Analysis of Movement Related EEG Signal by Time Dependent Fractal Dimension and
Neural Network for Brain Computer Interface
NI NI SOE (D3)
Fractal and Chaos Informatics Laboratory (NLAB)
Background and Purpose
• A brain computer interface is a system that can learn to recognize patterns of increased activity in local areas of the brain.
• An Electroencephalogram (EEG) based Brain-Computer-Interface (BCI) to provides a new communication channel between the human brain and a computer.
• To classify stochastic, non-stationary, self-similar signals which originate from non-linear systems and may be comprised of multiple signals, using independent component based blind signal separation, a time dependent fractal dimension analysis and neural networks.
Introduction
• A brain computer interface is designed to recognize patterns in data extracted from the brain and associate the patterns with commands.
• Very often these patterns, or states, are referred to as thoughts, and accordingly, systems that rely on BCI techniques for input are described as being thought controlled.
Block diagram of the system
PreprocessingFeature
Extraction by FD(CEM)
Classification(Detection)
SignalAcquisition
ApplicationInterface
Brain Computer Interface
Signal Processing
Feedback
EEG Signal Acquisition Module
Where to place the electrodes?
• The electrodes are placed according to the international standard 10-20 system.
• The electrode position of 16 channels placed as shown in Fig. of frontal area (F3 and F4), central area (C3, Cz and C4), parietal area(P3, Pz and P4)
• The grounding electrode and referencing electrode are placed at forehead and right ear lobe respectively.
• EEG signal are digitized at 1024 samples/sec, resolution 16bit/sample.
• Signal were analog bandpass filtered between 1.5 and 100 Hz.
GND
GND
REF
T5
P3 PZ P4
T4T3 C4
F3 F4F8F7
FP1 FP2
O2O1
T6
FZ
CZC3
NASION
Preprocessing ModuleEEG raw data
Common AverageReferencing
Local LaplacianFiltering
REREFERENCING
Center data
Whiten data
Preprocessingbefore ICA
ICA Source Separation andArtifact Removing
Preprocessing
Only apply C3 and C4electrode position because of
four neighboring electrode
To make the data a zeromean variable
linearly transformed so thatthe components areuncorrelated and has unitvariance. Use (EVD).
Fast ICA method
j
1ˆ
s = source signal
s = CAR filtered signal ( j=i=1,2,...16 )
j j ii s
s s ss
Feature Extraction Module
• Feature extraction is by Fractal dimension.
• What is Fractal?– A fractal is defined as a set for which
Hausdorff-Besicovich dimension is strictly greater than the topological dimension.
– Fractal dimension is defining property in the study of textual analysis.
– We use fractal dimension as feature extraction.
Feature Extraction of EEG signal by CEM based Fractal dimension
• The method proposed here is based on fractal dimension by CEM.
• Let, the αth momment is Iα
• Where, P(u) = power spectral density,• u= normalized frequency whose lower cut off is 1.• Assume,
P(u)~u-β
(-<α <+ )
Feature Extraction of EEG signal by CEM based Fractal dimension(cont.)
• Then,
• Where,
X=α-β+1, υ= logU• Then, αc , the critical value can get by satisfying the following
equation
1
1 1
1 2( 1) exp( )sinh( )
2 2
U UX X X X
I u du u du UX X
33 3
3 3
1 1log 2 2 cos ( )cosh( ) 0, ( 0)
8 2 2
d X XI ech X
d X
β=αc+1=2H+1 , D=2-H=2-αc/2 , D is the fractal dimension
Classification of EEG by Probabilistic Neural Networks
Class Category
1 Right hand
2 Left hand
3 Foot
Summing layer
Pattern layer
Output layer
Input Layer
CH 1
CH 16
Max
.
.
.
1
2
3
An Introduction to Probabilistic NeuralNetworks
• A probabilistic neural network (PNN) is predominantly a classifier– Map any input pattern to a number of classifications– Can be forced into a more general function approximator
• A PNN is an implementation of a statistical algorithm called kernel discriminant analysis in which the operations are organized into a multilayered feed forward network with four layers:– Input layer– Pattern layer– Summation layer– Output layer
• Actual and Imaginary movement of two subjects.
• 3 tasks• 7 trials for each tack• 1 trial took 30sec
• 5 trials is for training pattern and 2 is for testing pattern. We also tried the following sequence of task for testing our PNN.
Experimental setup
Experimental result of CEM
TDFD
1
1.2
1.4
1.6
1.8
2
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96 101 106 111
no. of observed window
Fra
cta
l D
imen
sio
n
Experimental result of PNN
• The following table shows the result of the imaginary tasks of subject 1.
Tasks Classification rate atσ=0.05 (%)
Classification rate atσ=0.025 (%)
Foot 97.463 99.154
Right hand 97.88 94.50
Left hand 99.365 99.154
Window size = 2048 , step = 256
• The long term goal of this research is to apply the BCI system in controlling of environmental navigation system of Humanoid robot.
• The experimental results show that it is possible to recognize quite reliably ongoing mental movement imaginary in the application area of humanoid robot control.
• In future, we are going to utilize Lyapunov exponent and chao neuron mpdel for classification.
Discussion and future works
Conclusions
• we have demonstrated fractal based TDFD based feature extraction and PNN classifier.
• The results of the experiment indicate that the EEG signals do contain extractable and classifiable information about the performed movements, during both physical and imagined movements.
• The experimental results show that the proposed classification systems are capable of classifying non-stationary, self-similar signals, such as EEG signal, with average accuracies up to 90%.