Proceedings of 2015 IEEE 12th International Conference on Networking, Sensing and Control
Howard Civil Service International House, Taipei, Taiwan, April 9-11, 2015
Noise Reduction and Brain Mapping based Robust
Principal Component Analysis
Arjon Turnip
Technical Implementation Unit for Instrumentation Development
Indonesian Institute of Sciences, Indonesia
E-mail: arjon.turnip@lipi.go.id
Abstract—Analysis of EEG activity usually raises the problem
of differentiating between genuine EEG activity and that which is
introduced through a variety of external influence. These
artifacts may affect the outcome of the EEG recording. In this
paper, wavelet denoising and band pass filter for preprocessing
and a robustprincipal component analysis algorithm for
extraction are proposed to remove the artifacts. The algorithm is
designed to adaptively derive a relatively small number of
decorrelated linear combinations of a set of random zero-mean
variables while retaining as much of the information from the
original variables as possible. The proposed method was tested in
real EEG records acquired from eight subjects. The experimental
result show that the proposed method can effectively remove the
artifacts from all subjects.
Keywords—Noise, Artifacts, Principal Component Analysis,
EEG, BCI
I. Introduction
When brain cells (neurons) are activated, the synaptic
currents are produced within the dendrites. This current
generates a magnetic field measurable by electromyogram
(EMG) machines and a secondary electrical field over the
scalp measurable by Electroencephalogram (EEG)systems
[1].EEG is the noninvasive measurement of the electrical
activity on the scalp over multiple areas of the brain. The
measured of currents that flow during synaptic excitations of
the dendrites of many pyramidal neurons in the cerebral cortex
is called EEG signal. EEG signal, which is important in
clinical application(i.e., for diagnosing, monitoring, and
[2] latar managing neurological disorders) and in research field (i.e.,
belakang brain computer interface (BCI) application), is widely affected
prgf 1 by a variety of large signal contaminations or noise. Artifacts
can dramatically alter the signal recorded at all scalp sites,
especially those closest to the source of the noise [2-7]. Hence,
a necessary stage in EEG processing is artifact or noise
cancelation. Determining what is considered artifact, how
much artifact is excessive, and removing artifacts from real
data require a skilled human or a sophisticated artificial
system [7].
In a BCI application that removed all data containing
artifact might be left with too little clean data to be of practical
use. Therefore,the use of noise cancelation method to the EEG
records and leaving clean data would be of tremendous
978-1-4799-8069-7/15/$31.00 ©2015 IEEE
[2]
value(there has been an ample amount of research toward this
goal)[8-13]. Many of these newer approaches involve
techniques including independent component analysis, neural
networks, Kohonen maps, principal component analysis
(PCA),and other methods which were either unavailable or
much less well known during the early days of EEG signal
processing. Techniques for online cancelation of EEG noise
without rejecting other data are remains difficult in EEG
processing software but necessary one for BCI designers.In
this paper, we introduce a novel method (i.e., robust principal
component analysis (RPCA)) for automatic noise cancelation
which can significantly enhance prominence of the spike in
the clean EEG signals. This method is based on separation of
the raw multichanel EEG recordings into a relatively small
number of decorrelated linear combinations of a set of
random zero-mean variables while retaining as much of the
information from the original variables as possible.
II. DataAcquisition
In the experiment, eight healthy adult subjects (all men,
ranging in age between 20-22 old years) participated in this
study, which are 5 subjects have slight hair and the rest are
thick hair. All of the subjects do same experiment with three
stimuluswhich are baseline, closed eye, and blink eye
condition. It has been made sure that every participant was not
in a stress mental condition while the experiment was being
conducted. All participant picture show in Fig.1.During the
experiment, allparticipant were sitting in a comfortable chair
in front of 14” monitor at a distance of about 1 m. To build the
experiment setting used in the EEG sample collecting process,
we utilize the OpenVibe software to perform the data
acquisition, stimulus visualization, and EEG recording with
the function block that already built in. This experiment
consists of two sessions, first session had a period of 130
seconds with 2stimulus(i.e., baseline (eye normal) and closed
eye condition). Second sessions had a period of 66 seconds
with 2 others stimulus (i.e., baseline (eye normal) and blink
eye condition).The EEG signals are recorded continuously
using six electrodes (channels) at F7, F8, T7, T8, O1 and O2
that represent the visual of human brain and digitized at a 128
Hz sampling rate.The six electrodes configuration is shown in
Fig. 2.
1
550
 ∑ , (i=1,2,...,n) (2)

where , , .The
minimization of these cost function ns by the gradient descent
technique leads to an adaptive learn
ning algorithm

(3)

for any (0)=0, (i=1,2,...,n), where

w 0,
, ,…, , ,
. ., tanh ⁄ ) , . Usually, the
Figure 1. Eight subjects participated in the eexperiment. second term in the right hand side of
o Equation (3) is relatively
small and can be neglected. Th hus, yielding a simplified
version of the learning algorithm fo
or the extraction of the first
m PCs:

(4)

or in a discrete-time version as

1 , (5)

with 0 0, 0.

Figure 2. Six channels of electrode configuration in the experiment.

IV. Simulation and
a Analysis
Preparatory to an analysis of o the features of ERPs
III. Robust PCA components from EEG signals, actu ual signals were recorded in
a six-channel (F7, F8, T7, T8, O1 and a O2) configuration. The
Let m be the number of measured EEG siggnals and n be the raw data(see Fig. 2) were first prre-processed using using a
number of unknown input sources. Then, the m measured signals sixth-order band-pass filter with cu ut-off frequencies of 2 Hz
at channeli, , can be represented as a linnear combination (i.e., to remove the trend from low w frequency bands) and 16
of n unknown mutually statistically independeent source signals Hz (i.e., to remove unimportan nt information from high
, j=1, 2, ..., n, as followsindependent sourrce signals. frequency bands), respectively (see Fig. 3). Since the ERPs
∑ (1) power (signal) to the EEG powerr (noise) ratio is small, a
method of extracting and classify ying the ERPs component
from the EEG is desirable. One way y of gaining further insights
Through equation (1), the ith principal coomponent of the
into EEG signals is by applying RPCA techniques. In this
measured signals are extracted and next thessource signals is algorithm, the convergence speed is increased by employing
reconstructed from by using the transfoormation the moving average approach to enaable faster adaptation at the
.The weight vector , ,..., should separation step. The extracted sign nal and coresponding spike
be adaptively updated in such a way so thhat the estimated in 2D plots are given in Fig. 4 and Fig. 5, respectively. Fig. 4
vector will reconstruct the input measurred signals shows that after separation using proposed
p method, channels
as correctly as possible. The general learning algorithms are 1, 2 and 5 are still contaminated by y noise. In Fig. 5 corespond
optimal only for a Gaussian distribution of thhe input data and to Fig. 2, it is clear that channel 1 - 6 corespond to T8, T7, F7,
they are rather sensitive to impulsive noise oor outliers. Many F8, O1 and O2, respectively. Sincce channels 1, 2 and 5 as
approaches can be taken to increase the robbustness of PCA shown in Fig. 4 are still embended by b noise, therefore its spike
with respect to noise and outliers. F Firstly, outlying are not directly corespond to T8, T7 7, and O1, respectively. It is
measurements can be eliminated from thee data; secondly indicated that the consentration of the1st subject was effected
outliers can be suppressed or modified by replacing them with by surounding conditions such voice and any movement.
more appropriate values; and finally, more roobust criteria can
be applied. In order to derive a robust alggorithm, we can
formulate a cost function as[14]:

2
551
 Since the artifacts can randomly occur and are unexpected,
F8 they are difficult to identify. Thus, instead of detecting and
removing artifacts, our approach is to extract the event-related
T8
components based on a global pattern that encapsulates
C hannels

O2 models for signals of interest. Neuroscientists have found that,

O1
for the inhibition task, most event-related components project
strongly around Fz channel and have two fixed significant
T7 ERP patterns [15]. The signal before and after our artifact
F7 removal techniques are shown in Figs. 2 and 5 from the 1st
subject. As shown in the diagram, after artifact separation, a
0 500 μ volt 20 40 60 80 100 120 view of ERP is clearly reconizable in the each channels except
Time [s] channels P7and P8 (i.e., contaminated by noise or artifacts). It
st
Figure 2. Recorded EEG signals from 1 subject. demonstrates that the proposed algorithm can effectively
extract the ERP from even when there are no clearly visible
ERPs in the original signals.
F8

T8

O2 V. Conclusion
Channels

O1 In this paper, we described a robust PCA based moving

average algorithm to automatically extract event-related
T7
components and refine raw EEG signals. The preprocessed
F7
EEG signals were filtered using a sixth-order band-pass filter
with cut-off frequencies of 3 Hz and 13 Hz to remove to
0 20 40 60 80 100 120
remove the trend from low frequency bands and unimportant
Time [s] information from high frequency bands, respectively. Finally,
Fig. 3 Filtered EEG signals using BPF. the results using the proposed method illustrate the
effectiveness of the proposed algorithm removing the artifacts
and other non-even-related sources, and increasing the
6 visibility of the ERPs on all subjects.
5

4
Channels

Acknowledgment
3
This research was supported by the thematic program
2 through the Bandung Technical Management Unit for
Instrumentation Development (Deputy for Scientific Services)
1
and the excellent program through the Research Center for
Physics (Deputy for Engineering Sciences) funded by
0 1 μ volt 20 40 60
Time [s]
80 100 120
Indonesian Institute of Sciences, Indonesia.

Fig. 4 Extracted EEG signals using robust PCA based moving average. References

Ch1 Ch2 Ch3 [1] S. Sanei, J.A. Chambers, EEG Signal Processing, USA: John Wiley and
Sons, 2007.
[2] A. Jafarifarmand, M.A Badamchizadeh, “Artifacts removal in EEG
signal using a new neural network enhanced adaptive filter”,
Neurocomputing, vol. 103, pp. 222–23, 2013.
[3] G. Geetha and S. N. Geethalakshmi,“Artifact Removal from EEG using
Spatially Constrained Independent Component Analysis and Wavelet
Denoising with Otsu’s Thresholding Technique”, Procedia Engineering,
vol. 30, pp. 1064 – 1071, 2012.
Ch4 Ch5 Ch6
+ [4] S. O’Regan, S. Faul, and W. Marnane,“Automatic detection of EEG
artefacts arising from head movements using EEG and gyroscope
signals”, Medical Engineering & Physics, vol. 35, pp. 867– 874, 2013.
[5] K. J. Mullinger, J. Havenhand, and R. Bowtell,“Identifying the sources
of the pulse artefact in EEG recordings made inside an MR scanner”,
NeuroImage, vol. 71,pp. 75–83, 2013.
- [6] M. E. H. Chowdhury, K. J. Mullinger, P. Glover, and R.
Bowtell,“Reference layer artefact subtraction (RLAS): A novel method
Fig. 5 Coresponding spike of waveforms of extracted EEG signals (in Fig. 5).

3
552
 of minimizing EEG artefacts during simultaneous fMRI”, NeuroImage,
vol. 84,pp. 307–319, 2014.
[7] B. Allison, P3 or not P3: Toward a better P300 BCI. University of
California, San Diego, 2003.
[8] C. Babiloni, F. Vecchio, F. Infarinato, P. Buffo, N. Marzano, D. Spada,
S. Rossi, I. Bruni, P. M. Rossini, & D. Perani, “Simultaneous recording
of electroencephalographic data in musicians playing in ensemble”,
Cortex, 47(9), pp. 1082-1090, 2011.
[9] J. Escudero, R. Hornero, D. Abasolo, A. Fernandez, “Quantitative
evaluation of artifact removal in real magnetoencephalogram signals
with blind source separation”, Annals of Biomedical Engineering, 39(8),
pp. 2274-2286, 2011.
[10] W. O. Tatum, B. A. Dworetzky & D. L. Schomer, “ Artifact and
recording concepts in EEG”, Journal of Clinical Neurophysiology,
28(3), pp. 252-263, 2011.
4
553
[11] A. Turnip & K.-S. Hong, “Classifying mental activities from EEG-P300
signals using adaptive neural network”, Int. J. Innov. Comp. Inf.
Control, 8(7), 2012.
[12] A. Turnip, K.-S. Hong & M.-Y. Jeong, “Real-time feature extraction of
P300 component using adaptive nonlinear principal component
analysis”, BioMedical Engineering OnLine, 10(83), 2011.
[13] K. I. Diamantaras and S.Y. Kung, Principal Component Neural
Networks. Theory and Applications.Adaptive and Learning Systems for
Signal Processing, Communications and Control.John Wiley & Sons
Inc., New York, 1996.
[14] A. Cichocki and R. Unbehauen, “Robust estimation of principal
components in real time”, Electronics Letters, 29(21): pp. 1869–1870,
1993.
[15] M. Falkenstein, J. Hoormann, and J. Hohnsbein,“ERP components in
Go/Nogo tasks and their relation to inhibition”, Acta Psychologica, vol.
101, No. 2-3, pp. 267-291, 1999.