Eduardo Ianez, José Maria Azorin,
Andrés Ubeda, José Manuel Ferrandez, Eduardo Fernandez
Robotics and Autonomous Systems, 58,
2010
Summary
Brain Computer Interface (BCI) methods allow
operators to generate control commands through EEG signals
(Electroencephalography), through the registration of “mental tasks”.
Brain activity product
electrical activity, magnetic signals (measured through magnetoencephalography,
MEG) electrical signals and metabolic signals (measured through the Positron
Emission Tomography, PET, or by Functional Magnetic Resonance Imaging, fMRI).
Techniques are divided in invasive and evasive, in the first case
microelectrodes are implanted directly in the brain (used for animals and
considered to be more precise), while in the second case are only having
cutaneous contact (used for human application being ethically preferred).
BCI can further be divided
in synchronous and asynchronous, the first when check the EEG and take the
decision, taking a certain amount of time, the second allowing actions to be
thought any time and therefore it’s chosen by the authors. The proposed BCI
uses Wavelet Transform to extract the relevant characteristics of the EEG, with
a new Linear Discriminant Analysis (LDA) for classifying the different EEG
patterns.
The experiments requires
16 scalp electrodes, with 9 concentrated on the top center, the rest are
supporting preprocessing. The information arriving from the Wavelet Transform
is then processed by the LDA classifier, which is based on Fisher linear
discriminant, using statistics and automatic learning to find the best linear
combination, the method guarantees the maximum separability between classes,
but four different models are required for simultaneous classification among
the three different classes (rest state, right mental tasks and left mental
tasks). Model 1 implies class 1 being Right side and class 2 left side and rest
state, model 2 has class one being the left side and class 2 being the right
and rest state, model 3 has class 1 being the rest state and class two the
right and left state and model 4 has class 1 being the right state and class
two the left state.
For the four models then a
decision system is created according on to the following rule: for the models
from 1 to 3 (task involving the brain’s right side) one point is assigned if
the vector belongs to class 1, 0.5 if it belongs to class 2 and 0 in the
uncertainty region; for region 4 one point is assigned either if the vector
belongs to class 1 or 2 and 0 if in uncertainty (the algorithm is shown at page
1251).
A test has been performed,
experimenting simulation offline (moving a dot), online and practice online
(moving a FANUC LR Mate 200iB robotic arm with 6 d.o.f.).
In the offline more errors
are tracked and later on, on the same session, users performed other 50s
activity online, the third session see the use of the robotic arm obtaining
final scores with the method previously introduced.
The offline results show
clearly that it is the worst way of performing tasks, it results in greater
errors and success percentage is therefore low, since the user doesn’t receive
any feedback from the system
The online method, which
uses Matlab interface, appears to be more effective with less error reported,
the problem is that EEG signals are time invariant, reason for which there are
wrong adjustments of data at initial calculations. It has been demonstrated
that through training the final results achieved are much better.
The robot arm, programmed
using C++, appears to be controlled well, having the same problems of the
online simulation, being the EEG time invariant.
Key
Concepts
EEG control systems
Key Results
The paper has introduced the
different steps involving BCI with Wavelet Transform, LDA-based classifier and
in the end the algorithm for decision making. For future implementations the
time invariance of EEG signals has to be adjusted and more degrees of freedom
should be added to command the robotic arm. The method proposed it thought
mainly for people with some physical disabilities.
No comments:
Post a Comment