Suffering from dirty strong supersonic attacks (39Ϊ½)
γΊO1-V
oπ KΑͺ―Ά°(ΘΥΕ) Ϊ½x Ϊ[ρ
2: YAMAGUTIseisei 2019/04/24(
)09:17 ID:5ZbN1Z79Q(1/32) BE AAS
This is the html version of the file OΨέΈ:arxiv.org
. Google
Page 1
arXiv:1809.07356v1 [eess.SP] 19 Sep 2018
GENERIC COLORIZED JOURNAL, VOL. XX, NO. XX, XXXX 2017
1
Predictive Model for SSVEP Magnitude Variation:
Applications to Continuous Control in Brain-Computer Interfaces
Phairot Autthasan, Xiangqian Du, Binggwong Leung, Nannapas Banluesombatkul, Fryderyk K l, Thanakrit Tachatiemchan, Poramate Manoonpong, Tohru Yagi and Theerawit Wilaiprasitporn,
Member, IEEE
3: YAMAGUTIseisei 2019/04/24(
)09:17 ID:5ZbN1Z79Q(2/32) BE AAS
Abstract
The steady-state visual-evoked potential-based brain-computer interface (SSVEP-BCI) is a typically recognized visual stimulus frequency from brain responses.
Each frequency represents one command to control a machine.
For example, multiple target stimuli with different frequencies can be used to control the moving speeds of a robot.
Each target stimulus frequency corresponds to a speed level.
Such a conventional SSVEP-BCI is choice selection paradigm with discrete information, allowing users to discretely control the speed of a movable object.
This can result in non-smooth object movement.
To overcome the problem, in this study, a conceptual design of a SSVEP-BCI with continuous information for continuous control is proposed to allow users to control the moving speed of an object smoothly.
A predictive model for SSVEP magnitude variation plays an important role in the proposed design.
Thus, this study mainly focuses on a feasibility study concerning the use of SSVEP magnitude prediction for BCI.
Θ4
4: YAMAGUTIseisei 2019/04/24(
)09:20 ID:5ZbN1Z79Q(3/32) BE AAS
This work was supported by The Thailand Research Fund and Office of the Higher Education Commission under Grant MRG6180028, Junior Science Talent Project, NSTDA, Thailand.
T.Wilaiprasitporn,P.Autthasan,N.Banluesombatkul and B.Leung are with Bio-inspired Robotics and Neural Engineering Lab,School of Information Science and Technology,Vidyasirimedhi Institute of Science & Technology, Rayong, Thailand.
theerawit.w at vistec.ac.th Fryderyk K l is with Munich School of Engineering,Technical University of Munich,Munich,Germany.
P.Manoonpong is with Bio-inspired Robotics and Neural Engineering Lab,School of Information Science and Technology,Vidyasirimedhi Institute of Science & Engineering,Rayong,Thailand and Embodied AI & Neurorobotics Lab
, Centre for BioRobotics,The Msk Mc-Kinney Mller Institute,The University of Southern Denmark,Odense M,DK-5230,Denmark.
X. Du and T. Yagi are with Yagi Lab, Department of Mechanical Engineering, Tokyo Institute of Technology, Tokyo, Japan.
Thanakrit Tachatiemchan is with Department of Mathematics and Computer Science, Chulalongkorn University, Bangkok, Thailand
Θ2
5: YAMAGUTIseisei 2019/04/24(
)09:22 ID:5ZbN1Z79Q(4/32) BE AAS
I.
INTRODUCTION
THE dramatic decrease in the cost of electronic components and computational resources has made brain-computer interfaces (BCI) more fascinating to twenty-first-century researchers.
BCIs allow people to communicate with machines via brain responses or signals.
Consequently, the development of BCI-related technologies could benefit people with difficulty in executing motor functions [1].
Amyotrophic lateral sclerosis (ALS) is an example of such a disease.
BCI research focuses mainly on three types of brain responses: event-related potential (ERP), steady-state visual-evoked potential (SSVEP), and motor imagery (MI).
ERP and SSVEP are usually generated by visual, auditory or tactile stimulation of the human sensory system.
On the other hand, to generate MI signals, one has to imagine executing motor functions (such as hand or foot movements), without actually performing any movement.
The most widespread method for measuring brain responses is electroencephalography (EEG), mainly because it is non-invasive and somewhat cheaper in comparison to others.
Θ2
6: YAMAGUTIseisei 2019/04/24(
)09:23 ID:5ZbN1Z79Q(5/32) BE AAS
Out of the three aforementioned responses (ERP, SSVEP, and MI), SSVEP is the most practical since it is easy to obtain.
Recently, a research group conducted a series of experiments to answer the following question: How many (and what kinds of) people can use an SSVEP-based BCI? [2].
The experimental results of the research show that most participants could use an SSVEP-based BCI with acceptable accuracy, even if they had no prior experience with BCIs.
Participants were not annoyed by the flickering of the stimulus in any way.
Furthermore, the experiments were conducted in a noisy environment, confirming the practicality of SSVEP-based BCIs.
This study focuses on the exploitation of visual stimulation for SSVEP-based BCIs toward continuous and smooth brain-machine interaction which remains a challenging problem.
However, before discussing the specifics of this study, it is important to look at some of the milestones in SSVEP-based BCI research.
Page 2
2
GENERIC COLORIZED JOURNAL, VOL. XX, NO. XX, XXXX 2017
Θ2
7: YAMAGUTIseisei 2019/04/24(
)09:23 ID:5ZbN1Z79Q(6/32) BE AAS
In 2007, a state-of-the-art SSVEP recognition technique called canonical correlation analysis (CCA) was developed for use in SSVEP-based BCIs [3].
Then, in 2008, the application of SSVEP-based BCIs for controlling an electrical prosthesis was proposed by a pioneering BCI research group [4].
This system involved a typical choice selection SSVEP-based BCI to generate movements or gestures of a prosthetic arm.
Another pioneering group conducted a detailed study on the performance of SSVEP-based BCIs involving people with disabilities who cannot perform gaze shifting.
This group proposed stimulus patterns which could help such people use the system without gaze shifting [5].
In 2010, another research group focused on a user-friendly design for SSVEP-based BCIs and proposed a new algorithm named stimulus-locked inter-trace correlation (SLIC) [6].
The main concept behind SLIC is to combine ERP and SSVEP detection.
In 2012, a hybrid BCI system to improve the performance of BCIs was proposed.
A hybrid BCI makes use of various types of brain responses.
A hybrid ERP/SSVEP-based BCI was introduced for continuous, simultaneous, two-dimensional cursor control [7].
Θ5
8(1): YAMAGUTIseisei 2019/04/24(
)09:26 ID:5ZbN1Z79Q(7/32) BE AAS
The existing SSVEP-BCIs mainly rely on frequency recognition from EEG responses.
To develop a novel SSVEP-BCI paradigm for a brain-machine control system which allows users to continuously increase/decrease the moving speed of the application (i.e.
speed robot movements), this study hypothesizes that magnitude variation would help attain the goal.
Inspired by neuroscientific studies on human attention levels and SSVEP gains [13], feasibility studies are performed on the practicality of using SSVEP stimulus intensity to manipulate SSVEP magnitude.
In this experiment, the researchers varied the SSVEP stimulus intensity while keeping the stimulus frequency fixed.
Moreover, only a single-channel EEG is used here.
Using an experimental recorded EEG, the researchers conducted a comparative study of three predictive models for SSVEP magnitude variation.
Polynomial regression (Poly), random forest regression (RF), and neural network (NN) are proposed as potential models.
Leave-one-subject-out cross validation is performed to evaluate the mean square error (MSE) of prediction.
The results present that the predictive model for SSVEP magnitude variation using the RF approach outperforms both Poly and NN in terms of computational-time prediction with low MSE.
Θ3
9: YAMAGUTIseisei 2019/04/24(
)09:28 ID:5ZbN1Z79Q(8/32) BE AAS
The remainder of this paper consists of a section on the conceptual design of SSVEP-BCI with magnitude prediction (section II).
Section III presents the data acquisition and two experimental studies.
Finally, the results, discussion, and conclusion are contained in sections IV, V and IV, respectively.
II.
SSVEP-BCI WITH MAGNITUDE PREDICTION
In order to use a conventional SSVEP-BCI paradigm to control movable speed machines or robots, a visual stimulation must be designed as depicted in Figure 1(a).
There are seven target stimuli on a black screen, which are flickering at different frequencies.
The higher numbers represent higher levels of speed, and the stimulus with a hand icon is used to stop objects from moving.
In summary, users of the conventional SSVEP-BCI paradigm can constantly control movable speed machines by attending to the target speed number (target stimulus) on the screen.
However, by using a novel design to exploit the benefits of this paper (predictive models for SSVEP magnitude variation), the complexity of visual stimulation can be reduced as shown in Figure 1(b).
Θ6
10: YAMAGUTIseisei 2019/04/24(
)09:36 ID:5ZbN1Z79Q(9/32) BE AAS
2)
Due to the small number of stimulation frequencies, those unattended are less likely to irritate or disturb the user when focusing on the target stimulus.
Therefore, the frequency recognition rate of the CCA approach is unlikely to decrease.
Page 3
AUTHOR et al.: PREPARATION OF PAPERS FOR IEEE TRANSACTIONS AND JOURNALS (FEBRUARY 2017)
3
3)
Lower visual stimulation complexity can reduce eye fatigue in the user.
One important issue is that the conceptual design must use a predictive model for SSVEP magnitude variation to handle the system.
Thus, this paper covers the first step toward the SSVEP-BCI with a magnitude prediction.
Θ11
11: YAMAGUTIseisei 2019/04/24(
)09:36 ID:5ZbN1Z79Q(10/32) BE AAS
A.
Data Acquisition The participants of this experiment were ten healthy people aged between 20 and 25 (n = 10).
The experiments followed the Helsinki Declaration of 1975 (as revised in 2000), approved by the internal review board of the Tokyo Institute of Technology, Japan.
1)
EEG recording: In this study, an open source and low-cost EEG amplifier were used with a 250 Hz sampling rate, namely OpenBCI [14].
For practical purposes, a single-channel EEG (Oz) was used for recording data during all experiments.
2)
Stimulation protocol: To ensure practicality of the study outcomes in the continuing development of real-world applications, the experiments were conducted in a normal environment (a room without electromagnetic shielding).
The subjects were asked to sit in front of a 17-inch monitor, put their heads on a chin-rest 30 cm away from the screen, and pay constant attention to the center of the screen.
Figure 2 illustrates the SSVEP stimulus protocol.
Θ9
12: YAMAGUTIseisei 2019/04/24(
)09:37 ID:5ZbN1Z79Q(11/32) BE AAS
Fig.2:
Four stimulus conditions presented randomly to the subjects, each lasting for 50 seconds.
A black screen and a conditional cue were both shown for four seconds each before the beginning of every condition.
2)
The same square starts flickering at an intensity level of 105.
The light intensity is then increased by three levels per second for 50 seconds.
This condition is supposed to help the subjects increase their SSVEP magnitude (cond.2).
3)
The same square starts flickering at the maximum intensity (225).
The light intensity is then decreased by three levels per second for 50 seconds.
Θ8
13: YAMAGUTIseisei 2019/04/24(
)09:44 ID:5ZbN1Z79Q(12/32) BE AAS
B.
Experiment I:
Predictive Model for SSVEP Magnitude Variation
1)
Data Preparation:
From the aforementioned SSVEP responses of 10 subjects, each with four conditions, EEG signals were randomly selected from nine subjects for a training set and the one remaining for the test set (1 subject ~ 4 conditions ~ 50-second long data).
As shown in Figure 3, the EEG signal from each subject was calculated for two sets of data.
The first set consisted of input signals (a).
The EEG signal from each subject was first converted into a sequence of subsamples or blocks with a three-second (3s ~ 250Hz = 750 data points) sliding window and a two-second overlap.
Page 4
Θ3
14: YAMAGUTIseisei 2019/04/24(
)09:44 ID:5ZbN1Z79Q(13/32) BE AAS
After step 1 was completed, 48 blocks of data were obtained for the following structure: 4 conditions ~ 48 blocks ~ 750 data points or features.
Then, for step 2 , each data point was squared to obtain the input signals of this subject.
The second set of data consisted of target signals (b).
Starting from 3 , the average value of each block was calculated from the input signals.
Hence, 48 values were obtained for each condition.
After that, in step 4 , curve fittings were performed with polynomial functions (either a quadratic function (poly2) or a cubic function (poly3)) for each condition.
Finally, the target signals were 4 conditions ~ 48 target values (four curves with 48 points each).
After the data preparation was completed, the input signals and target signals of each subject were used for SSVEP amplitude prediction using the following approaches.
2)
Neural Networks Approach (NN): A recurrent neural network (RNN) is extended from a conventional feed-forward neural network and has the ability to extract essential features from time series data, such as EEG, due to its recurrent hidden state.
Θ4
15: YAMAGUTIseisei 2019/04/24(
)09:45 ID:5ZbN1Z79Q(14/32) BE AAS
After step 1 was completed, 48 blocks of data were obtained for the following structure: 4 conditions ~ 48 blocks ~ 750 data points or features.
Then, for step 2 , each data point was squared to obtain the input signals of this subject.
The second set of data consisted of target signals (b).
Starting from 3 , the average value of each block was calculated from the input signals.
Hence, 48 values were obtained for each condition.
After that, in step 4 , curve fittings were performed with polynomial functions (either a quadratic function (poly2) or a cubic function (poly3)) for each condition.
Finally, the target signals were 4 conditions ~ 48 target values (four curves with 48 points each).
After the data preparation was completed, the input signals and target signals of each subject were used for SSVEP amplitude prediction using the following approaches.
2)
Neural Networks Approach (NN): A recurrent neural network (RNN) is extended from a conventional feed-forward neural network and has the ability to extract essential features from time series data, such as EEG, due to its recurrent hidden state.
Θ4
16: YAMAGUTIseisei 2019/04/24(
)09:46 ID:5ZbN1Z79Q(15/32) BE AAS
In order to train the NN model with the data, as shown in Figure 3, the input signals were reshaped (a) into number of samples ~ time steps ~ features.
The researchers decided to consider each block in each condition as one sample.
Moreover, since the 750 data points in each sample actually represented time series data of three seconds, the number of time steps was 750.
In addition, instead of using only the value of each data point as a feature, its block ID was also used, ranging from 0 to 47 in each condition.
For example, each time step had two features including its data point value and block ID.
Therefore, for each subject, 192 samples ~ 750 time steps ~ 2 features were obtained.
For the target signals (b), each block in each condition was also considered as one sample.
Therefore, these were reshaped into 192 samples ~ 1 target value for each subject.
The NN model in this study was implemented using Keras [16] which was tuned until the model gave the best parameter configurations, as set out below:
A layer of GRU with 256 units.
Θ12
17: YAMAGUTIseisei 2019/04/24(
)09:46 ID:5ZbN1Z79Q(16/32) BE AAS
Fig.4:
Illustration of an experimental protocol for an online-like brain-controlled robot.
The experimental protocol is designed to demonstrate the advantages of the proposed case over the conventional SSVEP-BCI.
The protocol begins with an increasing speed period, followed by the constant maximum speed, decreasing speed, and the constant minimum speed period.
The final period stopped the movement.
Afterward, the prediction is performed using majority voting or averaging the results from these decision trees.
In order to insert the data into the RF model as shown in Figure 3, each block of input signals was reshaped into one sample (a) as in the previous approach.
Therefore, from each subject, 192 samples were obtained (4 conditions ~ 48 blocks) with 750 features per sample.
After that, its block ID was added, ranging from 0 to 47 for each condition, as the last feature.
Consequently, the final shape of input was 192 samples ~ 751 features.
Θ10
18: YAMAGUTIseisei 2019/04/24(
)09:50 ID:5ZbN1Z79Q(17/32) BE AAS
5)
Evaluation: Leave-one-person-out cross validation was used to train and evaluate all models.
Thus, 10 folds, each consisting of nine subjects for training (1,728 samples) and one subject for testing (192 samples).
Since there were two types of target signal, this study used the following predictive models: Poly poly2, Poly poly3, RF poly2, RF poly3, NN(GRUs) poly2 and NN(GRUs) poly3 for comparison.
Therefore, the performance of each approach was measured using two values.
The first was the accuracy of each model, calculated using MSE, and the second its computational-time for the prediction.
To compare these three approaches, the one-way repeated measures analysis of variance (ANOVA) was used, based on the assumption of sphericity (statistical analysis of the experimental results).
Correction was applied when the data violated the sphericity assumption.
Bonferroni correction and pairwise comparison were performed for post hoc analysis.
19: YAMAGUTIseisei 2019/04/24(
)09:51 ID:5ZbN1Z79Q(18/32) BE AAS
C.
Experiment II: Brain-Controlled Robotic Simulator In this part, the researchers aim to demonstrate the advantages of the proposed SSVEP-BCI over the conventional case via a brain-controlled robotic stimulator.
As in the results of Experiment I, the RF poly2 model was found to be the most suitable predictive model for SSVEP magnitude variation, in terms of both small error and short computation-time for control applications.
Hence, only the RF poly2 approach was incorporated into the V-Rep simulator [18].
The Vortex physics engine mode inside the V-Rep was constructed to evaluate the concept of the proposed SSVEP-BCI against the conventional one.
Here, a modified Pioneer P3DX robot was used, equipped with an 80 ~ 80 cm plate and a 10 cm3 cube with 60 kg on top of the robot.
In order to set the robotic simulator for an online-like study, the velocity profiles (VPs) were created following the experimental protocol (Figure 4).
Using the gathered datasets from Experiment 1, the predicted signals from cond.2 and cond.3 were connected to online-like SSVEP responses (the brain signals).
These brain signals were then used as speed controller inputs.
The controller applied the standard moving average (MA) algorithm to the brain signals.
Θ4
20: YAMAGUTIseisei 2019/04/24(
)09:54 ID:5ZbN1Z79Q(19/32) BE AAS
The numbers in bold are significantly higher than the others, p<0.01.
Finally, the results of the MA are used as VPs to vary the moving speed of the robot.
The controller worked alternatively between the two states according to the experimental protocol for the increasing and decreasing speed periods, using the following rules:
An increasing speed period: the current velocity value would be updated if the incoming value was higher, otherwise the value remains stable.
A decreasing speed period: the current velocity value would be updated if the incoming value was lower, otherwise the value remains stable.
21: YAMAGUTIseisei 2019/04/24(
)09:54 ID:5ZbN1Z79Q(20/32) BE AAS
The aforementioned experimental protocol was used to set up the brain-controlled robotic simulators to explore the performance of the proposed SSVEP-BCI in Studies I and II, as shown in the following subsections.
1)
Study I:
Trade-off between processing window length and smooth movement: The aim of this study is to find the most suitable processing window length for the speed controller to help users move the object smoothly at an acceptable speed.
The experiment protocol is shown in Figure 4.
The robot was supposed to move following the outputs of the speed controller when completing a transportation task.
To create the outputs, the brain signals were converted into velocity profiles (VPs) using both the RF poly2 predictive model and the simple MA in accordance with two rules, as explained earlier.
The researchers considered both the ability to maintain stability in the carrying of objects (or the box in this experiment) and the average moving speed.
The stability of the box was measured by the deviation of a central mass in the box on the robotic plate space (2D plane).
The deviation of the box in this study was the Euclidean distance between the original and final positions of the box on the 2D plane.
Θ3
γΊO1-VΦΚΒυέxπ
½Ϊξρ ΤΪ½o ζΪ½o πΜ’Η½Ϊ AA»ΡΘ²Ω
Κ±Μθ Κ±TOP 0.278s*