CNN实现的运动想象脑电分类及人-机器人交互

基于脑电的脑机交互能帮助肢体运动障碍患者进行日常生活和康复训练,但是,由于脑电信号存在信噪比较低、个体差异性大等问题,导致脑电特征的提取与分类还需要进一步提高准确性和效率.因此,在减少脑电采集通道数目、增加分类数目的前提下,基于卷积神经网络对运动想象中的脑电信号进行分类.首先,基于已有方法进行探索实验,建立由3层卷积层、3层池化层和2层全连接层构成的卷积神经网络;然后针对想象左手、右手、脚的运动和静息态设计与开展了实验,获取了相关脑电数据;之后,利用脑电数据训练出基于卷积神经网络的分类模型,测试结果表明,该模型平均分类识别率达到了82.81%,且高于已有的相关分类算法;最后,将已建立的分类模型应用于运动想象信号的在线分类,设计与开发了脑机交互应用原型系统,驱动人-机器人之间的实时交互,帮助用户利用运动想象控制仿人机器人的抬手、前进等运动状态.进一步的测试结果表明,机器人对用户控制命令的平均识别率达到了80.31%,从而验证了所提方法可以对运动想象脑电数据进行较为精确的实时分类,可以促进脑机接口技术在人-机器人交互中的应用.     

多任务运动想象脑电特征的融合分类研究

针对运动想象脑-机交互任务模式单一、识别精度低、实用性较差等问题,采用改进的共空间模式（CSP）的特征提取方法,并利用支持向量机（SVM）与CSP融合分类方法对多类任务运动想象脑电特征进行分类识别。首先,选择特定导联上的脑电信号进行小波分解与重构,去除冗余信息;其次,利用特征参数做差的方法,得到较为明显的脑电特征;最后,通过SVM融合CSP的分类模式,对脑电特征进行多任务分类。利用BCI竞赛数据,对左手,右手,舌和脚四类运动想象任务的脑电进行识别。结果表明：分类正确率最高达到90.9%,平均正确率为86.8%,Kappa系数为0.8867,信息传输速率可达0.68 bit/trial,能够有效的获得脑电特征并较好的实现多任务运动想象脑电识别。     

基于HHT运动想象脑电模式识别研究

脑机接口是一种变革性的人机交互, 其中基于运动想象(Motor imagery, MI)脑电的脑机接口是一类非常重要的脑机交互. 本文旨在探索有效的运动想象脑电特征模式提取方法. 采用在时域、频域同时具有很高分辨率的希尔伯特--黄变换(Hilbert-Huang transform, HHT),进而提取自回归(Auto regressive, AR)模型参数并计算运动想象脑电平均瞬时能量,从而构造特征向量, 最后利用能较好地适应运动想象脑电单次试验分类的支持向量机(Support vector machine, SVM)进行分类. 结果表明在Trial的5.5～7.5s期间, HHT特征提取方法平均分类正确率为81.08%, 具有良好的适应性;最高分类正确率为87.86%, 优于传统的小波变换特征提取方法和未经HHT的特征提取方法;在Trial的8～9s期间, HHT特征提取方法显著优于后两种特征提取方法. 本研究证实了HHT对运动想象脑电这一非平稳非线性信号具有很好的特征提取能力, 也再次验证了运动想象事件相关去同步(Event-related desynchronization, ERD)现象, 同时也表明运动想象脑电的脑--机交互系统性能与被试想象心理活动的质量密切相关. 本文可望为基于运动想象脑电的在线实时脑机交互控制系统的研究打下坚实的基础.     

基于ECoG的运动想象脑-机接口分类方法

脑—机接口BCI(Brain-Computer Interface)技术是近年来国际上的研究热点之一,它通常利用脑电EEG(electroencephalo-gram)来实现无动作的人机交互,运动想象是其中一种重要的BCI实验范式。有关研究表明,脑皮层电位ECoG(electrocorticogram)具有更好的信噪比与频带特性。研究基于ECoG的运动想象BCI系统,针对ECoG信号的特点,改进了信号处理方法,提取数据的公共空间模式CSP(Common Spatial Pattern)特征,并利用支持向量机SVM(Support Vector Machines)进行分类器设计,提高了运动意向的识别正确率。用相应方法处理2005年脑-机接口竞赛中的一组实验数据,正确率达到92%,相比于当时参赛时所用的方法提高了6%。实验还发现,支持向量机在克服"维数灾难"和"过拟合"方面具有更好的鲁棒性。     

基于SSVEP与运动想象的实时脑控阿凡达系统

针对当前在线脑机接口交互系统识别正确率比较低、模式单一、算法复杂度高等问题,设计基于SSVEP和运动想象的实时脑控实时算法并应用于脑控阿凡达系统。通过对6名受试者进行离线训练和在线测试,实验结果表明该系统较好地反映受试者的控制意图,所有受试者均可以基本准确地控制机器人,可实现多指令的控制,识别准确率高,可进一步推动脑机接口在实际领域的应用和发展。     

基于模糊化符号复杂度的脑电运动想象识别算法

提出一种基于模糊化符号复杂度的运动想象脑电信号特征提取与识别方法。在脑电信号的复杂度细粒化多符号度量中引入模糊算法,用sigmoid函数模糊化处理,逻辑判断得到模糊化符号复杂度。取细粒化指数n为2,提取模糊化符号复杂度作为特征值,最后利用支持向量机对脑电运动想象任务进行分类识别。实验结果表明,以模糊化符号复杂度为特征的分类方法,对左右手运动想象脑电信号的分类识别率最高达88.67%,优于二值化Lempel-Ziv复杂度算法。     

运动想象脑电的块选择共空间模式特征提取

共空间模式(CSP)作为一种空间滤波方法已在脑电信号(EEG)的特征提取上得到了广泛应用,而对脑电信号的通道和频带进行合理选择可以有效改善共空间模式特征在运动想象脑机接口(BCI)中的分类性能.针对已有选择方法中未充分考虑通道间差异性的问题,本文提出一种对通道和频带同时进行选择的块选择共空间模式(BS–CSP)特征提取方法.首先针对每个通道进行频带划分从而构建数据块,然后根据时频特征计算对应的Fisher比表征每个块的分类能力,并设置阈值选出一定数量的最优块,最后用CSP和支持向量机(SVM)分别进行特征提取与分类.在对BCI Competition Ⅲ Datesate Ⅳa和BCI Competition Ⅳ Datesate Ⅰ两个二分类运动想象任务的分类实验中,平均分类精度达到了90.25%和83.78%,表明了所提出的特征提取方法的有效性和鲁棒性.     

基于SSVEP直接脑控机器人方向和速度研究

直接用思维意图来控制机器人而没有大脑外周神经和肌肉的参与是人类的一个梦想,目前这一研究已成为国际前沿热点和突破点.传统的脑控机器人（Brain-controlled robot,BCR）主要控制其方向,而本文旨在探讨能够同时脑控机器人方向和速度的有效方法.采用可分类目标数多、单次识别率高且训练时间短的稳态视觉诱发电位（Steady state visual evoked potentials,SSVEP）脑机交互（Brain-computer/machine interaction,BCI/BMI）方法,为脑控机器人运动规划了向左、向右、前进和后退4个方向,设计了低速、中速和高速3级运动速度并组合了9个脑控指令;进而比较并优化了SSVEP刺激目标布局间距以及刺激目标闪烁时间,采用典型相关分析（Canonical correlation analysis,CCA）进行识别.结果表明恰当设置SSVEP刺激目标数及其布局间距和刺激目标闪烁时间,可以有效提高被试/用户直接脑控机器人的性能;优化的SSVEP刺激范式三结合适应SSVEP解码的典型相关分析,8名被试脑控机器人到达终点平均用时为2分40秒,最少用时1分29秒;同时,在脑控机器人运动过程中触碰障碍平均次数为0.88,最少碰触次数为0.本研究显示基于SSVEP的脑机交互可以作为直接脑控机器人灵活运动的一种可选方法,能够实现对机器人多个运动方向和多级速度的控制;也证实了适当增加刺激目标间距可以有效提高SSVEP-BCI脑控指令识别的正确率,说明了该脑控方法的性能与刺激被试的范式有关;再次验证了CCA算法在基于SSVEP的脑机交互中具有优良的效果.最后,为克服单一SSVEP范式存在的局限,本研究也尝试把该范式与运动想象相结合的混合范式用于脑控机器人方向和速度,并进行了初步的研究,表明可以进一步改善控制速度和提高被试舒适度.本文可望为基于SSVEP或与运动想象混合的脑机交互应用于分级或精细控制机器人方向和速度提供思路,并为直接脑控机器人技术推向实际应用打下一定的基础.     

异步BCI系统中运动想象诱发脑电特征识别

针对异步运动想象脑机交互（Brain Computer Interface,BCI）系统中空闲状态检测和不同想象任务分类的问题,在小波变换提取脑电信号特征基础上,设计了阈值判别结合支持向量机的二级分类器。由于大脑想象单侧肢体运动时,会导致同侧和对侧运动皮层脑区EEG信号在μ节律上分别出现事件相关同步和去同步,而大脑处于空闲状态时则无此现象。基于大脑活动的这一特性,提出了小波能量阈值判别法,进行空闲状态检测,径向基核函数和交叉检验的支持向量机方法,进行左、右手运动想象任务分类。结果表明该分类器最佳分类正确率达到了80.7%,且整个时间消耗仅为3.0 s,可以较好地满足异步在线运动想象BCI系统的应用。     

一种基于时–空–频联合选择与相关向量机的 运动想象脑电信号分析算法

研究表明:不同受试者由于个体差异,会引起在执行相同运动想象任务时,产生与受试者关联的特定脑电信号特征,这是设计脑机接口系统面临的一个实际问题.为解决这个问题,本文提出了一种基于时–空–频联合特征的提取方法.首先,对原始118导联的EEG进行空间特征分析,从中提取出与运动想象相关脑区对应的55导联EEG信号.进一步,在训练集上,通过7–折交叉验证,训练出与受试者匹配的时间窗和频带.其次,利用8个共空域滤波器进行特征提取.最后,将获得基于样本的运动想象特征,采用相关向量机进行分类.仿真结果表明:该算法在第3届脑机接口竞赛数据集Data IVa分类上获得5位受试者平均分类精度为94.49%,结果优于当年第1名94.17%.此外,与其他3种常用的方法比较亦具有明显优势.本文提出的基于样本的时–空–频特征提取方法和相关向量机的结合,该算法整体性能优越,为基于运动想象的脑机接口在线系统设计提供了一种新方法.     

Optimal work–rest cycles for an isometric intermittent gripping task as a function of force,posture and grip span

The aim of this study was to investigate the maximum acceptable contraction frequencies (i.e. work–rest cycles) for an isometric-intermittent handgrip task as a function of grip span, applied force and shoulder posture using psychophysical and physiological approaches. Twelve healthy males served as subjects. The three grip spans investigated were the optimal, 2 cm narrower than the optimal, and 2 cm wider than the optimal. The grip force levels studied were 15% and 30% of maximum voluntary grip force and the two shoulder postures were 25° flexion and 30° abduction. The psychophysical results indicate that subjects work faster with the narrower grip span at 15% of maximum voluntary grip force level in comparison to working with the optimal and the wider spans. However, when the task required 30% of maximum grip force level, the subjects worked faster with the optimal grip span. These findings were supported by the results of electromyography, heart rate, blood pressure and perceived discomfort. The study suggests that grip span of a tool is an important factor to be considered in predicting optimal work–rest cycles for hand grip tasks, and the optimum setting of grip span of the hand-tool depends on the required task force level. That is, the optimality is relative rather than absolute. In addition, it appears that weaker subjects can work at a higher rate than stronger ones at the same relative force level.     

Grip force and force sharing in two different manipulation tasks with bottles

Abstract

Grip force and force sharing during two activities of daily living were analysed experimentally in 10 right-handed subjects. Four different bottles, filled to two different levels, were manipulated for two tasks: transporting and pouring. Each test subject’s hand was instrumented with eight thin wearable force sensors. The grip force and force sharing were significantly different for each bottle model. Increasing the filling level resulted in an increase in grip force, but the ratio of grip force to load force was higher for lighter loads. The task influenced the force sharing but not the mean grip force. The contributions of the thumb and ring finger were higher in the pouring task, whereas the contributions of the palm and the index finger were higher in the transport task. Mean force sharing among fingers was 30% for index, 29% for middle, 22% for ring and 19% for little finger.

Practitioner Summary: We analysed grip force and force sharing in two manipulation tasks with bottles: transporting and pouring. The objective was to understand the effects of the bottle features, filling level and task on the contribution of different areas of the hand to the grip force. Force sharing was different for each task and the bottles features affected to both grip force and force sharing.     

Optimal work-rest cycles for an isometric intermittent gripping task as a function of force, posture and grip span

The aim of this study was to investigate the maximum acceptable contraction frequencies (i.e. work-rest cycles) for an isometric-intermittent handgrip task as a function of grip span, applied force and shoulder posture using psychophysical and physiological approaches. Twelve healthy males served as subjects. The three grip spans investigated were the optimal, 2 cm narrower than the optimal, and 2 cm wider than the optimal. The grip force levels studied were 15% and 30% of maximum voluntary grip force and the two shoulder postures were 25 degrees flexion and 30 degrees abduction. The psychophysical results indicate that subjects work faster with the narrower grip span at 15% of maximum voluntary grip force level in comparison to working with the optimal and the wider spans. However, when the task required 30% of maximum grip force level, the subjects worked faster with the optimal grip span. These findings were supported by the results of electromyography, heart rate, blood pressure and perceived discomfort. The study suggests that grip span of a tool is an important factor to be considered in predicting optimal work-rest cycles for hand grip tasks, and the optimum setting of grip span of the hand-tool depends on the required task force level. That is, the optimality is relative rather than absolute. In addition, it appears that weaker subjects can work at a higher rate than stronger ones at the same relative force level.     

Dependence of safety margins in grip force on isometric push force levels in lateral pinch

This study examined the relationship between safety margin and force level during an isometric push task in a lateral pinch posture. Ten participants grasped an object with an aluminium- or rubber-finished grip surface using a lateral pinch posture and exerted 20%, 40%, 60%, 80% and 100% of maximum push force while voluntary grip force was recorded. Then minimum required grip force was measured for each push force level. Mean safety margin, the difference between voluntary and minimum required grip forces, was 25% maximum voluntary contraction (MVC) when averaged for all push levels. Safety margin significantly increased with increasing push force for both grip surfaces. Grip force used during maximum push exertion was only 74% lateral pinch grip MVC. Possible underlying mechanisms for increasing safety margin with increasing push force are discussed as well as the implication of this finding for ergonomic analysis. This study demonstrates that ergonomic analyses of push tasks that involve friction force should account for safety margin and reduced grip strength during the push. Failure to consider these can result in overestimation of people's push capability.     

An EEG Brain–Computer Interface Approach for Classifying Vigilance States in Humans: A Gamma Band Focus Supports Low Misclassification Rates

This research investigated a brain–computer interface (BCI) for classifying vigilance states. A BCI approach based on event-related potential (ERP) and spectral features of electroencephalographic (EEG) data derived from an auditory oddball paradigm was employed. A 128-channel EEG system recorded potentials while participants simultaneously completed a prolonged visual match-to-sample task. It was hypothesized that better classification rates would be found for a group of “bored” participants, as compared to the not-bored participants. The best classification rates were found by extracting EEG features from gamma frequencies at middle and late latencies (for the standard tones). Critically, the BCI paradigm was most effective for the bored participants (mean misclassification rate = 13%), as compared to the not-bored participants (mean misclassification rate = 21%). These findings extend our knowledge of reliable latencies and spectral features of ERP data most salient to the classification of vigilance states.     

Contributions and co-ordination of individual fingers in multiple finger prehension

The contributions and co-ordination of external ringer grip forces were examined during a lifting task with a precision grip using multiple fingers. The subjects ( n = 10) lifted a force transducer-equipped grip apparatus. Grip force from each of the five fingers was continuously measured under different object weight (200 g, 400 g, and 800 g) and surface structure (plastic and sandpaper) conditions. The effect of five-, four-, and three-finger grip modes was also examined. It was found that variation of object weight or surface friction resulted in change of the total grip force magnitude; the largest change in finger force, was that for the index finger, followed by the middle, ring, and little fingers. Percentage contribution of static grip force to the total grip force for the index, middle, ring, and little fingers was 420%, 27·4%, 17·6% and 12·9%, respectively. These values were fairly constant for all object weight conditions, as well as for all surface friction conditions, suggesting that all individual finger force adjustments for light loads less than 800 g are controlled comprehensively simply by using a single common scaling value. A higher surface friction provided faster lifting initiation and required lesser grip force exertion, indicating advantageous effect of a non-slippery surface over a slippery surface. The results indicate that nearly 40% force reduction can be obtained when a non-slippery surface is used. Variation in grip mode changed the total grip force, i.e., the fewer the number of fingers, the greater the total grip force. The percent value of static grip force for the index, middle, and ring fingers in the four-finger grip mode was 42·7%, 32·5%, and 24·7%, respectively, and that for the index and middle fingers in the three-finger grip mode was 43·0% and 56·9%, respectively. Therefore, the grip mode was found to influence the force contributions of the middle and ring fingers, but not of the index finger.     

Interfering effects of the task demands of grip force and mental processing on isometric shoulder strength and muscle activity

The purpose of this study was to examine the interfering effects of physical and mental tasks on shoulder isometric strength in different postures. Fifteen volunteers (seven women, eight men) performed a series of isometric shoulder exertions at 30°, 60° and 90° of both shoulder flexion and abduction alone and with the addition of a 30% grip force, a mental task (Stroop test) and both additional tasks simultaneously. The shoulder tasks were completed either at maximal intensity, or while maintaining a shoulder posture without any additional effort. Surface electromyography (EMG) from seven muscles of the shoulder girdle and shoulder moment were collected for each 6 s shoulder exertion. When normalized to maximum exertion, no differences were found between genders and no differences existed between conditions when subjects maintained each posture without exerted force. In the maximal shoulder exertion trials, an increase in shoulder angle (in either plane) resulted in an increase in EMG in most muscles, while shoulder moment decreased in flexion and remained constant in abduction. Shoulder moments and muscle activation were greatest in the shoulder exertion alone condition followed by adding a 30% grip and the Stroop test, with the addition of both tasks further reducing the exerted shoulder moment and EMG. However, muscle activity did not always decrease with shoulder strength and remained elevated, indicating a complex coactivation pattern produced by an interfering role of the tasks. Overall, it was found that a mental task can have the same or greater effect as a concurrent grip and should be considered when assessing muscular loading in the workplace, as typical biomechanical modelling may underestimate internal loads. The results not only provide valuable shoulder strength data but also practical strength values, depending on additional tasks.     

运动相关电位分类算法比较和语义范式分析

运动相关电位（MRPs）机理复杂、形式多变,使得对基于MRPs的脑电信号的特征提取和数据挖掘工作很具有挑战性。本文目的是要将多种机器学习和语义范式模型应用于对脑电信号的数据挖掘,以应对上述挑战。本文采用多种机器学习算法和信号处理方法进行分析和实验对比,并给出对应不同场景、目标的最佳模型。为了将跨度较大的模糊性的电生理信号、兼容多种信号的深度学习和明确的语义模型各领域无缝地衔接,实现了一个以脑电信号数据为研究对象的语义范式框架,赋予复杂信号以文法、语法和语义内涵,为深度神经网络构筑了语义解释。通过该范式框架能够找出脑电信号中特定语义的信息块以及这些信息块之间的语义组合,自动学习出高效的滤波器,达到准确率高、传输通量大、普适性强的效果。     

Finger gesture input utilizing the rear camera of a mobile phone: A perspective of mobile CAD

Various interfaces have been suggested for mobile devices, including touch gesture and embedded sensor‐based interfaces. However, if a user's task requires a thorough look, these interfaces hinder sight of view in display and thus can be inappropriate for the tasks. This problem is more important especially in the mobile computer aided design (CAD) context, which performs visually demanding tasks on the limited screen. To address this point, this study suggests a mobile interface utilizing the rear camera of the device and describes the function mapping for CAD. The suggested interface detects finger attachment and movement direction. The number of colors in camera vision and local binary pattern are used as features, and a support vector machine (SVM) is used for feature classification. A prototype application is designed for validation, and appropriate SVM models are selected through benchmarking tests. The validation results show that the suggested interface can perform with high accuracy and low computational resources.