signals can be used for the survey of the responses to audile stimulations.

1.6 Objective

Our aim is to analyse the EEG signals and sort the EEG informations into different categories. Our chief mark is to better the truth of EEG signals. In our undertaking we have besides applied optimisation techniques to cut down the calculation complexness of the web without impacting the truth of the categorization.

Figure. 4 shows the block diagram for the EEG signal processing. For the categorization intent we have taken the natural EEG signals available at [ 9 ] . From the five informations sets available we have selected 3 sets ( put A, set D, set E ) . In set A EEG signals were recorded from healthy voluntaries. In set D recordings were taken from within epiletogenic zone but during seizure free interval while set E contained merely seizure activity. For each of these informations we extracted the characteristics utilizing distinct ripple transform and Auto-Regressive coefficients method. After characteristic extraction the different informations sets were classified utilizing commission nervous web trained with back extension algorithm. To cut down the dimensionality of the characteristics we have used Fisher ‘s ratio based optimisation technique.

Figure 1.4: Block diagram of EEG signal categorization

In chapter 2 our proposed technique for the categorization intent is discussed, so in the chapter 3 F-ratio based technique to cut down computational complexness. In chapter 4 we conclude our thesis along with future work.

1.7 Data Choice

The EEG informations used in this survey was obtained from the database available with the Bonn University. This information is publically available at [ 9 ] . The complete dataset consists of five categories ( A, B, C, D, E ) each of which contains 100 individual channel EEG sections of 23.6s continuance. Each section was selected and cut out from uninterrupted multichannel EEG recordings after ocular review for artefacts e.g. due to oculus motion or musculus activity.

Sets A and B consisted of signals taken from surface EEG recordings that were carried out on five healthy voluntaries utilizing a standardised electrode arrangement strategy ( International 10-20 system ) . Volunteers were relaxed in an awake province with eyes unfastened ( A ) and eyes closed ( B ) , severally. Sets C-E originated from the EEG archive of presurgical diagnosing. Electroencephalogram from five patients were selected, all of whom had achieved complete ictus control after resection of one of the hippocampal formations, which was hence right diagnosed to be the epileptogenic zone. Signals in set D were recorded from within the epileptogenic zone, and those in set C from the hippocampal formation of the antonym hemisphere of the encephalon. While sets C and D contained merely activity measured during seizure-free intervals, set E merely contained ictus activity.

Using an mean common mention, all EEG signals were recorded with the same 128-channel amplifier system. The informations were digitized at 173.61 samples per 2nd utilizing 12 spot analog-to-digital convertor. The scenes of the set base on balls filter were 0.53-40 Hz ( 12 dB/oct. ) [ 10 ] . In the present survey sets A, D, E was used.

1.8 Raw EEG signal

From the information available at [ 9 ] , a rectangular window of length 256 distinct informations was selected to organize a individual EEG section. The secret plan of section of the three categories ( A, D, E ) is shown belowA.bmp

( a )

D.bmp

( B )

E.bmp

( degree Celsius )

Figure 1.5 ( a ) EEG section of category A ( B ) EEG section of category D ( degree Celsius ) EEG section of category Tocopherol

Chapter 2

Categorization of EEG Signal

2.1 Wavelet Transform

The transform of a signal is merely another manner of stand foring a signal as it does n’t alter any information content of a signal. Although short clip Fourier transform ( STFT ) can be used to analyse non-stationary signals, it has a changeless declaration at all frequences. The ripple transform gives a time-frequency representation and in this transform different frequences are analyzed with different declarations.

Wavelet transform uses ripples of finite energy. Ripples are localized moving ridges which are suited to analyse transients since their energy is concentrated in clip and infinite [ 11 ] .C: UsersmanguDesktopTHESIS WORKwave.png

( B )

Figure 2.1 Representation of ( a ) moving ridge ( B ) ripple

The ripple transform gives us multi-resolution description of a signal. It addresses the jobs of non-stationary signals and hence is peculiarly suited for feature extraction of EEG signals [ 12 ] . At high frequences it provides a good clip declaration and for low frequences it provides better frequence declaration, this is because the transform is computed utilizing a female parent ripple and different footing maps which are generated from the female parent ripple through grading and interlingual rendition operations. Hence it has a varying window size which is wide at low frequences and narrow at high frequences, therefore supplying optimum declaration at all frequences.

2.1.1 Continuous ripple transform

The uninterrupted ripple transform is defined as

Where ten ( T ) is the signal to be analyzed, ? ( T ) is the female parent ripple or the footing map ? is the interlingual rendition parametric quantity and s is the scale parametric quantity.

The Continuous ripple transform performs the whirl operation of the footing map and the signal. The female parent ripple is chosen depending upon the features associated with the signal. The interlingual rendition parametric quantity ? relates to the clip information nowadays in the signal and it is used to switch the location of the ripple map in the signal. The scale parametric quantity s correspond to the frequence information is defined as the opposite of frequence. Scaling expands or contracts a signal, therefore big graduated tables expand the signal and give the concealed local information while little graduated tables contract a signal and supply planetary information [ 11 ] .

2.1.2 Discrete ripple transform

The calculation of CWT consumes a batch of clip and resources and consequences in big sum of informations, hence Discrete ripple transform, which is based on sub-band cryptography is used as it gives a fast calculation of ripple transform. In DWT the time-scale representation of the signal can be achieved utilizing digital filtering techniques. The attack for the multi-resolution decomposition of a signal ten ( n ) is shown in Fig. 2.2. The DWT is computed by consecutive low base on balls and high base on balls filtering of the signal ten ( n ) . Each measure consists of two digital filters and two downsamplers by 2. The high base on balls filter g [ ] is the distinct female parent ripple and the low base on balls filter H [ . ] is its mirror version. At each degree the downsampled end products of the high base on balls filter produce the item coefficients and that of low base on balls filter gives the estimate coefficients. The estimate coefficients are farther decomposed and the process is continued as shown [ 13-14 ] .

Degree centigrades: UsersmanguDesktopTHESIS WORKdwt.png

Figure: 2.2 Discrete ripple transform block diagram [ 15 ]

The standard quadrature filter status is

where H ( omega ) is the Z-transform the low base on balls filter h. this filter can be used to stipulate all ripple transforms.

The complementary high base on balls filter is defined as

Now the sequence of filters can be obtained as

with the initial status H0 ( omega ) = 1. In clip sphere we have

The inferior [ . ] ^2k denotes upsampling by 2k. Here n is the discrete sampled clip.

The normalized ripple and scale footing map are defined as

where the factor 2k/2 is the interior merchandise standardization, K and cubic decimeter are the graduated table and interlingual rendition parametric quantity severally.

The DWT decomposition can be described as

where a ( K ) ( cubic decimeter ) and vitamin D ( K ) ( cubic decimeter ) are the estimate coefficients and the item coefficients at declaration K, severally [ 13-16 ] .

2.1.3 Wavelet households

There are a figure of basic maps that can be used as the female parent ripple for Wavelet transform. While taking the female parent wavelet the features of the signal should be taken into history since it produces the different ripples through interlingual rendition and dilation and hence determines the features of the ensuing transform. Figure 2.3 illustrates the normally used ripple maps. The ripples are chosen on the footing of their form and ability to analyse the signal for a peculiar application.

waveletfamily.png

Figure 2.3: Normally used ripple maps ( a ) Haar ( B ) Daubechies4 ( degree Celsius ) Coiflet1 ( vitamin D ) Symlet2 ( vitamin E ) Meyer ( degree Fahrenheit ) Morlet ( g ) Mexican Hat [ 11 ] .

2.2 Feature extraction utilizing distinct ripple transform

From the information available at [ 9 ] a rectangular window of length 256 distinct informations was selected to organize a individual EEG section. For analysis of signals utilizing Wavelet tranform choice of the appropriate ripple and figure of decomposition degree is of extreme importantce. The ripple coefficients were computed utilizing daubechies ripple of order 2 because its smoothing characteristics are more suited to observe alterations in EEG signal. In the present survey, the EEG signals were decomposed into inside informations D1-D4 and one estimate A4. For the four item coefficients we get 247 coefficients ( 129+66+34+18 ) and 18 for the estimate coefficient. So a sum of 265 coefficients were obtained for each section [ 15 ] .

To cut down the figure of characteristics following statistics were used:

Maximum of ripple coefficients in each bomber set

Minimum of ripple coefficients in each bomber set

Mean of ripple coefficients in each bomber set

Standard divergence of ripple coefficients in each bomber set [ 22 ]

Therefore the dimension of DWT coefficients is 20.

The tabular array for DWT coefficients of an EEG section of categories A, D, E is shown below

Table 2.1 Statistics of ripple coefficients

Dataset

Extracted Features

Ripple coefficients

Sub sets

D1

D2

D3

D4

Put A

Maximum

23.44104

65.70968

177.9029

83.28186

Minimum

-15.1013

-52.1523

-141.268

-171.945

Mean

-0.16537

-1.21974

-7.68956

-4.04699

Std. divergence

6.86211

24.09339

61.35413

63.63968

Set B

Maximum

9.763284

24.68172

72.32882

194.7452

Minimum

-8.08096

-28.8256

-84.667

-118.222

Mean

0.078406

0.366535

2.012951

9.170777

Std. divergence

2.820835

8.466435

36.18396

84.91487

Set C

Maximum

63.49726

309.0024

816.6531

1366.084

Minimum

-110.733

-317.317

-868.665

-1180.19

Mean

-0.69357

2.200813

-41.2569

-99.0486

Std. divergence

28.63028

117.3747

479.7756

712.5626

The elaborate ripple coefficients of set A, set D, set E EEG sections at the first decomposition degree is shown in the undermentioned figures.

dwta.bmp

( a )

dwtd.bmp

( B )

dwte.bmp

( degree Celsius )

Figure 2.4 ( a ) , ( B ) , ( degree Celsius ) Plot for Discrete ripple coefficients of category A, D and E severally

2.3 Feature Extraction utilizing Autoregressive Coefficients

The Autoregressive ( AR ) Power spectral denseness appraisal of the EEG signals of set A, set B and set C was computed. The Power spectral denseness is the distribution of power with regard to the frequence. Power spectral denseness Rxx of the random stationary signal can be expressed by multinomials A ( omega ) and B ( omega ) holding roots that fall inside the unit circle in the z-plane [ Pr ] as shown in the given expression [ 23 ]

,

where ?w is the discrepancy of the white Gaussian noise tungsten ( n ) . Now the additive filter H ( omega ) for bring forthing the random procedure x ( n ) from the white Gaussian noise tungsten ( n ) can be written as

,

Therefore the end product x ( n ) can be related to the input by utilizing the undermentioned difference equation:

If b0 = 1, berkelium = 0, K & A ; gt ; 0 so the additive filter H ( omega ) can be written as 1/A ( omega ) . Now the difference equation for the AR procedure can be reduced to

If ak = 0, k ? 1 so the additive filter H ( omega ) = B ( omega ) and the difference equation for the moving norm ( MA ) procedure can be written as follows:

In instance of Autoregressive traveling norm ( ARMA ) procedure additive filter H ( omega ) = B ( omega ) /A ( omega ) has both finite poles and nothings in the z-plane [ 23 ] .

Autoregressive coefficients are really of import characteristics as they represent the PSD of the signal which is really common. Since the method characterizes the input informations utilizing an all-pole theoretical account, the right pick of the theoretical account order P is of import. We can non take the value of theoretical account order excessively big or excessively little as it gives hapless appraisal of PSDs. We can pattern any stochastic procedure utilizing AR theoretical account. The spectrum of the stochastic procedure can be given as

There are assorted methods available for AR patterning such as traveling norm ( MA ) theoretical account, autoregressive traveling norm ( ARMA ) theoretical account, Burg ‘s algorithm [ 24 ] . ARMA method of AR theoretical account is usually used to acquire good truth. Burg algorithm estimates the contemplation coefficient Alaska. we can utilize Burg method to suit a pth order autoregressive ( AR ) theoretical account to the input signal, x, by minimising ( least squares ) the forward and backward anticipation mistakes while restraining the AR parametric quantities to fulfill the Levinson-Durbin recursion [ 25 ] . The Burg method is a recursive procedure.

In this paper we have followed the Burg ‘s method to happen the AR coefficients. The theoretical account order is taken to be equal to 10. We have used the Burg algorithm to happen the AR coefficients utilizing MATLAB.

AR coefficients and the Power spectral densitywere obtained by utilizing MATLAB. Since the theoretical account order is 10 we have 11 AR coefficients. The secret plan for the power spectral denseness is shown in the undermentioned figures:

powAfig.bmp

( a )

powDfig.bmp

( B )

powEfig.bmp

( degree Celsius )

Figure 2.5 ( a ) , ( B ) , ( degree Celsius ) Plot for power spectral denseness of category A, D and E severally

2.4 Feature Vector

The 20 distinct ripple coefficients and 11 Auto-regressive coefficients were appended to organize characteristic vector of dimension 31. These characteristic coefficients are shown as follows:

feature_a.bmp

( a )

feature_d.bmp

( B )

feature_e.bmp

( degree Celsius )

Figure 2.6 ( a ) , ( B ) , ( degree Celsius ) Feature vector of dimension 31 of category A, D and E severally

Greenwich Mean Time

s usage 64 or 128 electrodes ) , we can acquire some thought of where the ERP constituents are strongest. This does n’t truly

2.5 Artificial Neural Network

An unreal nervous web can be defined as a machine that is modelled on a human encephalon. The cardinal structural components of the encephalon are nerve cells which are besides the basic information treating units of an ANN. The nervous web is formed by a monolithic interconnectedness of these nerve cells. The web so formed has the capableness of larning i.e geting cognition from the environment by executing calculations. The synaptic weights which are the interneuron connexion strengths are used to hive away this acquired cognition. In the acquisition procedure synaptic weights can modified harmonizing to many algorithms to accomplish the coveted design aim. Fig 2.7 shows the theoretical account of a individual nerve cell [ 18 ] .

aneu

Figure 2.7: Model of a nerve cell

A typical nervous web consists of the undermentioned beds

1. Input bed

2. Hidden bed

3. Output bed

The input bed consists of beginning nodes which supply the input vector ( activation form ) i.e the input signals to the following bed.

A nervous web can hold one or more concealed bed. This bed consists of concealed nerve cells. These nerve cells are important to execute certain higher order statistics and calculations to execute a specific undertaking. They intervene between the input and end product bed in some utile mode. The ability of concealed nerve cells to pull out higher order statistics is valuable when the size of the input bed is big. The end product signals of a peculiar bed act supplied to the following bed.

The end product bed is the last bed of a web construction. The set of end product signals of the nerve cells in the end product bed is gives the overall response of the web to the input vector supplied at the input bed. Fig 2.8 shows a typical web construction with one hidden bed. This web has 4 nerve cells in the input bed, three nerve cells in the concealed bed and a individual out neuron. A web is said to be to the full connected when each node in each bed of the web is connected to every other node in the next forward bed, otherwise if some connexions are losing so is said to be partly connected.

MLP

Figure 2.8: Structure of a nervous web

2.6 Learning Procedure

Learning is a procedure in which the nervous web undergoes alterations in its free parametric quantities when it is stimulated by the environment. As a consequence of this larning it construction alterations and it responds in a new manner to its environment. Gradually its public presentation improves through this procedure. There are many types of acquisition regulations, some of it are mentioned below [ 18 ] .

Error- rectification acquisition in which the mistake signal actuates a control mechanism so as to do accommodations in to the synaptic weights. These alterations make the end product signal come closer to the mark value in a measure by measure mode. The mistake signal is the difference between the desired end product and the end product from the web. The aim in this type of acquisition is to minimise the cost map or the index of public presentation. The cost map is the instantaneous value of the mistake energy.

Memory based learning- here all the past values of right classified input-output illustrations are stored. When a new trial form is applied to the web this acquisition algorithm responds by recovering and analysing the preparation informations in the local vicinity of the trial form. Nearest neighbour regulation and K-nearest classifier are two popular algorithm in this type of acquisition.

Hebbian learning-It is the oldest and most celebrated of all acquisition regulations. It is based on hebbian synapse which is defined as a synapse with time-dependent, extremely local, and strongly synergistic mechanism to increase synaptic efficiency as a map of the correlativity between presynaptic and post synaptic activities. [ Brown et al.,1990 ] . In other words if the nerve cells on either side of a synapse are at the same time activated so the strength of that synapse is selectively increased and if activated asynchronously so the synapse is selectively weakened or eliminated [ Stent,1973 ; Changeux and Danchin 1976 ] .

Competitive learning-As the name implies the end product nerve cells of the web compete among themselves to acquire become active. At a clip merely one nerve cell is activated. The set of nerve cells are all same but for some indiscriminately distributed synaptic weights, there is a mechanism in topographic point so that for the given input pattern merely one nerve cell is fired i.e. the nerve cell that wins the competition is called a winner-takes-all-neuron [ Rumel Hart and Zipser 1985 ] . So this regulation is suited for characteristic sensing and pattern acknowledgment intents.

2.7 MLPNN and Back Propagation algorithm

The multilayer perceptron is the most popular and normally used nervous web construction. It is an extension of the individual bed perceptron. Basically an MPLNN consists of a set of beginning nodes called the input bed, one or more bed of concealed nerve cells and an end product bed. These type of webs have been used to work out many pattern acknowledgment jobs by developing them in a supervised mode by utilizing a extremely popular algorithm based on mistake rectification regulation called the mistake back extension algorithm.

The back extension algorithm is based on delta regulation and gradient descent of mistake surface in weight infinite. Harmonizing to delta govern the synaptic weight alteration of a nerve cell is relative to the larning rate parametric quantity and the gradient of the cost map at the peculiar weight in multidimensional weight infinite. Basically this algorithm consists of two base on ballss – forward and backward through different beds of the web. In the forward base on balls an input signal is applied to the input bed. This signal is propagated in the forward way bed by bed by executing calculations at each and every node. In this base on balls the synaptic weight remain unchanged. At the end product bed we get a response for each activity form applied. In the backward base on balls the mistake signal is computed as the difference between the mark value and end product value. This signal is responsible for alterations in weights bed by bed harmonizing to the delta regulation so that the response of the web moves closer to the desired response in a statistical sense.

The stairss involved in the back extension algorithm are given below [ 18 ] .

Initialization- the synaptic weights and prejudices are given random values which are picked from a unvarying distribution with nothing mean.

Presentation of input patterns- the web is nowadayss with input forms which act as preparation vectors. These forms are used to calculate the forward base on balls and so the backward base on balls.

Forward pass- Let the input and mark of a preparation illustration is ( ten ( n ) , d ( n ) ) , the induced local field vj ( cubic decimeter ) ( n ) can be formulated as below

The end product signal of neuron J in bed cubic decimeter can be given as below

If the nerve cell J is in the first hidden bed ( i.e. cubic decimeter = 1 )

If the nerve cell J is in the end product bed ( i.e. cubic decimeter = L )

Mistake can be computed as

Backward pass- The local gradient ( ?s ) can be computed by

Where ? ‘ ( . ) denotes the distinction with regard to the statement. Now adjust the synaptic weights utilizing the generalised regulation

Iteration- Now we can repeat the forward and backward calculation.

2.8 Committee Neural Network

Committee nervous web is an attack that reaps the benefits of its single members. It has a parallel construction that produces a concluding end product [ 18-19 ] by uniting consequences of its member nervous webs. In the present survey the proposed technique consists of 3 stairss ( 1 ) choice of appropriate inputs for the single member of the commission ( 2 ) preparation of each member ( 3 ) determination doing based on bulk sentiment.

The commission web consists of member nervous webs which are multi bed perceptron nervous web trained with back extension algorithm. The available information is divided into preparation and proving informations. From the preparation informations characteristics were extracted utilizing ripple transform and AR coefficients. The input characteristic set is divided every bit among all the nervous webs for developing intent. The different webs have different nerve cells and initial weights. After the preparation stage is completed the webs are tested with proving informations. All the nervous webs were trained utilizing gradient descent back extension algorithm utilizing MATLAB package bundle. Out of the different webs employed for the initial preparation phase the best acting webs were selected to organize the commission. For the categorization purpose the bulk determination of the commission formed the concluding end product.Fig 2.9 shows the block diagram of the commission

Untitled

Figure 2.9: Block diagram of Committee nervous web [ 17 ]

2.9 Classification utilizing Committee Neural Network

The commission nervous web was formed by three independent members each trained with different characteristic sets. Prior to recruitment in the commission many webs incorporating different concealed nerve cells and initial weights were trained and the best acting three were selected. The determination merger was obtained utilizing bulk vote. In order to make a steadfast bulk determination odd figure of webs is required. The ensuing truth of single and the commission is shown in table 2.9.

Table 2.2 Confusion matrix of Neural Network 1

Classifier

Desired consequence

Output consequence

Put A

Set D

NN 1

Put A

612

32

Set D

26

577

Set E

2

31

Table 2.3 Confusion matrix of Neural Network 2

Classifier

Desired consequence

Output consequence

Put A

Set D

NN 2

Put A

621

27

Set D

14

586

Set E

5

27

Table 2.4 Confusion matrix of Neural Network 3

Classifier

Desired consequence

Output consequence

Put A

Set D

NN 3

Put A

628

60

Set D

11

538

Set E

1

42

Table 2.5 Statistical parametric quantities of Neural Network 1

Statistical Parameters

Valuess ( % )

Specificity

95.63

Sensitivity ( ictus free epileptogenic zone sections )

90.16

Sensitivity ( epileptic ictus section )

94.06

Entire categorization truth

93.28

Table 2.6 Statistical parametric quantities of Neural Network 2

Statistical Parameters

Valuess ( % )

Specificity

97.03

Sensitivity ( ictus free epileptogenic zone sections )

91.56

Sensitivity ( epileptic ictus section )

95.16

Entire categorization truth

94.52

Table 2.7 Statistical parametric quantities of Neural Network 3

Statistical Parameters

Valuess ( % )

Specificity

98.13

Sensitivity ( ictus free epileptogenic zone sections )

84.06

Sensitivity ( epileptic ictus section )

96.09

Entire categorization truth

92.76

Table 2.8 Statistical parametric quantities of Committee Neural Network

Statistical Parameters

Valuess ( % )

Specificity

98.02

Sensitivity ( ictus free epileptogenic zone sections )

91.82

Sensitivity ( epileptic ictus section )

96.09

Entire categorization truth

95.31

Table 2.9 Accuracy of Committee nervous web

ANN

Accuracy

NN1

93.28

NN2

94.52

NN3

92.76

CNN

95.31

Chapter 3

Dimension Reduction utilizing F-Ratio based Technique

3.1 F-Ratio

F-Ratio is a statistical step which is used in the comparing of statistical theoretical accounts that have been fit to informations set to place the theoretical account that best fits the population from which the informations were sampled [ 21 ] . We can see a multi bunch informations as shown in fig 3.1. F-ratio can be formulated as

Figure 3.1: Diagram for multi-cluster informations

Suppose there are thousand Numberss of bunchs each holding n figure of informations points. If xij is an ith component of the jth category so the mean of the jth category µj can be expressed as [ 26 ]

The mean of all µj is called the planetary mean of the informations and can be expressed as µ0

The f-ratio can be expressed as [ 26 ]

If the f-ratio additions so the bunchs move off from each other or the bunch size psychiatrists. We can use this f-ratio based optimisation technique in instance of EEG signals to cut down the dimensionality of the characteristic vector.

3.2 Optimization utilizing F-Ratio

The characteristics holding low value of F-Ratio are less of import as compared to the characteristics holding high value of F-Ratio. To cut down the computational complexness of the web we can cancel the characteristics holding lesser values of F-ratio. By canceling these characteristics the truth of categorization does non diminish much. In this paper we have deleted characteristics one by one and each clip we have analyzed the categorization truth at the same time. If the categorization truth is cut downing by more than 0.5 % anyplace we did non cancel that peculiar characteristic. Finally the difference between the categorization truth utilizing all the characteristics and the categorization truth after canceling the appropriate characteristics should non be greater than 0.5 % . In this manner we can cut down the characteristic dimension and hence we can optimise the web without impacting the categorization truth. In some instances the truth was found to increase on omission of characteristics.

3.3 F-Ratio of the extracted characteristics

The F-Ratio corresponding to each characteristic is shown in table 3.1.

Table 3.1 F-Ratio of the extracted characteristics

Consecutive No.

Coefficient No.

Coefficients F-ratio

Consecutive No.

Coefficient No.

Coefficients F-ratio

1

21

1.082

17

1

0.226

2

22

0.9316

18

13

0.1972

3

12

0.5243

19

14

0.1958

4

8

0.4748

20

31

0.1634

5

24

0.4714

21

28

0.1555

6

9

0.4464

22

27

0.0459

7

6

0.4073

23

20

0.0444

8

10

0.401

24

26

0.0407

9

23

0.3981

25

17

0.0036

10

4

0.3915

26

18

0.0023

11

25

0.3708

27

19

0.0011

12

5

0.3699

28

11

0.0004

13

30

0.365

29

15

0.0003

14

16

0.301

30

3

0.0002

15

29

0.2875

31

7

0.0001

16

2

0.2766

ten

ten

ten

Now in order to cut down the dimension of the characteristic vector features holding less F-Ratio were deleted. In the table 3.2 we have shown that the truth after the omission of characteristics.

Consecutive No.

No. of coefficients taken

Network construction

Accuracy %

1

31

31-93-3

95.31

2

30

30-90-3

94.91

3

29

29-87-3

95.00

4

28

28-84-3

94.71

5

27

27-81-3

94.79

6

26

26-78-3

95.03

7

25

25-75-3

95.32

8

24

24-72-3

95.02

9

23

23-69-3

95.12

10

22

22-66-3

95.37

11

21

21-63-3

94.95

12

20

20-60-3

95.16

13

19

19-57-3

95.45

14

18

18-54-3

95.29

15

17

17-51-3

95.70

16

16

16-48-3

95.83

17

15

15-45-3

95.15

18

14

14-42-3

95.04

19

13

13-39-3

94.35

Table 3.2 dimension decrease utilizing F-Ratio

Hence on the footing of F-Ratio based optimisation technique 18 characteristics were deleted. Thus computational complexness was reduced without impacting the truth much.

Chapter 4

Decisions and Future Work

The EEG signals was collected from [ 9 ] , ocular review of the three categories does non supply much information sing the wellness of single. So we have proceeded with the undermentioned methods

Feature extraction was done utilizing distinct ripple transform and the power spectral denseness was estimated by Burg ‘s algorithm for AR theoretical account.

To cut down the figure of wavelet coefficients we have used the statistics, viz. upper limit, lower limit, mean, and standard divergence for each of the item and the approximative coefficients.

The AR coefficients were appended to the distinct ripple coefficients to organize the characteristic vector.

We have used Committee nervous web for the categorization intent. The Committee nervous web consisted of three member nervous webs which were trained utilizing mistake back extension algorithm.

The Committee nervous web gives robust public presentation as compared to the single webs.

We have used the F-Ratio based optimisation technique to cut down the dimension of characteristic vector.

Finally utilizing our proposed technique we have successfully classified the EEG signals and reduced the computational complexness of the classifier.

EEG signal processing promises to be a huge country of research. The technique of commission nervous web is a fresh attack to better the categorization truth. The procedure of uniting the end products of each member from the commission implemented in this undertaking was based on bulk determination. There are many new methodological analysiss that can be implemented in this country. Further different types of classifiers can be tested utilizing different EEG database. Other feature extraction techniques can be tried so that a best possible set of characteristics can be used besides to cut down the computational complexness other dimensionality decrease techniques can be applied to the characteristic vectors.

Post Author: admin