A Brain Computer Interface

Think. Done!
A Brain Computer Interface (BCI)
By
Muhammad Bilal Khalid

Intisar Rizwan-i-Haque
Sarmad Munir
Farhan Tahir
Submitted to the Faculty of Electrical Engineering Department,
National University of Sciences and Technology, Rawalpindi in partial fulfillment for
the requirements of a B.E Degree in Telecommunication Engineering
August 2009
Think. Done! A Brain Computer Interface (BCI)

CERTIFICATE
Certified that the contents and form of project report entitled “Think. Done! A Brain
Computer Interface (BCI)” submitted by 1) Muhammad Bilal Khalid, 2) Intisar Rizwan-i-
Haque, 3) Sarmad Munir and 4) Farhan Tahir have been found satisfactory for the
requirement of the degree.
Supervisor: ____________________
Maj. Dr. Naveed Iqbal Rao
Military College of Signals,
NUST, Rawalpindi.

ABSTRACT
A Brain Computer Interface (BCI) is a system that provides an individual a non-muscular
way to control his surroundings by working on the activity signals collected from his
brain. These signals are in the form of Electroencephalogram (EEG).
We have developed a complete a BCI system from signal acquisition to control interface.
The system we have designed lets a user control a cursor on computer screen and play a
video game (car game developed in Java) without any physical movement. Another and
most interesting interface is with a toy car through parallel port.
For signal acquisition a complete five channel (including one for reference) EEG machine
has been designed. This machine is light weight, low cost and ultra portable.
Windowing and time domain normalization have been used for preprocessing. Wavelet
Transform (WT) and Fractional Fourier Transform (FRFT) have been used to extract
feature vectors. Neural networks are used as classifiers. Finally heuristic weight
adjustment of networks has been introduced to improve classification accuracy.
All the algorithms used have been published in form of two research papers in
international conferences [23], [24].

DEDICATION
In the name of Allah, the Most Gracious, the Most Merciful
To
Our Beloved Parents and Siblings
&
Our Respectful Teachers

ACKNOWLEDGEMENTS
We would like to take this opportunity to pay our humble gratitude to Almighty Allah
(S.W.A.T), the most merciful.
We are extremely thankful to our Project Supervisor Dr. Naveed Iqbal Rao who
supervised the project in a very encouraging and helpful manner. This project has been
completed with funding from Higher Education Commission (HEC) Pakistan at Image
Processing Center (IPC).
Following people have been very helpful at different stages of this project;
Mr. Kashif Siraj Research Associate at Image Processing Center MCS, Dr Abdul Ghafoor EE
Dept. MCS, Mr. Athar Mohsin Zaidi of CS Dept.MCS, Brig. Mowaddat Rana of MH
Rawalpindi, Ammad Junaid of CS Dept. and Image Processing Center staff.

Contents
DEDICATION ..................................................................................................................... 3
In the name of Allah, the Most Gracious, the Most Merciful ........................................... 3
Our Respectful Teachers..................................................................................................... 3
1. INTRODUCTION ........................................................................................................ 10
1.1 Motivation for this project ..................................................................................... 10
2. Introduction to Brain Computer Interface (BCI) ................................................... 14
2.1 History ......................................................................................................................15
2.2 Classifications ..........................................................................................................15
2.2.1 Synchronous Vs. Asynchronous 15
2.2.2 Online Vs. Offline 16
2.2.3 Invasive Vs. Non-invasive 16
2.3 Structure of Brain ................................................................................................... 16
2.3.1 Cerebrum 17
2.3.2 Cerebellum 17
2.3.3 Brain Stem ........................................................................................................... 18
2.4 Motor Cortex .......................................................................................................... 18
2.5 Electroencephalogram (EEG) .................................................................................... 20
2.6 Modules/Blocks of BCI system .................................................................................. 20
3. Literature Review .................................................................................................... 22
3.1 Algorithms for preprocessing ................................................................................ 22
3.2 Algorithms for feature extraction ......................................................................... 22
3.2.1 Wavelet transform ............................................................................................... 22
3.2.2 Fractional Fourier transform .............................................................................. 23
3.3 Classification ........................................................................................................... 23
4. What we have done in each module ...................................................................... 24
4.1 Signal Acquisition ....................................................................................................... 24
4.1.1 Online 24
EEG MACHINE HARDWARE DESIGN ........................................................................... 25
Basic elements ............................................................................................................. 25
INPUT PROTECTION RESISTANCE 25
Defining the Ideal Operational Amplifier 27
A Summing Point Restraint 28
The Desirability of Feedback 29
Two Important Feedback Circuits 30

Voltage Follower 31
Non-Inverting Amplifier 33
Inverting Amplifier 34
Integrator 36
Voltage Adder 36
Low Pass Filter 38
High Pass Filter 40
INSTRUMENTATION AMPLIFIER 40
Signal Amplification and Common-Mode Rejection 41
Common-Mode Rejection: Op Amp vs. In-Amp 44
High AC (and DC) Common-Mode Rejection 47
Low Offset Voltage and Offset Voltage Drift 47
A Matched High Input Impedance 48
Low Input Bias and Offset Current Errors 48
Low Noise 49
Low Nonlinearity 49
Simple Gain Selection 49
Adequate Bandwidth 49
Differential to Single-Ended Conversion 50
Rail-to-Rail Input and Output Swing 50
Power vs. Bandwidth, Slew Rate, and Noise 50
Modulation 52
Carrier Generation 54
4.1.2 Offline 55
4.2 Preprocessing............................................................................................................. 56
4.2.1 Windowing ........................................................................................................... 56
4.2.2 Time domain normalization ................................................................................ 57
4.3 Feature Extraction ..................................................................................................... 57
4.3.1 Wavelet Transform (WT)..................................................................................... 57
4.3.2 Fractional Fourier Transform............................................................................. 60
Interpretation of the Fractional Fourier Transform................................................... 61
4.4 Classification .............................................................................................................. 63
4.4.1 Neural networks as classifiers ............................................................................ 63
4.4.2 Heuristic Weight Adjustment for Final Decision Making ...............................64
4.5 Control Interface ........................................................................................................64
4.5.1 Cursor control ......................................................................................................64

4.5.2 Car game in Java ................................................................................................. 65
4.5.3 Toy car control ................................................................................................... 66
5. Applications ..............................................................................................................69
EEG Bio feedback ........................................................................................................ 69
Computer Control and Communications................................................................... 69
Entertainment, Virtual Reality .................................................................................... 69
Education and Research ............................................................................................. 69
Military Commercial .................................................................................................... 69
5.1 EEG Biofeedback ..................................................................................................... 70
5.2 Computer Control & Communication ................................................................... 70
5.3 Entertainment, Virtual Reality ................................................................................71
5.4 Education, Research ...............................................................................................71
5.5 Military, Commercial ............................................................................................. 72
6. References ................................................................................................................ 73

LIST OF FIGURES
Figure 1: Block Diagram of BCI system
Figure 2: Our brain—part wise distribution
Figure 3: Motor cortex region of brain
Figure 4: Modules of BCI system
Figure 5: Our EEG machine—bird eye view
Figure 6: OPAMP—circuit diagram
Figure 7: Open loop OPAMP
Figure 8: Feedback Circuit
Figure 9: Voltage Follower
Figure 10: Voltage Follower with feedback
Figure 11: Non-inverting AMP
Figure 12: Inverting AMP
Figure 13: Integrator
Figure 14: Voltage summer
Figure 15: Response Curve
Figure 16: Response Curve 2
Figure 17: Response Curve 3
Figure 18: LPF
Figure 19: Internal structure AD620
Figure 20: Pin configuration AD620

Figure 21: Pin configuration AD633
Figure 22: Circuit diagram AD 633
Figure 23: Carrier Generator
Figure 24: Time frequency axis
Figure 25:Car game in Java
Figure 26:Toy car and its interface
Figure 27:Parallel port configuration

1. INTRODUCTION
1.1 Motivation for this project
Interaction with any machine requires a very smooth and efficient interface between the
machine and its users. Humans using computers are familiar with a number of interfaces
such as keyboard, mouse, touch screens etc. All these interfaces have one thing in
common: they require physical movement by the user. This physical/muscular
movement may not be possible in some situations such as Virtual Reality environment,
daily routine works for physically disabled people etc. A Brain Computer Interface (BCI) is
such an interface in which a person uses his/her brain to control the machine to be used,
this machine can be a computer, a wheel chair or anything.
Brain Computer Interfacing is an exciting new technology. It provides an individual a non-
muscular way to control his surroundings. It enables the individual using it to relay
command signals directly to the machine instead of first relaying it to a body part such as
hands or feet. Its applications range from empowering physically paralyzed people (and
normal people too) to use computers without moving a single muscle, to applications in
Virtual Reality. Due to its tremendous potential it is attracting huge investments and
research activities from around the world. World leading universities are establishing
‘Brain Computer Interface Centers’ to accelerate development in this field.

Brain Computer Interfacing or sometimes called Man Machine Interfacing is the
technology of future. Imagine being able to navigate through your mobile phone using
your mind... Imagine driving your car by just thinking about it!! Yes Brain Computer
Interfacing promises all these exciting rewards.
This document describes the work done by this group in this field. We have developed an
asynchronous real time complete BCI system that enables a user to control a cursor on
screen, move a toy car and play a video game, all in real time with a reasonable accuracy.
Since this is a project report, it does not serve the purpose of defining BCI terms.
However some terms are explained as required by the context. Definitions flow with the
work we have done, i.e. the context remains the work we have done in this field.
1.2 How can this be accomplished?
This task, as the reader might have anticipated, is not a trivial one. Like the activity
signals generated by our heart, which are called ECG signals, similarly our brain also
generates its own activity signals Electroencephalogram (EEG) signals(it actually
generates a lot many other signals too but we are interested in EEG only), each time it
processes something. These activity signals are captured in the form of micro-volt level
signals. More about EEG signals is presented is following sections.

Each time we do something or we ‘think about doing something’, our brain generates
distinct signals. Each activity by us generates corresponding signals on our brain, so by
reading these signals one can guess what the person is thinking at the moment!!
Though these signals are highly random, but still one can find some pattern in them, i.e.
the pattern corresponding to the activity the user is doing or just ‘thinking about doing’,
like moving his arm or leg. Finding this pattern is the main task of any BCI. This requires
some very smart classification algorithm which once trained can work for a particular
person in a particular domain of thought.
A BCI picks these signals from the brain of user in the form of EEG. Then after classifying
what the user is thinking, it translates these signals into meaningful commands, e.g. it
translates thoughts about hand movements in to corresponding cursor movements.
Figure1: Block diagram of BCI system
1.3 What we have accomplished

We have developed a complete BCI system from signal acquisition to control interface.
Signal acquisition module is a light weight, low cost and ultra portable EEG machine
developed by this group.
All the algorithms we have implemented are novel and more efficient than previous
ones. These algorithms have been published by us in international conferences. Detail of
these algorithms is presented in following sections.
We have interfaced our system to three applications. First one is a cursor control system
which enables an individual to control a cursor in four directions in real time. Second
interface is a video game which is a car game in Java. And third and the most interesting
one is a toy car which is interfaced to a PC through a parallel port.
This project has won numerous national level competitions including NESCON 09 at FAST
Islamabad, NUTEC 09 at FAST Peshawar and COMPPEC 09 at College of EME Rawalpindi.
One of our publications also won second best paper award in category at IC-4 conference
at Karachi.

2. Introduction to Brain Computer Interface (BCI)
A Brain Computer Interface (BCI) is a system that enables an individual to control his /her
surroundings without any physical movements. This is accomplished by working on the
signals acquired directly from the brain in the form of Electroencephalogram (EEG)
signals.
The technology inside an EEG-based communication device is comprised of four main
components; a signal capture system, a signal processing system, a pattern recognition
system, and a device control system. Nowadays, these systems can simply be four
different software modules running on a fast PC.
The signal capture system includes the electrodes themselves, the isolated electronic
amplifiers (they're isolated from the mains so that the only electricity that flows is from
the subjects head to the computer and NOT vice versa !) and possibly a Digital Signal
Processing (DSP) board. The signal processing system often used to be on a dedicated
DSP board but now PCs are fast enough to do everything on the main processor. The
pattern recognition system often used to be composed of a linear classifier such as a
logistic discriminant or a classical nonlinear classifier such as the Bayes quadratic
classifier or linear vector quantiser (LVQ). Nowadays, neural networks are most
commonly used. Again, although specialist neural net chips are available, most PCs have
enough power to do everything in the main CPU.

2.1 History
Research on BCIs began in the 1970s at the University of California Los Angeles (UCLA)
under a grant from the National Science Foundation followed by a contract from DARPA.
These papers also mark the first appearance of the expression brain–computer interface
in the scientific literature.
The field has since blossomed spectacularly, mostly toward neuroprosthetics applications
that aim at restoring damaged hearing, sight and movement. Thanks to the remarkable
cortical plasticity of the brain, signals from implanted prostheses can, after adaptation,
be handled by the brain like natural sensor or effecter channels. Following years of
animal experimentation, the first neuroprosthetic devices implanted in humans
appeared in the mid-nineties.
2.2 Classifications
There are many classifications of BCI systems. These classifications can be based on signal
acquisition, sensor placement and timing etc. Some classifications are listed here;
2.2.1 Synchronous Vs. Asynchronous

Synchronous BCIs can detect user activity automatically and take action accordingly. User
is not required to follow cues by system. These systems are more user friendly but are
much more difficult to build. Main issues fuelling this difficulty are sensitivity of the
sensors and efficiency of algorithm to detect between meaningful activity and gibberish.
Asynchronous BCIs on the other hand require system cues to function. In this
arrangement system provides the user regular cues to perform activity. System only

responds to activity during cue time and not otherwise. The system we have developed is
a synchronous BCI.
2.2.2 Online Vs. Offline

Online systems work in real time. For this they need to connect to some signal
acquisition machine all the time. We have developed our own EEG machine for this
purpose, the details of which can be found in corresponding chapter.
Offline analysis is performed on pre-recorded data. This is done to test efficiency of
algorithms. There are a number of sources available online.
2.2.3 Invasive Vs. Non-invasive

Electrodes (sensors) may be placed on the skin on scalp or can be dug in to the skin.
When placed on the skin surface, with the help of a conducting gel, the signals picked are
relatively weak. When implanted with in the skin, they provide better signals and more
accuracy. System described in this report is non-invasive.
With improvements in technology non-invasive techniques are becoming more and more
efficient. However high accuracy BCIs use invasive methods.
2.3 Structure of Brain

At this stage it is imperative to take a look at the structure of brain, the very organ which
makes this whole thing possible. Our brain is perhaps the most complex machine in this
world, and to describe it in a few pages is not possible. Here I will only provide
information necessary to understanding to BCI systems. Also covered is the topic of EEG.

Brain has three parts
• Cerebrum
• Cerebellum
• Brain Stem
2.3.1 Cerebrum
The cerebrum or cortex is the largest part of the human brain, associated with higher
brain function such as thought and action. The cerebral cortex is divided into four
sections, called "lobes": the frontal lobe, parietal lobe, occipital lobe, and temporal lobe
• Frontal Lobe- associated with reasoning, planning, parts of speech, movement,
emotions, and problem solving
• Parietal Lobe- associated with movement, orientation, recognition, perception of
stimuli
• Occipital Lobe- associated with visual processing
• Temporal Lobe- associated with perception and recognition of auditory stimuli,
memory, and speech
A deep furrow divides the cerebrum into two halves, known as the left and right
hemispheres. The two hemispheres look mostly symmetrical yet it has been shown that
each side functions slightly different than the other. Sometimes the right hemisphere is
associated with creativity and the left hemispheres is associated with logic abilities.
The corpus callosum is a bundle of axons which connects these two hemispheres.
2.3.2 Cerebellum
The cerebellum, or "little brain", is similar to the cerebrum in that it has two hemispheres
and has a highly folded surface or cortex. This structure is associated with regulation and
coordination of movement, posture, and balance.

2.3.3 Brain Stem
This structure is responsible for basic vital life functions such as breathing, heartbeat, and
blood pressure. Scientists say that this is the "simplest" part of human brains because
animals' entire brains, such as reptiles (who appear early on the evolutionary scale)
resemble our brain stem.
The brain stem is made of the midbrain, pons, and medulla.
2.4 Motor Cortex

Movement related data is picked from the motor cortex.
All of the body's voluntary movements are controlled by the brain. One of the brain areas
most involved in controlling these voluntary movements is the motor cortex. The motor
cortex is located in the rear portion of the frontal lobe, just before the central sulcus
(furrow) that separates the frontal lobe from the parietal lobe. The motor cortex is
divided into two main areas, Area 4 and Area 6. Area 4, also known as the primary
motor cortex, forms a thin band along the central sulcus. Area 6 lies immediately forward
of Area 4. Area 6 is wider and is further subdivided into two distinct sub-areas.
Figure2: Our brain

In 1870, Hitzig and Fritsch electrically stimulated various parts of a dog's motor cortex.
They observed that depending on what part of the cortex they stimulated, a different
part of the body contracted. Then they found that if they destroyed this same small area
of the cortex, the corresponding part of the body became paralyzed. This is how it was
discovered that every part of the body has a particular region of the primary motor
cortex that controls its movement.
But what is remarkable about this motor map is that certain parts of the body—those
that can make the finest movements—take up much more space than others. These
parts of the body are shown larger than the others in the illustration here.
Figure3: Motor Cortex

2.5 Electroencephalogram (EEG)
The EEG is recorded between electrodes placed in standard positions on the scalp and
has a typical amplitude of 2-100 microvolts and a frequency spectrum from 0.1 to 60 Hz.
Most activity occurs within the following frequency bands; delta (0.5 - 4 Hz), theta (4-8
Hz), alpha (8-13 Hz), beta (13-22 Hz) and gamma (30-40 Hz).
The potential at the scalp derives from electrical activity of large synchronised groups of
neurons inside the brain. The activity of single neurons or small groups is attenuated too
much by the skull and scalp to be detected at the scalp surface
EEG activity in particular frequency bands is often correlated with particular cognitive
states. Signals in the alpha band, for example, are associated with relaxation. Thus, an
electrode placed over the visual cortex that detects alpha band signals is detecting visual
relaxation. An electrode over the motor cortex picking up alpha band signals is detecting
motor relaxation (the mu rhythm).
2.6 Modules/Blocks of BCI system

A BCI comprises of five main parts. First part is an acquisition system, which directly
captures EEG or any other signals from the brain. It also performs noise elimination,
artifact processing etc. Secondly there is a pre-processing block. The aim of this block is
to bring the signals into a suitable form for further processing. Third is the feature
extractor, which extracts features from the signals, forming feature vectors, upon which
classification can be done. Then comes the classification block, which classifies the
signals using their feature vectors. Finally comes the control interface which translates
the classification of the previous block into meaningful commands for any device

connected, such as a wheel chair or a computer. Non-linear time domain normalization
for the second block has been proposed.
Signal Acquisition
Preprocessing
Feature Extraction
Classification
Control Interface
Figure4: Modules

3. Literature Review
This section provides an overview of techniques and tools used by other researches
as compared to our algorithms in different modules of BCI systems.
This work proposes a novel approach for the second, third and fourth blocks of a BCI
system.
3.1 Algorithms for preprocessing

Non-linear time domain normalization for the second block has been proposed.
Other techniques commonly used are Common Average Referencing (CAR), Surface
Laplacian (SL), Independent Component Analysis (ICA), Common Spatial Patterns
(CSP), and Principal Component Analysis (PCA) [19]. Neural networks perform
well when their input signal range is limited. We do this by using non-linear
normalization used by [1], but in time domain. Time domain normalization requires
much less computing power and provides for faster processing which is a critical
factor in online implementation of BCI systems.
3.2 Algorithms for feature extraction

Current trends in feature extraction indicate usage of Fast Fourier Transform (FFT)
[1], [3], Parametric modeling and spectral parameters [2], Signal envelope cross
correlation [4].
3.2.1 Wavelet transform

In this work we propose using an adapted averaged wavelet filter for transformation.
This adapted wavelet depicts EEG signal characteristics, such as drift, trends, abrupt
changes, and beginnings and ends of events, much better and hence improve feature
extraction. Wavelet transform has also been used by [5] and others [19], but this and
other techniques that use wavelet transform, use a predefined wavelet from some

wavelet family like Haar, Meyer, Morlet and Daubechies wavelets, although pattern
adapted wavelets have been used [20], [21], [22] and others [19], but the difference
between this method and previous ones is that this wavelet is adapted not only to the
individual tasks under consideration but also their specific time segments, thus
helping create a very efficient filter.
3.2.2 Fractional Fourier transform

Fractional Fourier Transform (FRFT) has also been used. The Brain Computer
Interface (BCI) using the Fractional Fourier Transform (FRFT) arises as the
generalization of works done on thought classification through the Fourier
Transform (FT) [1], [2], given that the FT is a particular case of the FRFT, thus adding
new dimensions that increase the accuracy of the correct thought classification.
Conventional Fourier analysis treats frequency and time as orthogonal variables and
consequently is only suitable for the analysis of signals with stationary
frequency content [10].
3.3 Classification
Heuristic weight adjustment of the Neural network classifiers has also been
proposed. This weight adjustment, as described in subsequent sections, provides for
a much better classification accuracy by utilizing the vital information of accuracy of
individual electrodes feeding the networks.
To summarize, this work has three contributions; first in preprocessing we have
used non-linear normalization in the time domain. Secondly wavelet transform with
pattern adapted wavelet filter because it provides better resolution in time-
frequency plane, thirdly heuristic weight adjustment of neural network classifiers.

The selection of exact radial basis network for classification provided an improved
efficiency in terms of time and memory usage of the system.
4. What we have done in each module

This section describes the main work we have done in each section. It explains in detail
the algorithms and techniques used by us in each module.
4.1 Signal Acquisition

Signal acquisition module is the interface with the user. It consists of sensors in the form
of a cap and electrodes placed on user’s head. After EEG signals are picked up by the
sensors they are passed through amplification and common noise rejections units.
4.1.1 Online
For online real time analysis we have designed a light weight, low cost and ultra portable
EEG machine. The EEG machine we have designed is described below. It is a five channel
machine, including one for reference. Cost of this prototype is less than PKR 300000.

Figure5: EEG machine
EEG MACHINE HARDWARE DESIGN
Basic elements
INPUT PROTECTION RESISTANCE
Interface amplifiers are often subjected to input overloads, i.e., voltage levels in excess of
their maximum for the selected gain range or even in excess of the supply voltage. These
overloads usually occur for only a fraction of a second and fall into two general classes:
steady state and transient (ESD, etc.).
Standard practice is to place current-limiting resistors in each input, but adding this
protection also increases the circuit’s noise level. A reasonable balance needs to be
found between the protection provided and the increased resistor (Johnson) noise

introduced. Circuits using in-amps with a relatively high noise level can tolerate more
series protection without seriously increasing their total circuit noise.
Of course, the less added noise the better, but a good guideline is that circuits needing
this extra protection can easily tolerate resistor values that generate 30% of the total
circuit noise. For example, a circuit using an in-amp with a rated noise level of 20nV/√Hz
can tolerate an additional 6nV/√Hz of Johnson noise.
Use the following cookbook method to translate this number into a practical resistance
value. A single 1kΩ would raise the maximum dc level to approximately 22.5V above each
supply or ±37.5V with ±15V supplies. The Johnson noise of a 1kΩ resistor is
approximately 4nV/√Hz. This value varies as the square root of the resistance. So, a
100kΩ resistor would have √100 times as much noise as the 1kΩ resistor, which is
40nV/√Hz (10 × 4nV/√Hz). Because both inputs need to be protected, two resistors are
needed, and their combined noise will add as the square root of the number of resistors
(the root sum of squares value). In this case, the total added noise from the two 100kΩ
resistors will be 56.56nV/√Hz (40 × 1.414).
With 3-op amp in-amp designs, when operating at low gains (10 or less), the gain resistor
acts as a current-limiting element in series with their resistor inputs. At high gains, the
lower value of RG may not adequately protect the inputs from excessive currents. The
AD620 uses its 400Ω internal resistor and a single set of diodes to protect against
negative input voltages. For positive voltage overloads, it relies on its own base-emitter
input junction to act as the clamping diode, and will safely withstand input overloads of

up to ±15V or ±60mA for several hours. This is true for all gains and power on and off,
which is particularly important since the signal source and amplifier may be powered
separately. For longer time periods, the current should not exceed 6mA (IIN ≤ VIN/400Ω).
REFERENCE:
Low Cost, Low Power Instrumentation Amplifier AD620 DATASHEET
A Designer’s Guide to Instrumentation Amplifiers, 3RD EDITION ANALOG DEVICES
OPERATIONAL AMPLIFIER
Figure6: OPAMP
Defining the Ideal Operational Amplifier

• Gain: The primary function of an amplifier is to amplify, so the more gain the
better. It can always be reduced with external circuitry, so we assume gain to be
infinite.

• Input Impedance: Input impedance is assumed to be infinite. This is so the
driving source won’t be affected by power being drawn by the ideal operational
amplifier.
• Output Impedance: The output impedance of the ideal operational amplifier is
assumed to be zero. It then can supply as much current as necessary to the load
being driven.
• Response Time: The output must occur at the same time as the inverting input so
the response time is assumed to be zero. Phase shift will be 180. Frequency
response will be flat and bandwidth infinite because AC will be simply a rapidly
varying DC level to the ideal amplifier.
• Offset: The amplifier output will be zero when a zero signal appears between the
inverting and non-inverting inputs.
A Summing Point Restraint

An important by-product of these properties of the ideal operational amplifier is that the
summing point, the inverting input, will conduct no current to the amplifier. This
property is to become an important tool for circuit analysis and design, for it gives us an
inherent restraint on our circuit – a place to begin analysis. Both the inverting and non-
inverting inputs must remain at the same voltage, giving us a second powerful tool for
analysis.

The Desirability of Feedback
Figure7: Open loop OPAMP
Consider the open loop amplifier used in the circuit shown in figure above. Note that no
current flows from the source into the inverting input - the summing point restraint-
hence, there is no voltage drop across RS and ES appears across the amplifier input.
When ES is zero, the output is zero. If ES takes on any non-zero value, the output voltage
increases to saturation, and the amplifier acts as a switch.
The open loop amplifier is not practical - once an op amp is pushed to saturation, its
behavior is unpredictable. Recovery time from saturation is not specified for op amps
(except voltage limiting types). It may not recover at all; the output may latch up. The
output structure of some op-amps, particularly rail-to-rail models, may draw a lot of
current as the output stage attempts to drive to one or the other rail.

Two Important Feedback Circuits
Figure8: Feedback circuits
Figure shows the connections and the gain equations for two basic feedback circuits. The
application of negative feedback around the ideal operational amplifier results in another
important summing point restraint: The voltage appearing between the inverting and
non-inverting inputs approaches zero when the feedback loop is closed. Consider either
of the two circuits shown in figure. If a small voltage, measured at the inverting input
with respect to the non-inverting input, is assumed to exist, the amplifier output voltage
will be of opposite polarity and can always increase in value (with infinite output
available) until the voltage between the inputs becomes infinitesimally small. When the
amplifier output is fed back to the inverting input, the output voltage will always take on
the value required to drive the signal between the inputs toward zero.
The two summing point restraints are so important that they are repeated:
1. No current flows into either input terminal of the ideal operational amplifier.

2. When negative feedback is applied around the ideal operational amplifier, the
differential input voltage approaches zero.
Voltage Follower
Figure9: Voltage follower
The circuit in figure above demonstrates how the addition of a simple feedback loop to
the open loop amplifier converts it from a device of no usefulness to one with many
practical applications.
Analyzing this circuit, we see that the voltage at the non-inverting input is EI, the voltage
at the inverting input approaches the voltage at the non-inverting input, and the output
is at the same voltage as the inverting input. Hence, EO = EI, and our analysis is complete.
Our result also may be verified by mathematical analysis very simply. Since no current
flows at the non-inverting input, the input impedance of the voltage follower is infinite.
The output impedance is just that of the ideal operational amplifier itself, i.e. zero. Note
also that no current flows through the feedback loop, so any arbitrary (but finite)
resistance may be placed in the feedback loop without changing the properties of the

ideal circuit, shown in figure below. No voltage would appear across the feedback
element and the same mathematical analysis would hold.
Figure10: Voltage follower with feedback
The feedback resistor is of particular importance if the op amp selected is a current-
feedback type. The stability of current-feedback op amps is dependent entirely on the
value of feedback resistor selected, and the designer should use the value recommended
on the data sheet for the device.
Unity gain circuits are used as electrical buffers to isolate circuits or devices from one
another and prevent undesired interaction. As a voltage following power amplifier, this
circuit will allow a source with low current capabilities to drive a heavy load.
The gain of the voltage follower with the feedback loop closed (closed loop gain) is unity.
The gain of the ideal operational amplifier without a feedback loop (open loop gain) is
infinity. Thus, we have traded gain for control by adding feedback. Such a severe sacrifice
of gain – from infinity to unity - is not necessary in most circuits.

In our design voltage follower was used to buffer data from electrodes and after IC
ICL8038CCPD used for carrier generation. The purpose of using buffer after electrodes
was to improve the input impedance, although it can be removed if Instrumentation
Amplifier immediately after it has sufficient input impedance. Also if this buffer is not
used then protection resistances at the start can also be removed resulting in significant
reduction in noise.
OP177GP IC was used as voltage follower with no feedback resistance right after
electrodes and TL061 was used after ICL8038CCPD. TL061 can be replaced by TL062
which has two op-amps within single package reducing design size and space. A voltage
follower must be used after ICL8038CCPD because of impedance mismatch between
ICL8038 and AD633 IC which is used for amplitude modulation.
Non-Inverting Amplifier
Figure11: Non-inverting amp

Voltage follower is simply a special case of non-inverting amplifier. Since no current flows
into the inverting input, RO and RI form a simple voltage divider. The same voltage must
appear at the inverting and non-inverting inputs, so that:
From the voltage division formula:
The input impedance of the non-inverting amplifier circuit is infinite since no current
flows into the inverting input. Output impedance is zero since output voltage is ideally
independent of output current.
Closed loop gain can be any desired value above unity.
Inverting Amplifier
Figure12: Inverting amp

The inverting amplifier is shown in figure above. This circuit and its many variations form
the bulk of commonly used operational amplifier circuitry. Single ended input and output
versions were first used, and they became the basis of analog computation. Today’s
modern differential input amplifier is used as an inverting amplifier by grounding the
non-inverting input and applying the input signal to the inverting input terminal. Since
the amplifier draws no input current and the input voltage approaches zero when the
feedback loop is closed (the two summing point restraints), we may write:
Hence:
Input impedance to this circuit is not infinite as in the two previous circuits, the inverting
input is at ground potential so the driving source effectively “sees” RI as the input
impedance. Output impedance is zero as in the two previous circuits.

Integrator
Figure13: Integrator
If a capacitor is used as the feedback element in the inverting amplifier, shown in figure
above, the result is an integrator. Current through the feedback loop charges the
capacitor and is stored there as a voltage from the output to ground. This is a voltage
input current integrator.
Voltage Adder
In a great many practical applications the input to the inverting amplifier is more than
one voltage. The simplest form of multiple inputs is shown in figure below.

Figure14: Voltage summer
Current in the feedback loop is the algebraic sum of the current due to each input. Each
source, E1, E2, etc., contributes to the total current, and no interaction occurs between
ିோ೚
them. All inputs “see” RI as the input impedance, while gain is ோ಺
. Direct voltage
addition may be obtained with RO = RI.
In case of our EEG Signal Acquisition Device Design, we have four channels of EEG data
modulated at different frequencies are added at the end with R1=R2=R3=R4=100kΩ and
Ro= 1kΩ. Operational Amplifier used was OP177GP with ±15V dual supply. The purpose
of attenuating the input signal is sound card. Input signal to sound card must not exceed

1V or it will be clipped. In our case we have ensured that signal will not exceed 0.6V.
OP177GP IC can be replaced with OP277 as they offer improved noise, wider output
voltage swing, and are twice as fast with half the quiescent current compared to OP177.
Low Pass Filter
Figure15: Response curve

Figure16: Response curve2
Figure17: Response curve3
Circuit Diagram:
Operational Amplifier used was OP177GP which is dual supply IC.

Figure18: LPF
High Pass Filter

Instrumentation Amplifier AD620 has a reference pin which provides AC and DC coupling
functionality. If Ref pin is grounded it performs DC coupling that is it compares the
output with ground and allows all frequencies to pass but if output is passed through a
low pass filter and output is returned to Ref pin it acts as a High Pass Filter as well. In our
design we designed a Low Pass Filter with R=330kΩ and C=1µF, which provides a cutoff
frequency of 0.48Hz approximately. When the output is fed to Ref pin of AD620, it cuts
off all frequencies below 0.48Hz.
REFERENCE:
HANDBOOK OF OPERATIONAL AMPLIFIER APPLICATIONS, TEXAS INSTRUMENTS
INSTRUMENTATION AMPLIFIER
An instrumentation amplifier is a closed-loop gain block that has a differential input and
an output that is single-ended with respect to a reference terminal. Most commonly, the

impedances of the two input terminals are balanced and have high values, typically 109Ω,
or greater. The input bias currents should also be low, typically 1nA to 50nA. As with op
amps, output impedance is very low, nominally only a few milliohms, at low frequencies.
Unlike an op amp, for which closed-loop gain is determined by external resistors
connected between its inverting input and its output, an in-amp employs an internal
feedback resistor network that is isolated from its signal input terminals. With the input
signal applied across the two differential inputs, gain is either preset internally or is user
set (via pins) by an internal or external gain resistor, which is also isolated from the signal
inputs.
In its primary function, the in-amp will normally reject the common-mode dc voltage, or
any other voltage common to both lines, while amplifying the differential signal voltage,
the difference in voltage between the two lines. In contrast, if a standard op amp
amplifier circuit were used in this application, it would simply amplify both the signal
voltage and any dc, noise, or other common-mode voltages. As a result, the signal would
remain buried under the dc offset and noise. Because of this, even the best op-amps are
far less effective in extracting weak signals.
Signal Amplification and Common-Mode Rejection

An instrumentation amplifier is a device that amplifies the difference between two input
signal voltages while rejecting any signals that are common to both inputs. The in-amp,
therefore, provides the very important function of extracting small signals from
transducers and other signal sources.

Common-mode rejection (CMR), the property of canceling out any signals that are
common (the same potential on both inputs), while amplifying any signals that are
differential (a potential difference between the inputs), is the most important function
an instrumentation amplifier provides. Both dc and ac common-mode rejection are
important in-amp specifications. Any errors due to dc common-mode voltage (i.e., dc
voltage present at both inputs) will be reduced 80 dB to 120 dB by any modern in-amp of
decent quality.
However, inadequate ac CMR causes a large, time-varying error that often changes
greatly with frequency and, therefore, is difficult to remove at the IA’s output.
Fortunately, most modern monolithic IC in-amps provide excellent ac and dc common-
mode rejection.
Common-mode gain (ACM), the ratio of change in output voltage to change in common-
mode input voltage, is related to common-mode rejection. It is the net gain (or
attenuation) from input to output for voltages common to both inputs. For example, an
in-amp with a common-mode gain of 1/1000 and a 10V common mode voltage at its
inputs will exhibit a 10 mV output change. The differential or normal mode gain (AD) is
the gain between input and output for voltages applied differentially (or across) the two
inputs. The common-mode rejection ratio (CMRR) is simply the ratio of the differential
gain, AD, to the common-mode gain.
Note that in an ideal in-amp, CMRR will increase in proportion to gain.

Common-mode rejection is usually specified for full range common-mode voltage (CMV)
change at a given frequency and a specified imbalance of source impedance (e.g., 1kΩ
source imbalance, at 60 Hz).
Mathematically, common-mode rejection can be represented as
where:
AD is the differential gain of the amplifier; VCM is the common-mode voltage present at
the amplifier inputs; VOUT is the output voltage present when a common-mode input
signal is applied to the amplifier. The term CMR is a logarithmic expression of the
common-mode rejection ratio (CMRR).
That is, CMR = 20 log10 CMRR.
To be effective, an in-amp needs to be able to amplify microvolt-level signals while
rejecting common-mode voltage at its inputs. It is particularly important for the in-amp
to be able to reject common-mode signals over the bandwidth of interest. This requires
that instrumentation amplifiers have very high common-mode rejection over the main
frequency of interest and its harmonics.
At unity gain, typical dc values of CMR are 70 dB to more than 100 dB, with CMR usually
improving at higher gains. While it is true that operational amplifiers connected as
subtractors also provide common-mode rejection, the user must provide closely matched

external resistors (to provide adequate CMRR). On the other hand, monolithic in-amps,
with their pre-trimmed resistor networks, are far easier to apply.
Common-Mode Rejection: Op Amp vs. In-Amp

Op amps, in-amps, and difference amps all provide common-mode rejection. However,
in-amps and diff amps are designed to reject common-mode signals so that they do not
appear at the amplifier’s output. In contrast, an op amp operated in the typical inverting
or non-inverting amplifier configuration will process common-mode signals, passing
them through to the output, but will not normally reject them.
Consider an op amp connected to an input source that is riding on a common-mode
voltage in inverting configuration. Because of feedback applied externally between the
output and the summing junction, the voltage on the “–” input is forced to be the same
as that on the “+” input voltage. Therefore, the op amp ideally will have zero volts across
its input terminals. As a result, the voltage at the op amp output must equal VCM, for zero
volts differential input. Even though the op amp has common-mode rejection, the
common-mode voltage is transferred to the output along with the signal. In practice, the
signal is amplified by the op-amp’s closed-loop gain, while the common-mode voltage
receives only unity gain. This difference in gain does provide some reduction in common-
mode voltage as a percentage of signal voltage. However, the common-mode voltage still
appears at the output, and its presence reduces the amplifier’s available output swing.
For many reasons, any common-mode signals (dc or ac) appearing at the op amp’s
output is highly undesirable.

Figure19: Internal structure AD620
Now consider a 3-op amp in-amp operating under the same conditions. Note that, just
like the op amp circuit, the input buffer amplifiers of the in-amp pass the common-mode
signal through at unity gain. In contrast, the signal is amplified by both buffers. The
output signals from the two buffers connect to the subtractor section of the IA. Here the
differential signal is amplified (typically at low gain or unity) while the common-mode
voltage is attenuated (typically by 10,000:1 or more). Contrasting the two circuits, both
provide signal amplification (and buffering), but because of its subtractor section, the in-
amp rejects the common-mode voltage. In-amps find their primary use amplifying
signals from low level output transducers in noisy environments.
Since an ideal instrumentation amplifier detects only the difference in voltage between
its inputs, any common-mode signals (equal potentials for both inputs), such as noise or
voltage drops in ground lines, are rejected at the input stage without being amplified.

Either internal or external resistors may be used to set the gain. Internal resistors are the
most accurate and provide the lowest gain drift over temperature.
One common approach is to use a single external resistor, working with two internal
resistors, to set the gain. The user can calculate the required value of resistance for a
given gain, using the gain equation listed in the in-amp’s spec sheet. This permits gain to
be set anywhere within a very large range. However, the external resistor can seldom be
exactly the correct value for the desired gain, and it will always be at a slightly different
temperature than the IC’s internal resistors. These practical limitations always contribute
additional gain error and gain drift.
Sometimes two external resistors are employed. In general, a 2-resistor solution will have
lower drift than a single resistor as the ratio of the two resistors sets the gain, and these
resistors can be within a single IC array for close matching and very similar temperature
coefficients (TC).
Conversely, a single external resistor will always be a TC mismatch for an on-chip resistor.
The output of an instrumentation amplifier often has its own reference terminal, which,
among other uses, allows the in-amp to drive a load that may be at a distant location.
Of course, power must be supplied to the in-amp. As with op amps, the power would
normally be provided by a dual-supply voltage that operates the in-amp over a specified
range. Alternatively, an in-amp specified for single-supply (rail-to-rail) operation may be
used.

An instrumentation amplifier may be assembled using one or more operational
amplifiers, or it may be of monolithic construction. Both technologies have their
advantages and limitations.
In general, discrete (op amp) in-amps offer design flexibility at low cost and can
sometimes provide performance unattainable with monolithic designs, such as very high
bandwidth. In contrast, monolithic designs provide complete in-amp functionality and
are fully specified and usually factory trimmed, often to higher dc precision than discrete
designs. Monolithic in-amps are also much smaller, lower in cost, and easier to apply.
High AC (and DC) Common-Mode Rejection

At a minimum, an in-amp’s CMR should be high over the range of input frequencies that
need to be rejected. This includes high CMR at power line frequencies and at the second
harmonic of the power line frequency.
Low Offset Voltage and Offset Voltage Drift

As with an operational amplifier, an in-amp must have low offset voltage. Since an
instrumentation amplifier consists of two independent sections, an input stage and an
output amplifier, total output offset will equal the sum of the gain times the input offset
plus the offset of the output amplifier (within the in-amp). Typical values for input and
output offset drift are 1µV/oC and 10µV/oC, respectively. Although the initial offset
voltage may be nulled with external trimming, offset voltage drift cannot be adjusted
out. As with initial offset, offset drift has two components, with the input and output
section of the in-amp each contributing its portion of error to the total.

As gain is increased, the offset drift of the input stage becomes the dominant source of
offset error.
A Matched High Input Impedance

The impedances of the inverting and non-inverting input terminals of an in-amp must be
high and closely matched to one another. High input impedance is necessary to avoid
loading down the input signal source, which could also lower the input signal voltage.
Values of input impedance from 109Ω to 1012Ω are typical. Difference amplifiers, such as
the AD629, have lower input impedances, but can be very effective in high common-
mode voltage applications.
Low Input Bias and Offset Current Errors

Again, as with an op amp, an instrumentation amplifier has bias currents that flow into,
or out of, its input terminals; bipolar in-amps have base currents and FET amplifiers have
gate leakage currents. This bias current flowing through an imbalance in the signal source
resistance will create an offset error. Note that if the input source resistance becomes
infinite, as with ac (capacitive) input coupling, without a resistive return to power supply
ground, the input common-mode voltage will climb until the amplifier saturates. A high
value resistor connected between each input and ground is normally used to prevent this
problem. Typically, the input bias current multiplied by the resistor’s value in ohms
should be less than 10mV. Input offset current errors are defined as the mismatch
between the bias currents flowing into the two inputs. Typical values of input bias
current for a bipolar in-amp range from 1nA to 50nA; for a FET input device, values of
1pA to 50pA are typical at room temperature.

Low Noise
Because it must be able to handle very low level input voltages, an in-amp must not add
its own noise to that of the signal. A minimum input noise level of 10nV/√Hz @ 1 kHz
(gain > 100) referred to input (RTI) is desirable. Micro-power in-amps are optimized for
the lowest possible input stage current and, therefore, typically have higher noise levels
than their higher current cousins.
Low Nonlinearity
Input offset and scale factor errors can be corrected by external trimming, but
nonlinearity is an inherent performance limitation of the device and cannot be removed
by external adjustment. Low nonlinearity must be designed in by the manufacturer.
Nonlinearity is normally specified as a percentage of full scale, whereas the manufacturer
measures the in-amp’s error at the plus and minus full-scale voltage and at zero. A
nonlinearity error of 0.01% is typical for a high quality in-amp; some devices have levels
as low as 0.0001%.
Simple Gain Selection

Gain selection should be easy. The use of a single external gain resistor is common, but
an external resistor will affect the circuit’s accuracy and gain drift with temperature. In-
amps, such as the AD621, provide a choice of internally preset gains that are pin-
selectable, with very low gain TC.
Adequate Bandwidth
An instrumentation amplifier must provide bandwidth sufficient for the particular
application. Since typical unity gain, small-signal bandwidths fall between 500 kHz and 4
MHz, performance at low gains is easily achieved, but at higher gains bandwidth

becomes much more of an issue. Micro-power in-amps typically have lower bandwidth
than comparable standard in-amps, as micro-power input stages are operated at much
lower current levels.
Differential to Single-Ended Conversion

Differential to single-ended conversion is, of course, an integral part of an in-amp’s
function: A differential input voltage is amplified and a buffered, single-ended output
voltage is provided. There are many in-amp applications that require amplifying a
differential voltage that is riding on top of a much larger common-mode voltage. This
common-mode voltage may be noise, or ADC offset, or both. The use of an op amp
rather than an in-amp would simply amplify both the common mode and the signal by
equal amounts. The great benefit provided by an in-amp is that it selectively amplifies
the (differential) signal while rejecting the common-mode signal.
Rail-to-Rail Input and Output Swing

Modern in-amps often need to operate on single-supply voltages of 5 V or less. In many
of these applications, a rail-to-rail input ADC is often used. So-called rail-to-rail operation
means that an amplifier’s maximum input or output swing is essentially equal to the
power supply voltage. In fact, the input swing can sometimes exceed the supply voltage
slightly, while the output swing is often within 100mV of the supply voltage or ground.
Careful attention to the data sheet specifications is advised.
Power vs. Bandwidth, Slew Rate, and Noise

As a general rule, the higher the operating current of the in-amp’s input section, the
greater the bandwidth and slew rate and the lower the noise. But higher operating
current means higher power dissipation and heat. Battery operated equipment needs to

use low power devices, and densely packed printed circuit boards must be able to
dissipate the collective heat of all their active components. Device heating also increases
offset drift and other temperature-related errors. IC designers often must trade off some
specifications to keep power dissipation and drift to acceptable levels.
For many years, the AD620 has been the industry standard, high performance, low cost
in-amp. The AD620 is a complete monolithic instrumentation amplifier offered in both 8-
lead DIP and SOIC packages. The user can program any desired gain from 1 to 1000 using
a single external resistor. By design, the required resistor values for gains of 10 and 100
are standard 1% metal film resistor values.
Figure20: pin configuration AD620
The AD620 is a second-generation version of the classic AD524 in-amp and embodies a
modification of its 3-op amp circuit. Laser trimming of on-chip thin film resistors, R1 and
R2, allows the user to accurately set the gain to 100 within ±0.3% max error, using only
one external resistor. Monolithic construction and laser wafer trimming allow the tight
matching and tracking of circuit components.

The internal gain resistors, R1 and R2, are trimmed to an absolute value of 24.7kΩ,
allowing the gain to be programmed accurately with a single external resistor. The gain
equation is then
So that
Where resistor RG is in kΩ.
Instrumentation Amplifier used in our design was AD620. We experimented with INA114,
INA118, INA128, INA129, INA126 and AD620.
REFERENCE:
A Designer’s Guide to Instrumentation Amplifiers, 3RD EDITION ANALOG DEVICES
Modulation
The generation of AM signals consists simply of the addition of the carrier waveform to
the DSB signal, for this reason AM modulation is referred to as DSB-LC (DSB with large
carrier). The message signal x(t) is modulated by a carrier waveform Ac cos(2πfct). The
amplitude modulated waveform xc(t) is
xc(t) = Acx(t) cos(2πfct) + Ac cos(2πfct)
= Ac[1 + x(t)] cos(2πfct)
= A(t) cos(2πfct)

where |A(t)| is the envelope of the AM signal.
We have used AD633 IC for Amplitude Modulation. It is a dual supply IC operated
typically at ±15V. The AD633 is a functionally complete analog multiplier. It includes high
impedance, differential X and Y inputs and a high impedance summing input (Z).
Figure21: pin configuration AD633
Pin 2 and Pin 4 were grounded. Pin 3 was connected with Pin 6 to implement amplitude
modulation.
Circuit Diagram:

Figure22: Circuit Diagram AD633
Carrier Generation
The ICL8038 waveform generator is a monolithic integrated circuit capable of producing
high accuracy sine, square, triangular, sawtooth and pulse waveforms with a minimum of
external components. The frequency (or repetition rate) can be selected externally from
0.001Hz to more than 300kHz using either resistors or capacitors, and frequency
modulation and sweeping can be accomplished with an external voltage.
Figure23: Carrier generator

4.1.2 Offline
The first dataset (dataset1) used was acquired from Colorado University website [6],
which was actually obtained by [8]. This dataset has also been used by [1], [3], [9] and
many others. Seven Non-Invasive active electrodes C3, C4, P3, P4, O1, O2 and EOG
were used to record EEG signals at a sampling frequency of 250Hz from seven
individuals, six males and one female. EOG which does not appear in this figure is
used for measuring movement of eyeballs. Therefore for a 10sec recording 250×10 =
2500 samples were recorded. Placement of electrodes on the scalp is shown in
Figure 2 by [1]. EOG which was used to record the movement of Eye ball is not shown.
The Dataset comprised of 5 Mental Tasks and each task was repeated five times on
two days, amounting to total trials of 50. These Tasks were
• Baseline Task
• Multiplication Task
• Letter Composing Task
• 3-D Rotation Task
• Counting Task
The second dataset (dataset2) was obtained from Department of Medical
Informatics, Institute for
Biomedical Engineering, University of Technology Graz. (Gert Pfurtscheller) [9]. This
dataset was recorded from a normal subject (female, 25y) during a feedback session.
The subject sat in a relaxing hair with armrests. The task was to control a feedback
bar by means of imagery left or right hand movements. The order of left and right
cues was random.

The experiment consists of 7 runs with 40 trials each. All runs were conducted on
the same day with several minutes break in between. Given are 280 trials of 9s length.
The first 2s was quite, at t=2s an acoustic stimulus indicates the beginning of the trial, the
trigger channel (#4) went from low to high, and a cross “+” was displayed for 1s;
then at t=3s, an arrow (left or right) was displayed as cue. At the same time the
subject was asked to move a bar into the direction of a the cue. The feedback was based
on AAR parameters of channel #1 (C3) and #3 (C4), the AAR parameters were
combined with a discriminant analysis into one output parameter. The recording was
made using a G.tec amplifier and Ag/AgCl electrodes. Three bipolar EEG channels
(anterior „+‟, posterior „-„) were measured over C3, Cz and C4. The EEG was sampled
with 128Hz, it was filtered between 0.5 and 30Hz.
4.2 Preprocessing
Preprocessing is required to bring the signals into a suitable ‘shape’ before feeding them
to subsequent blocks.
4.2.1 Windowing
To reduce the execution time and memory utilization in processing 10secs EEG
signal, each signal of dataset1 was segmented into windows of 1sec with an overlap of
0.5sec. Overlap was introduced so as to cater for the loss at the edges of the window and
to make BCI response fast as each new window was available after 0.5sec.. This
segmentation does not effect any cross validation effort, because this much overlap
does not make signals identical to each other or dependent on each other over
times greater than the apodization time. Thus cross validation can be effectively

performed because the proposed scheme divides the signal in to windows, and training
and testing can be done on separate halves of the signal as done by [1].
4.2.2 Time domain normalization

Neural Networks when used for classification perform far better if the input data lies in a
small range of values. Often important information is present in small samples which
could be overwhelmed by large insignificant sample values if not appropriately scaled.
For that purpose the data was non- linearly normalized using a normalization technique
proposed by [5]. Mathematically
log( x − x min )
y= (1)
log( x max − x max )
Where in (1) y is the normalized sample, x is the current sample, and xmax and xmin are
maximum and minimum values in the 1 sec window under consideration. It must be
remembered that this normalization was done in time domain, prior to feature
extraction, as opposed to [1], which proposes normalization of the FFT coefficients in
frequency domain.
4.3 Feature Extraction

Feature extraction is perhaps the most important signal processing step in a BCI. We
have introduced two techniques for feature extraction, Wavelet transform (WT) and
Fractional Fourier Transform (FRFT).
4.3.1 Wavelet Transform (WT)

This section describes the logic behind the selection of wavelet analysis as feature
extraction tool. In transforming to the frequency domain using Fourier transform, time

information is lost. When looking at a Fourier transform of a signal, it is impossible to tell
when a particular event took place. If the signal is a stationary signal, this drawback isn't
very important. However the EEG signals contain numerous non stationary or transitory
characteristics: drift, trends, abrupt changes, and beginnings and ends of events. Wavelet
analysis provides a much better resolution, that is, it allows the use of long time intervals
where we want more precise low-frequency information, and shorter regions where we
want high frequency information. The major advantage afforded by wavelets is the
ability to perform local analysis, that is, to analyze a localized area of a larger signal.
Wavelet Transform is a representation of a signal x(t) by a basis function Ψ having such
properties as smoothness, localization of Time Frequency, orthogonality and symmetry
obtained by scaling and shifting the mother wavelet. Mathematically;
∞
X (a, b) = ∫ x(t )ψ (a, b, t )dt (2)
−∞
where α and β are real numbers representing scale and position of the wavelet
respectively. The appropriate selection of basis function Ψ, the wavelet, is significant as
it should approximate a given pattern in the signal. By merely looking at the brain signals
one might erroneously consider them to be highly random in nature, which in fact they
are not. There is a pattern associated with every task that human eye cannot decipher.
The fundamental assumption in pattern detection is that data is not random and that the
pattern repeats itself with a slight variation where pre-processing is required to remove
this variation to the extent possible and to make the pattern detectable. For that very
reason use of wavelet transform was made. The basic idea behind using an adapted

wavelet is to use a wavelet for analysis which actually depicts the signal characteristics
and thus provides for better feature extraction.
Pattern adapted wavelet using least Square Optimization was selected for pattern
detection of these signals. A wavelet was adapted for each window segment of data
using polynomial method and degree of polynomial was set to be five. Wavelet for each
task in both datasets was obtained for one second window using Matlab® Wavelet
toolbox command ‘pat2cwav’, i.e. ‘pattern to continuous wave’. This command
computes an admissible wavelet for CWT adapted to the given pattern. The principle for
designing a new wavelet for CWT is to approximate a given pattern using least squares
optimization under constraints leading to an admissible wavelet well suited for the
pattern detection using the continuous wavelet transform [16]. An average wavelet filter
was obtained using simple averaging of the adapted wavelets over all the tasks for both
datasets. Wavelets for each window segment, i.e., for each one second window was used
in the averaging processes to yield a filter for each of the window segments in the signal,
to be used for CWT of that window segment of the training and test data. Thus the CWT
coefficients were obtained using adapted wavelet filter for each time segmented window
of data. Entire process of synthesizing the filter was carried out with Matlab® wavelet
toolbox [18]. Another work, [17], also describes a similar filter for performing wavelet
transformation. However the filter described in [17] is adapted to a single neural activity.
The filter proposed in this paper takes it one step ahead by providing and testing a
wavelet filter prepared by averaging the adapted wavelets of all the tasks under
consideration for wavelet analysis. This filter is thus optimized for the tasks under

consideration. This wavelet adaptation scheme combined with data segmentation and
normalization yields excellent feature vectors for classification, the reason for it can be
explained as; when we make overlapping segments of the given signal, we get a
continuous flow of pattern in our resultant segmented signal.
Even though signal is time segmented, yet due to their overlapping nature each segment
picks up the pattern where the previous one dropped it, thus when we extract wavelets
from this pattern, we get extremely useful pattern information in these wavelets.
4.3.2 Fractional Fourier Transform

The Brain Computer Interface (BCI) using the Fractional Fourier Transform (FRFT)
arises as the generalization of works done on thought classification through the Fourier
Transform (FT) [1], [2], given that the FT is a particular case of the FRFT, thus adding
new dimensions that increase the accuracy of the correct thought classification.
Conventional Fourier analysis treats frequency and time as orthogonal variables and
consequently is only suitable for the analysis of signals with stationary frequency
content [10]. Exactly as the Fourier harmonic analysis employs sinusoidal function to
decompose periodic signals, Fractional Fourier techniques employ chirp harmonics for
the decomposition of signals with time-varying periodicity [11]. It can be
interpreted as the representation of a signal in neutral domain by means of the
rotation of the signal by the origin in counter-clockwise direction with rotational
angle in time-frequency domain [12].
In mathematics, in the area of harmonic analysis, the fractional Fourier transform (FRFT)
is a linear transformation generalizing the Fourier transform. It can be thought of as the

Fourier transform to the n-th power where n need not be an integer — thus, it can
transform a function to an intermediate domain between time and frequency. Its
applications range from filter design and signal analysis to phase retrieval and pattern
recognition.
Interpretation of the Fractional Fourier Transform

The usual interpretation of the Fourier transform is as a transformation of a time domain
signal into a frequency domain signal. On the other hand, the interpretation of the
inverse Fourier transform is as a transformation of a frequency domain signal into a time
domain signal. Apparently, fractional Fourier transforms can transform a signal (either in
the time domain or frequency domain) into the domain between time and frequency: it
is a rotation in the time-frequency domain. This perspective is generalized by the linear
canonical transformation, which generalizes the fractional Fourier transform and allows
linear transforms of the time-frequency domain other than rotation.
Actually, fractional Fourier transform is a rotation operation on the time frequency
distribution. From the definition above, for α=0, there will be no change after applying
fractional Fourier transform, and for α=π/2, fractional Fourier transform becomes a
Fourier transform, which rotates the time frequency distribution with π/2. For other
value of α, fractional Fourier transform rotates the time frequency distribution according
to α.
If the signal in the time domain is rectangular, it will become a sinc function in the
frequency domain. But if we apply the fractional Fourier transform to the rectangular
signal, the transformation output will be in the domain between time and frequency.

There are many approaches to defining FRFT, one of them is given below. The FRFT of
a s(x1) signal can be given as
ஶ
‫ ܨ‬௔ [‫ݏ‬ሺ‫ݔ‬ଵ ሻ = ܵሺ‫ݔ‬ሻ = න ‫ܭ‬௣ ሺ‫ݔ‬, ‫ݔ‬ଵ ሻ‫ݏ‬ሺ‫ݔ‬ଵ ሻ݀ሺ‫ݔ‬ଵ ሻ
ିஶ
Where kernel function is defined by:
௫ ௫
ି௜గ൬ଶ భ ି൫௫ మ ା௫భ మ ൯ୡ୭୲ ሺ௔ሻ൰
‫ܭ‬௣ ሺ‫ݔ‬, ‫ݔ‬ଵ ሻ = ‫ܥ‬௔ ݁ ୱ୧୬ ሺ௔ሻ
With ,‫ܥ‬௔ = ඥ1 − ݅ܿ‫ݐ݋‬ሺܽሻ , ܽ = ܽߨ/2 and “a” is order of the FRFT. Also observe
that, F0=I; F1=F; F2=P; F3=FP=PF; and F4=F0= I, where I is an identity operator and
P is a parity operator. According to this definition, the zero-order transform of a
function is the same as the function itself s(x1), the first order transform is the
Fourier transform of s(x1), and the ±2nd order transform equals s(-x1).
Figure24: Time Frequency axis
Periodic nature of FRFT operator implies that the range of α and the order of the
transform can be restricted to α∈[0, 2π) and [0, 4) respectively. Value of α was
selected on hit and trial basis.

4.4 Classification
Multilayer Exact Radial Basis (RBE) Neural Network were used for the purpose of
mental task classification. An RBE network contains one hidden radial basis layer of
as many neurons as input vectors and an output linear layer and has a property of
returning the exact target vectors if training and testing data is same. It has an added
advantage that its training time is noticeably less compared to other neural networks
4.4.1 Neural networks as classifiers

One Multilayer Exact Radial Basis (RBE) Neural Network was trained for each
electrode, giving us a total of six (one for each of C3, C4, P3, P4, O1, O2 electrodes)
networks for dataset1 and two networks for dataset2 corresponding to electrodes C3
and C4. We have used a parallel structure of networks for classification. This
parallel structure provides for a much better classification because each network
works on a separate electrode thus fully utilizing the montage characteristics of
individual electrodes.
RBE Neural Networks were trained with first five alternate segments of the
dataset1 of first day trials only (i.e. first five trials only), thus giving 5Tasks ×
5Trials × 5Segments = 125segments of data for training. Since segments overlap,
selection of alternate data segments improves training considerably. Target vector
was chosen to be (1, 0, 0, 0, 0) where the task under consideration was assigned a
high value and remaining all were assigned low values. Similar actions were taken for
dataset2 and real time data.

4.4.2 Heuristic Weight Adjustment for Final Decision Making
A task was classified as one if majority of the networks classified it to be that particular
task. The concept of weigh

weightt adjustment comes in when we assign weights to networks
while deciding. Since the different networks correspond to different electrodes which
are feeding them and for which they are trained. Electrodes were found to exhibit
different fidelities
ies in representing the different tasks, and based on these differences
weight adjustment was made, so as to provide the network (or the electrode) with
better accuracy, a greater share in classification. Accuracy without this step was found
to
o be 62.63% for dataset1 and 71% for dataset2. This technique improved the
correct classification accuracy of the system to 84% and 87% for datasets 1 and 2
respectively.
4.5 Control Interface

Three different control interfaces have be
been developed.
i. Cursor control
ii. Video game (car game in Java)
iii. Toy car game (parallel port interface)
4.5.1 Cursor control

Cursor control enables user to completely control a cursor
in four directions without touching the mouse. As obvious
such a system can be of immense value to physically
disabled people. Since this system works in real time, the user can get used to it easily
(claim based on tests).

This system defines four classes of directions, and the user can use a combination of
these directions to move the cursor in any desired direction.
Screen is divided into icon size square blocks so that any one movement by the user of
cursor doesn’t overshoot an icon. Depending upon the screen resolution the step size of
cursor movement changes. A java platform (robot) was used for this purpose. It was
loading directly in matlab, and then cursor control was performed by it. Java platform
was preferred on Matlab inbuilt control due to two reasons. One Java gives a virtually
moving affect, two it also has wider options available, which also include right and left
click control. Which can be used for further enhancement in the project.
4.5.2 Car game in Java

Car game in Java is a simple game in which the player has to avoid hurdles on the path by
moving the car left or right through his mind. Game runs on a separate PC form the main
BCI engine (i.e. the PC on which the classification and other processes are run) to avoid
sluggish responses.
The two PCs are connected through peer to peer connection. Now it seems simple to
connect two PCs and run Matlab on one (with BCI engine) which sends classification
results to the other PC (with Java game) and Java receives these values and moves the
car in the desired direction. Communication is done through UDP (User datagram
protocol; if you don’t know). A very intriguing problem arises here..that the two PCs
wont communicate simply with UDP. Problem is with UDP communications. Following
code was used to solve the issue.

Figure25: Car game in Java
4.5.3 Toy car control

And the most interesting interface is the toy car. We have interfaced the remote control
of a toy car to a parallel port of PC. User can move car in any direction (through
combination of two basic moves). Parallel Port Pin 6 Bottom Layer is Reference Pin.
When 15 or 16 is sent either pin 3 or 6 (top layer from left side) gets high (3.26V). This
signal is first amplified to 9V and then it controls the 4066 CMOS Analog Switch. When 9V
is applied at pin 13 control pin of 4066, voltage at pin 1 is transferred to pin 2 of 4066
which in turn turns on the relay. Relay closes the switch of remote control which enables
the car to move in one direction or another. To turn off relay a zero is sent.

Figure26: Toy car with remote control

Figure27: Parallel port configuration
Toy car is controlled through parallel port. Our first attempt was at serial port.

5. Applications
EEG Bio feedback
• Personal/self-improvement/meditation
• Therapist-guided relaxation, etc.
• Peak-performance training, "Brain calisthenics"
• Adjunct to EMG, GRS, etc.
• Treatment of ADD & other clinical disorders
Computer Control and Communications

• "Thought-controlled" cursor, switch
• Brainwave-controlled games
Entertainment, Virtual Reality

• Control of music, graphics
• Control of VR Displays
• Interface to Light/Sound machines
Education and Research

• Labs, experiments, demonstrations
• Monitoring classroom/audience attention
Military Commercial
• Screening & evaluating pilots, operators
• EEG-based Cockpit controls, pilot monitoring
• Assessing consumer reactions

Each of these application areas has its own requirements, history, and prospects for low-
cost
5.1 EEG Biofeedback

EEG was first explored as a biofeedback modality by Kamiya, Brown, and others,
beginning in the 1960's. Initial work led to a generalized relaxation model, based
primarily on the alpha rhythm. Training was often done solely for the strengthening of
the alpha rhythm, without regard for other variables, or other brain rhythms. It was
found that developing the alpha rhythm, in and of itself, had limited value. Continued
work has developed methods that use other rhythms, or combinations of rhythms, in
both encouragement and discouragement protocols, to teach users to control the
relative amounts of rhythms, providing much more precise control of the brain.
5.2 Computer Control & Communication

The use of the EEG to control or communicate with a computer is an application that has
undergone slow but steady development since the 1960's. Early work by Pinneo and
others attempted to record "latent motor potentials" caused by a person thinking a
work, such as "up" or "down." More recently, investigators have been looking for signals
that appear controllable, and adapting the system to them. One of these is the
"sensorimotor rhythm" (SMR), that has been found to be under a certain amount of
conscious control. Generally, the user uses "affective" thoughts, such as "feeling light" or
"stopping thinking" to cause the desired rhythm.

5.3 Entertainment, Virtual Reality
Entertainment application include EEG-controlled composition and performance,
including "audience-participation" situations. Virtual Reality displays can use EEG to
modulate, alter, or otherwise control any aspect of the virtual world. For example, a
system could be made sensitive to the individual's overall cognitive and emotional state,
to produce an appropriate world. This could include changing the colors or sizes of
objects, controlling sunrise and sunset, or causing the appearance or removal of
features, or even of the entire location.
5.4 Education, Research

EEG is not well known below the level of the graduate student. Very few colleges, and no
known high schools, offer any opportunity for students to record, study, and understand
the EEG. This is unfortunate, because it is becoming increasingly clear that a basic
understanding of the EEG and its properties, especially with personal experience of
recording (ideally one's own) EEG, can provide valuable insight into the brain, as well as
the mind. For example, individual differences can be seen in EEG patterns between
people, and EEG changes in various tasks or circumstances can also be revealing. It
would be desirable for a greater number of students, at and below the undergraduate
level, to have direct experience with, and understanding of, the EEG. There is a certain
amount of popular use of phrases like "left brain," "right brain," "being in synch," "alpha
waves," and other related concepts. We like to understand how the brain operates, but
often use concepts that we must for the most part, take for granted, because there is no
practical way to check any of these ideas out. With the availability of low-cost,

scientifically sound brainwave monitors, it becomes possible to effectively record and
see anyone's brain rhythms, their left and right-brain activity, balance, synchrony, and
other variables.
5.5 Military, Commercial

The military has a long history of studying the EEG; some of the earliest telemetric
monitoring and analysis was developed by NASA during the 1960's, in connection with
the space program. This was designed primarily to monitor the pilots' state of health and
consciousness. In addition, the Air Force has had a long-standing program to develop
EEG-based pilot controls for the cockpit. These include evoked-potential based system,
which attempt to rapidly detect and act upon changes in the pilot's gaze, or level of
attention, to a display item.
Another ongoing area is alertness monitoring for commercial and military transportation
systems. Initial work used the ongoing EEG, and more recent work uses event-related
potential, and neural networks to interpret the waveforms.
Purely commercial applications include studying subjects who are viewing advertising
material or evaluating products, primarily to assess their level of interest and/or arousal.

6. References
[1] Kenji Nakayama, Yasuaki Kaneda, Akihiro Hirano, “A Brain Computer Interface based
on FFT and Multilayer Neural Network-Feature Extraction and Generalization”,
Proceedings of ISPACS(2007) International Symposium on Intelligent Signal Processing
and Communication Systems Nov.28-Dec.1, 2007 Xiamen, China
[2] Kenji Nakayama Kiyoto Inagaki, “A Brain Computer Interface Based on Neural
Network with Efficient Pre-Processing”, (ISPACS2006)
[3] Burke D P, Kelly S P, de Chazal P, Reilly R B and Finucane C 2005 “A parametric
feature extraction and classification strategy for brain–computer interfacing” IEEE
Trans. Neural Syst. Rehabil.
[4] Wang Y, Zhang Z, Gao X and Gao S 2004c “Lead selection for SSVEP-based
brain–computer interface” Proc. 26th Annual Int.Conf. of the IEEE Engineering in
Medicine and Biology Society (San Francisco, CA)
[5] Sam Darvishi and Ahmed Al-Ani , “Brain-Computer Interface Analysis using
Continuous Wavelet transform and Adaptive Neuro-Fuzzy Classifier “Proceedings of the
29th Annual International Conference of the IEEE EMBS Cité Internationale, Lyon,
France August 23-26, 2007.
[6] Colorado State University: http://www.cs.colostate.edu/eeg/

[7] Zachary A. Keirn. “Alternative modes of communication between man and machine”.
Master‟s thesis, Purdue University, 1988.
[8] “EEG Signal Classification for Brain Computer Interface Applications “ Ecole
Polytechniq Federale De Lausanne
http/bci.epfl.ch/publications/baztarricadiplomaproject.pdf
[9] Department of Medical Informatics, Institute for Biomedical Engineering,
University of Technology Graz. (Gert Pfurtscheller)
[10] Balu Santhanam,” On Discrete Gauss Hermite Functions and Eigenvectors of the
Discrete Fourier Transform”
[11] Massimiliano (Max) Martone, Member, IEEE,” A Multicarrier System Based on
the Fractional Fourier Transform for Time–Frequency-Selective Channels”
[12] Huadong Sun, Fulin Su, Weiyi Wang and Shouhe Tang,”Parameter Estimation of
Moving Targets in Airborne SAR Based on Fractional Fourier Transform”
[13] Boqiang Liu, Junbo Gao, Zhongguo Liu, Zhenwang Zhang, Cong Yin,Cuiping
Peng and Jason Gu, ““Brainwave Classification based on Wavelet Entropy and Event-
related Desynchronization”. (CCECE 2007)
[14] Clemens Brunner, Reinhold Scherer, Bernhard Graimann, Gernot Supp, and Gert
Pfurtscheller, “Online Control of a Brain-Computer Interface Using Phase
Synchronization”, IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 53, NO.
12, DECEMBER 2006

[15] Nuri F. INCE, Ahmed H. TEWFIK, Sami ARICA, Member, IEEE, “A space time
frequency analysis for the classification motor imagery EEG recordings in a Brain
Computer Interface task”, 28th IEEE EMBS Annual International Conference New York
City, USA, Aug2006
[16] M. Misiti, Y. Misiti, G. Oppenheim, J.M. Poggi, "Les ondelettes et leurs applications”,
Hermes, 2003.
[17] E. L. Glassman, “A Wavelet-Like Filter Based on Neuron Action Potentials for
Analysis of Human Scalp Electroencephalographs” IEEE Transactions on Biomedical
Engineering, vol. 52, issue no. 11, pp. 1851-1862, 2005.
[18] Matlab Wavelet Toolbox. MathWorks.
[19] A. Bashashati, M. Fatourechi, R. K. Ward and G. E. Birch, “A survey of signal
processing algorithms in brain–computer interfaces based on electrical brain signals”,
Journal of Neural Engineering vol. 4, IOP Publishing, pp. 32-57, 2007.
[20] E. L. Glassman, “A Wavelet-Like Filter Based on Neuron Action Potentials for
Analysis of Human Scalp Electroencephalographs” IEEE Transactions on Biomedical
Engineering, vol. 52, issue no. 11, pp. 1851-1862, 2005.
[21] V. J. Samar, K. P. Swartz, and M. R. Raghuveer, “Multiresolution analysis of event-
related potentials by wavelet decomposition”, Brain and Cognition, Academic Press, vol.
27, pp. 398–438, 1995.

[22] J. O. Chapa and M. R. Raghuveer, “Optimal matched wavelet construction and its
application to image pattern recognition,” in Wavelet Applications: II—Proceedings of
Society Photo Instrumentation Engineering SPIE, vol. 2491, pp. 518–529,1995.
[23] Khalid, M.B.; Rao, N.I.; Rizwan-i-Haque, I.; Munir, S.; Tahir, F., “A Brain Computer
Interface (BCI) using Fractional Fourier Transform with Time Domain normalization and
heuristic weight adjustment”, Signal Processing, 2008. ICSP 2008. 9th International
Conference. pp. 2734 – 2737, 26-29 Oct. 2008.
[24] Khalid, M.B.; Rao, N.I.; Rizwan-i-Haque, I.; Munir, S.; Tahir, F., “Towards a Brain
Computer Interface using wavelet transform with averaged and time segmented adapted
wavelets”, Computer, Control and Communication, 2009. IC4 2009. 2nd International
Conference on 17-18 Feb. 2009 Page(s):1 - 4

A Brain Computer Interface

Transféré par

Informations du document

Description originale:

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

A Brain Computer Interface

Transféré par

Droits d'auteur :

Formats disponibles

Think. Done!

A Brain Computer Interface (BCI)

Muhammad Bilal Khalid

Submitted to the Faculty of Electrical Engineering Department,

National University of Sciences and Technology, Rawalpindi in partial fulfillment for

the requirements of a B.E Degree in Telecommunication Engineering

Think. Done! A Brain Computer Interface (BCI)

Computer Interface (BCI)” submitted by 1) Muhammad Bilal Khalid, 2) Intisar Rizwan-i-

requirement of the degree.

Maj. Dr. Naveed Iqbal Rao

Military College of Signals,

Think. Done! A Brain Computer Interface (BCI)

A Brain Computer Interface (BCI) is a system that provides an individual a non-muscular

brain. These signals are in the form of Electroencephalogram (EEG).

most interesting interface is with a toy car through parallel port.

adjustment of networks has been introduced to improve classification accuracy.

international conferences [23], [24].

Think. Done! A Brain Computer Interface (BCI)

In the name of Allah, the Most Gracious, the Most Merciful

Our Beloved Parents and Siblings

Our Respectful Teachers

Think. Done! A Brain Computer Interface (BCI)

(S.W.A.T), the most merciful.

Processing Center (IPC).

Rawalpindi, Ammad Junaid of CS Dept. and Image Processing Center staff.

Think. Done! A Brain Computer Interface (BCI)

Think. Done! A Brain Computer Interface (BCI)

Think. Done! A Brain Computer Interface (BCI)

Think. Done! A Brain Computer Interface (BCI)

Figure 1: Block Diagram of BCI system

Figure 2: Our brain—part wise distribution

Figure 3: Motor cortex region of brain

Figure 4: Modules of BCI system

Figure 5: Our EEG machine—bird eye view

Figure 6: OPAMP—circuit diagram

Figure 7: Open loop OPAMP

Figure 8: Feedback Circuit

Figure 9: Voltage Follower

Figure 10: Voltage Follower with feedback

Figure 11: Non-inverting AMP

Figure 12: Inverting AMP

Figure 13: Integrator

Figure 14: Voltage summer

Figure 15: Response Curve

Figure 16: Response Curve 2

Figure 17: Response Curve 3

Figure 18: LPF

Figure 19: Internal structure AD620

Figure 20: Pin configuration AD620

Think. Done! A Brain Computer Interface (BCI)

Figure 22: Circuit diagram AD 633

Figure 23: Carrier Generator

Figure 24: Time frequency axis

Figure 25:Car game in Java

Figure 26:Toy car and its interface

Figure 27:Parallel port configuration

Think. Done! A Brain Computer Interface (BCI)

1.1 Motivation for this project

common: they require physical movement by the user. This physical/muscular

this machine can be a computer, a wheel chair or anything.

Brain Computer Interfacing is an exciting new technology. It provides an individual a non-

‘Brain Computer Interface Centers’ to accelerate development in this field.

Think. Done! A Brain Computer Interface (BCI)