主体间模型可减少基于运动想象力的脑机接口的校准时间资源-CSDN文库

104 浏览量 2021-04-11 18:42:03 上传评论收藏 1.35MB PDF 举报

在本文中，作者探讨了一种新型框架，旨在通过使用其他主体的脑电图（EEG）信号知识，来减少基于运动想象的脑机接口（BCI）校准时间。BCI技术允许用户通过大脑活动直接与计算机或其他外部设备交互。主要使用的脑活动类型包括脑电图（EEG）、皮质电图（ECoG）、脑磁图（MEG）、功能磁共振成像（fMRI）和功能性近红外光谱成像（fNIR）。在众多脑信号类型中，基于EEG的BCI系统越来越多地被采用。目前，长时间的校准过程是阻碍BCI实际应用的一个主要因素。为了获得足够的训练试验，参与者必须在校准阶段花费很长时间。为了解决这一问题，研究者们提出了一个新的框架，即利用其他主体的EEG信号知识来减少校准时间。文章通过以下步骤详细介绍了这一框架： 1. 使用目标主体的EEG信号和其他主体的EEG信号来训练运动识别模型。 2. 提出了两种数据映射方法，分别减少了由于不同主体之间大脑激活区域的差异以及大脑激活强度的不同而引起的变异。 3. 在这些数据映射阶段之后，采用集成方法将所有主体的EEG信号聚合成最终模型。为了验证所提出方法的有效性，作者将它与其他减少校准时间的方法进行了比较。结果显示，所提出的方法使用非常少的训练试验（32个样本）就能达到令人满意的识别准确率。与使用少数训练试验的现有方法相比，所提出的方法实现了更高的准确率。关键词包括：脑机接口（BCI）、脑电图（EEG）、机器学习、运动想象、共同空间模式和主体间模型。文章的介绍部分指出，BCI技术允许用户通过大脑活动与计算机或其它外部设备直接交互。所使用的主要大脑活动类型包括脑电图（EEG）、皮质电图（ECoG）、脑磁图（MEG）、功能磁共振成像（fMRI）和功能性近红外光谱成像（fNIR）。在众多脑信号类型中，基于EEG的BCI系统因其无创性、便携性、成本低廉和非侵入性，被越来越广泛地采用。然而，BCI技术的一个主要应用障碍是长校准时间，即参与者在进行校准阶段时必须花费大量时间。为了克服这一挑战，作者在研究中提出了通过集成其他受试者的脑电波知识来训练目标受试者的运动识别模型，以减少校准所需的训练试验数量。作者还强调了集成学习方法在整合不同受试者EEG数据中的重要性，并通过实验证明了该方法在减少校准时间的同时提高了识别准确率。

资源推荐

资源详情

资源评论

ORIGINAL ARTICLE

An inter-subject model to reduce the calibration time for motion

imagination-based brain-computer interface

Yijun Zou

& Xingang Zhao

& Yaqi Chu

1,2

& Yiwen Zhao

& Weiliang Xu

& Jianda Han

Received: 15 November 2017 /Accepted: 19 October 2018

International Federation for Medical and Biological Engineering 2018

Abstract

A major factor blocking the practical application of brain-computer interfaces (BCI) is the long calibration time. To obtain

enough training trials, participants must spend a long time in the calibration stage. In this paper, we propose a new framework to

reduce the calibration time through knowledge transferred from the electroencephalogram (EEG) of other subjects. We trained

the motor recognition model for the target subject using both the target’s EEG signal and the EEG signals of other subjects. To

reduce the individual variation of different datasets, we proposed two data mapping methods. These two methods separately

diminished the variation caused by dissimilarities in the brain activation region and the strength of the brain activation in different

subjects. After these data mapping stages, we adopted an ensemble method to aggregate the EEG signals from all subjects into a

final model. We compared our method with other methods that reduce the calibration time. The results showed that our method

achieves a satisfactory recognition accuracy using very few training trials (32 samples). Compared with existing methods using

few training trials, our method achieved much greater accuracy.

Keywords Brain-computer interface (BCI)

Electroencephalogram (EEG)

Machine learning

Movement imagi nation

Common spatial pattern

Inter-subject model

1 Introduction

Brain-computer interfaces (BCIs) are a type of technology

that allows users to interact directly with computers or other

external devices through brain activity. The main brain activ-

ities that are used are electroencephalography [1], electrocor-

ticography (ECoG) [2], magnetoencephalography (MEG),

functional magnetic resonance imaging (fMRI) [3], and func-

tional near-infrared (fNIR) imaging [4]. Of the various types

of brain signals, the EEG-based BCI system has been increas-

ingly used because of its non-invasiveness, higher time reso-

lution, ease of use and lower price. The EEG-based BCI is a

promising technology for providing a method of communi-

cation for disabled users to control prostheses and wheel-

chairs or for guiding rehabilitation treatments [5–8]. BCIs

are also used by healthy individuals as a valuable tool for

monitoring physical status [9]orasawayofoperatingan

entertainment system. The EEG signal modalities used in

BCI includes steady-state visual evoked potential (SSVEP)

[10, 11], slow cortical potential (SCP), P300 [8], and motor

imagery (MI) [12]. Compared with other signal modalities,

MI has the advantage of self-pace, which means no external

stimulation is needed. For severely disabled patients, motor

imagery based BCI may be the most feasible communication

system. However, most of the current BCI applications are

still in the prototype stage. Few BCI systems can be used

outside the laboratory because the recognition accuracy does

not meet the needs of the actual systems, the information

transformation rate is limited, and the BCI system requires a

long calibration time [13].

The calibration time can be reduced by using semi-

supervised learning or data augmentation on a small training

dataset to obtain better motion recognition accuracy and by

using EEG data from other subjects to train a model for the

task of the target subject.

* Xingang Zhao

zhaoxingang@sia.cn

University of Chinese Academy of Sciences, Beijing 100049, China

Key Laboratory of Networked Control System, Shenyang Institute of

Automation, Chinese Academy of Sciences, Shenyang 110016,

China

Department of Mechanical Engineering, University of Auckland,

Auckland, New Zealand

Medical & Biological Engineering & Computing

https://doi.org/10.1007/s11517-018-1917-x

Semi-supervised learning is an important method that uses

an unlabelled dataset to improve the final result of the learning

algorithm. Many semi-supervised learning methods have been

proposed in BCI. Tu et al. [14] used the unlabelled dataset to

generate regular items for the spatial filter, which makes the

model more robust and adaptive to the testing set. Jiannan

Meng et al. [15] first developed a base model from the tagged

dataset and then applied it to the unlabelled dataset. Based on

the results and linear projection score of the unlabelled dataset,

they produced a new model. Chen et al. [16] designed a new

self-training method to select the most effective unlabelled

data for the extended training set. C ompared with classic

methods, these methods improve the classification perfor-

mance with a shorter calibration time. However, the improve-

ment is not obvious. Moreover, these methods focus on the

classifica tion of two types of movements. These methods

would be better evaluated in a four category motion imagery

situation.

In addition to the semi-supervised approach, other methods

have also been proposed to improve accuracy when using a

small training set. Lotte et al. [13, 17] designed a method of

generating artificial signals from an existing training set to

construct a large datas et. Lotte’s method obtained a much

better result than the semi-supervised study. However, this

result was also obtained using the two-motor classification

approach.

Limited by the size of the training dataset, the current

method that directly reduces the calibration time using the

user ’s own training data does not perform well. It is very

difficult to achieve good training results just by using a

small number of samples (such as 10 samples per type of

action). Therefore, using data from other subjects to im-

prove t he recognition result is an importa nt alternative

way of training the BCI.

The manner in which to take advantage of datasets from

other subjects is of increasing interest to researchers. The two

main ways to do so are through knowledge transfer and data

fusion. Knowledge transfer refers to the transfer of knowledge

from an existing dataset to new individuals. In the BCI sys-

tem, compared with the traditional common spatial patter

(CSP), the regularized common spatial pattern (RCSP) is an

effective method of knowledge transfer. The RCSP reduces

the over-fitting of the CSP filters based on the existing knowl-

edge [18–20]. Lu et al. [21] directly used the covariance of

EEG data from multiple subjects to develop a regularization

term. H. Kang et al. [22] adopted a weighted average method

to develop CSP regularization terms. Lotte et al. [13, 23, 24]

evaluated the distance between the covariance of EEG data

from different individuals using the Riemann distance and

designed the regular term of the RCSP based on this distance.

Tu et al. [25] calculated the CSP filters from multiple subjects

and applied a cross-validation method to obtain the best one

from the CSP filters from other people.

In addition to knowledge transfer, data fusion is another

way to use data from multiple subjects. The two main types

of fusion methods are the pooled data method and the ensem-

ble method. The pooled data method establishes the training

set directly from the data of multiple subjects. The ensemble

method establishes a model for the data from each subject and

then assembles these models into a final model. The key prob-

lem with each of these approaches is how to combine the

results of these models. Fazli et al. [26] divided the EEG signal

from multiple subjects into different frequency bands and then

trained the classification models for each of the frequency

bands. By choosing the optimal frequency bands for each

dataset, they established a user-independent motion recogni-

tion model. Tu et al. [27] designed a meta-classifier for the

results of each dataset and then assembled the outputs of the

models from multiple subjects.

However, these methods only take data from other subjects

as a reference, constrain criteria for the new subject’smodel,

or directly establish a user-independent model with a dataset

from other subjects. The former methods have a shortage of

insufficient information mining, while the latter methods ig-

nore the essen tial differences among different in dividuals.

Moreover, these methods are all evaluated in a two-category

motor classification problem. Once the type of motion in-

creases, these methods can no longer address the noises and

biases. To solve these problems, this paper focuses on the

analysis of the differences in the EEG signals from different

bjects. We eliminated the differences as much as possible

and exploited the commonalities between the EEG signals

from different individuals. When people execute or imagine

a motion, the corresponding area in the brain is activated.

These event-related phenomena present variations on specific

frequencies of the ongoing EEG activity. These variations

may either be a decrease or an increase in power in the given

frequency bands. The former case is called event-related

desynchronization (ERD) and the latter is called event-

related synchronization (ERS) [28]. In our research, the dif-

ference between the EEG signals from each user was divided

into the following: 1) the mismatch of the channel position for

different subjects; 2) the intensity of the difference of ERD/

ERS, and 3) the frequency band differences of ERD/ERS.

This study focused on eliminating the first two differences.

The commonly used bandpass filter effectively reduces the

impact of ERD/ERS band differences. Furthermore, for the

purpose of making full use of the data from multiple subjects,

we did not use each dataset separately but combined datasets

from multiple subjects and the data from the target user to-

gether. Hence, the final model has both the target user’sown

characteristics and the characteristics of multiple other sub-

jects. In addition, a confidence score for the fusion of models

from different users was developed. The confidence score was

based on the linear projection and projection variance of the

LDA. By using the developed confidence score, the final

Med Biol Eng Comput

model was a combination of the models of multiple subjects.

The key points of the proposed methods are two pre-

processing algorithms and a framework to address the data

from multiple subjects. All these methods do not work with

fixed feature extraction methods and classifiers. However,

even though we used CSP and LDA, our framework can also

be combined with other feature extraction methods and clas-

sifiers to get an even better performance.

2Methods

2.1 Dataset description

To analyze the effectiveness of the proposed method, we ex-

perimentally validated it with a public dataset. The dataset we

used in this study is dataset 2a from BCI competition IV

(http://www.bbci.de/competition/iv/). The dataset is provided

by the Graz’s group and contained data from nine subjects. All

the subjects were naïve subjects who were unfamiliar with the

experiments. Each subject performed two experiments in

2 days. In each session, the subject was instructed to carry

out four types of motion imagery tasks in a random order,

namely, left hand (class 1), right hand (class 2), foot (class

3) and tongue (class 4). Each subject performed 72 trials of

each motion, yielding a total of 288 trials per session. At the

beginning of each trial, a motion hint appeared on the screen.

Then, the subject imagined the desired movement. After 4 s,

the stop hint appeared. The EEG signals were recorded with

22 channels and digitized at 250 Hz. All the channels were

used in our experiments. In BCI competition IV, the

participants were required to use an EEG signal from each

subjects’ first day to build a model and then validated it on

the data from the second day. In our study, we used the first 32

trials of each experiment to build a model and tested it on the

remainder trials of the same experiment.

2.2 Framework

ERD/ERS provides a solid theoretical basis for motion

imagery-based BCI. Even so, the motion recognition accuracy

cannot be easily improved directly by theoretical analysis. For

a specific individual, the BCI system still requires many train-

ing data to construct the model based on machine learning

algorithms. This is because the EEG signal itself is complex

and there are lots of background noises. Therefore, it is nec-

essary to obtain a large training dataset to extract the target-

related components from the complex EEG signals. Second,

in principle, EEG signals from different individuals have sim-

ilar qualitative characteristics, but for a specific individual, the

quantitative characteristics reveal great differences. This inter-

subject variation prevents the construction of an inter-subject

model. We know that there is energy enhancement or

attenuation when people execute a motion. This is the quali-

tative analysis. However, before we can access the EEG sig-

nals, we do not know the level of enhancement or the location

of the strongest frequency band. Additionally, there is no for-

mula to compute it. This shortage prevented us from classify-

ing each trial by direct logical reasoning. This is why we had

to use machine learning, which requires many trials. Thus, it is

difficult to decrease the calibration time.

Here, we analyzed these inter-subject differences and com-

bined the signals from all the subjects. The framework of our

method is shown in Fig. 1. The workflow was to first process

the EEG signals from each subject according to the target

subject’s EEG signal. Then, models were generated from each

subjects’ processed signals and these models were assembled

into a final model, which was a model for the target subject.

We first analyzed the variation between the EEG data from

different subjects. The ERD/ERS gives a qualitative descrip-

tion and we could not obtain a quantitative analysis before we

accessed the EEG signal. In fact, the variation in the ERD/

ERS between individuals occurs because of the ERD/ERS

frequency band, its intensity and its location. ERD/ERS oc-

curs at different frequencies in different people. In general, the

mu and beta rhythm band (8–30 Hz) is the most important

frequency band of the ERD/ERS. However, the most intense

frequency band of the ERD/ERS varies from person to person.

ERD/ERS refers to an energy enhancement or attenuation for

a specific freque ncy band in a particular period of time.

However, there is no quantitative assessment of the enhance-

ment or attenuation, which also varies among individuals.

ERD/ERS occurs in different locations of the cap. Generally,

the C3 and C4 channels correspond to the right hand and left

hand movements, respectively. The corresponding location is

typically determined by measurements in each participant.

Taking the physiological variation of each individual’sbrain

into account, the typical position of ERD/ERS is different in

different people. Additionally, how different participants wear

the cap may also cause location bias. Therefore, to establish an

inter-individual EEG model, these inter-individual variations

must be considered.

Here, we reduced inter-individual variations and extracted

the common characteristics for different individuals. The key

process of this method was to modify the datasets from multiple

subjects according to the data from the target individual. By

comparing the differences between the two datasets, the dataset

from multiple subjects was pre-processed to minimize the dif-

ferences and maximize the commonalities. The data processing

stage is divided into two steps: spatial filter and data adjustment.

The spatial filter step eliminated variations caused by the chan-

nel position. The data adjustment process step suppressed var-

iations in the degree of the energy decrease or increase. After

the data were pre-processed, models based on CSP and LDA

were trained using data from each subject in the dataset. A final

model was obtained by comparing these models.

Med Biol Eng Comput

剩余13页未读，继续阅读

评论收藏

内容反馈

weixin_38654915

粉丝: 7
资源: 995

主体间模型可减少基于运动想象力的脑机接口的校准时间

脑机接口中非平稳神经解码的动态集成建模方法

基于准模型校准卡尔曼滤波的巡检机器人运动系统辨识研究.pdf

脑机接口总体愿景与关键技术研究报告.docx

matlab-EEG运动想象分类-左右手运动信号

基于stm32的获取校准时间项目

航模发动机运动仿真模型

基于C#做的时间校准工具，可以从NTP获取时间，同时自动更新到本地电脑

Matlab 新能源汽车Simulink模型+纯电模型+混动模型+燃料电池模型+仿真

基于ERP的脑机接口的时空判别分析

基于STM32的电动洗胃机校准装置的设计与实现.pdf

【基于Matlab】利用运动结构和Levinson方法校准Velodyne相机系统的Matlab方法

基于校准导航的广域电力系统负荷模型校准

脑电信号左右手运动想象识别.rar

北京时间校准工具

基于深度学习的公文自动校对系统.pdf

迷你时间校对程序

基于雌穗粒位间吐丝时间差异的玉米籽粒生长模型

基于MATLAB的旋耕机运动仿真分析.pdf

风力机模型的仿真

基于EKF二阶RC模型电池Soc预测仿真

基于Matlab模块化模型的水轮发电机组仿真研究.zip

无线传感网络主从时间同步模型及算法实现.zip

基于MATLAB仿真模拟的单污染源扩散模型的优化研究.pdf

时间校准.bat 校准Windows时间

最新资源