SIA OpenIR  > 机器人学研究室
基于运动想象的脑机接口及其应用研究
Alternative TitleResearch on Motor Imagery based Brain-Computer-Interface and the Application
邹宜君
Department机器人学研究室
Thesis Advisor徐卫良 ; 赵新刚
Keyword脑机接口 脑电信号 独立成分分析 共空间模式滤波
Pages89页
Degree Discipline机械电子工程
Degree Name博士
2019-11-25
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract本论文以运动想象脑电信号为研究对象,从脑电信号空间滤波和减少训练时间的角度开展了一系列研究,主要工作如下:首先,针对脑电信号中不同类型源信号彼此混合的问题,提出了一种有监督的独立成分分析算法,将运动想象信号的特性与独立性约束结合,提取出与运动意图相关的独立脑电成分。脑电信号具有空间分辨率低的特点,同时,由于头皮、头骨的传导效果,不同通道之间脑电具有高度相关性。而有限的通道信号是受大脑的所有意识行为影响。如何从这些信号中还原出与运动意图相关的成分便显得非常重要。因此,本文提出了一种提取运动意图相关独立脑电成分的空间滤波算法。该算法基于传统的独立成分分析算法(Independent Component Analysis, ICA),将ICA中的独立性约束与运动想象过程中的事件相关同步/去同步(Event related synchronization/desynchronization, ERD/ERS)现象相结合,建立了新的优化目标。在左右手运动想象问题中,该优化目标将EEG中的独立成分分为三类:与左手运动想象相关的独立成分、与右手运动想象相关的独立成分以及与运动无关的独立成分。然后经过迭代优化,求解该优化目标。得到一组空间滤波器,将原始信号映射为三类具有具体意义的独立成分。我们对与运动相关的独立成分进行特征提取后分类,得到了更好的分类精度。另一方面,我们从运动相关电位(Motion Related Potentials,MRP)的角度对得到的独立成分进行验证,发现提取到的与运动无关信号的MRP得到了极大的抑制。其次,针对采集训练数据所需时间过长的问题,提出了一种样本迁移算法,在目标用户小样本情况下利用其它用户数据建立有效的脑电模型,由于脑电信号具有高度的个体差异性,必须对每个用户单独建立识别模型。而运动想象实验中,单个样本采样耗时长,为了避免疲劳,样本采集间隔要留有足够的休息时间,因此采集训练数据就需要很长时间。这也大大降低了了脑机接口的实用性。针对这个问题,本文研究了如何利用数据库中已有的数据来实现小样本情况下的新用户建模,进而减少训练时间。从ERD/ERS的角度分析了不同用户运动想象时相关脑电信号的差异性,并重点关注大脑激活区域的差异性和ERD/ERS强度的差异性。针对不同用户,使用共空间模式滤波(Common Spatial Pattern, CSP)矩阵来表征运动想象时大脑激活的区域。并设计了一个正交矩阵来描述这种个体差异性,经过该正交矩阵的矫正,我们可以减少这种大脑激活区域带来的个体差异。然后,我们将不同用户进行相同动作的脑电信号变换到一个共空间之中,在该空间中对运动时信号的差异做抑制或增强。经过以上两步信号处理后,我们可以更好的使用其他用户的数据来建立目标用户的脑电模型。我们将我们的方法与减少校准时间的其他方法进行了比较。结果表明,我们的方法使用极少的训练试验(32个样本)达到了令人满意的识别效果。与使用少量训练试验的现有方法相比,我们的方法获得了更高的准确性。再次,针对卷积神经网络(Convolutional Neural Network, CNN)需要大量样本而脑电信号总体训练样本太少的问题,提出了对其他用户数据进行筛选与加权的算法,使其他用户数据满足训练目标用户模型的要求。本文针对运动想象动作识别问题建立了一个纯数据驱动的端到端的卷积神经网络模型,并在该模型的基础上建立了一套对已有数据库中数据进行筛选与加权的方法。通过分析其他人数据对目标个体模型的适应程度,我们可以清除那些对于目标模型贡献为负的样本。然后,在CNN网络的训练过程中,我们赋予每一个训练样本一个权值,并且在网络训练过程中同步更新该权值和CNN网络参数。实验结果表明,在仅使用单人数据时,我们的CNN网络能够达到目前主流特征提取方法FBCSP的精度。在使用多人数据并进行数据清洗和自适应样本加权技术时,我们的CNN网络能够得到比FBSCP更好的精度。最后,建立了基于左右手运动想象的机械臂在线控制系统。通过控制虚拟界面中小球的移动来评估脑电实时控制中存在的问题,在分析了小球控制实验结果的基础上,我们选择了最优的特征提取与分类算法,同时,比较了离线识别实验和在线控制实验在评估指标上的差异。最后我们使用选择的模型来控制机械臂完成移动到目标位置的任务。
Other AbstractThis essay takes the motion imagination EEG signal as the research object and carries out a series of studies centering on the spatial filtering and training time reduction needle of EEG signal. The main work is as follows: First of all, in order to solve the problem that different types of EEG signals mix with each other, a supervised independent component analysis algorithm was proposed, which combined the characteristics of motion imagination signals with independent constraints to extract independent EEG components related to motion intention. EEG signals have a low spatial resolution. At the same time, due to the conduction effect of scalp and skull, EEG of different channels has a high correlation. The limited channel signals are affected by all conscious actions of the brain. How to recover the components related to motion intention from these signals is very important. Therefore, we propose a spatial filtering algorithm to extract independent EEG components related to motion intention. The algorithm based on the traditional independent component points will be independent of the ICA constraints associated with sports events in the process of imagination/to sync (the Event related desynchronization/synchronization, the ERD/ERS) phenomenon, the combination of new optimization target is established. In the left and right hand motion imaging problem, the optimization objective divides the independent components in the EEG into three categories: independent components related to the left hand motion imaging, independent components related to the right hand motion imaging, and independent components independent of the right hand motion imaging. Then the iterative optimization is carried out to solve the optimization objective. A set of spatial filters is obtained to map the original signal into three kinds of independent components with specific meanings. The independent components related to motion are classified by feature extraction, and the classification accuracy is better. On the other hand, we verified the independent components obtained from the perspective of Motion Related Potentials (MRP), and found that the extracted MRP of non-motion Related signals was greatly inhibited. Secondly, aiming at the problem that collectiong traiing data required too long time, a sample migration algorithm is proposed to establish an effective EEG model with other user data in the case of small target user samples. due to the high individual differences of EEG signals, it is necessary to establish a separate recognition model for each user. In the exercise visualization experiment, it takes a long time to collect a single sample. In order to avoid fatigue, sufficient rest time should be left between sample collection, so it takes a long time to collect training data. This greatly reduces the usefulness of brain-computer interfaces. Aiming at this problem, this paper studies how to use the existing data in the database to realize the new user modeling in the case of small samples, so as to reduce the training time. From the perspective of ERD/ERS, the differences of relevant EEG signals in different users' exercise imagination were analyzed, and the differences in brain activation areas and ERD/ERS intensity were focused on. For different users, the Common Spatial Pattern (CSP) matrix is used to represent the areas activated by the brain during motor imagination. And an orthogonal matrix was designed to describe the individual differences. By correcting the orthogonal matrix, we could reduce the individual differences caused by the activated areas of the brain. Then, we transform the EEG signals of different users performing the same action into a common space in which the difference in signals during movement is suppressed or enhanced. After the above two-step signal processing, we can better use the data of other users to build the target user's EEG model. We compared our method with other methods to reduce calibration time. The results show that our method achieves satisfactory identification results with few training experiments (32 samples). Compared with the existing method which USES a small number of training experiments, our method achieves higher accuracy. Again, aiming at the problem that CNN needs a large number of samples while the total training samples of EEG signals are too small, an algorithm is proposed to filter and weight other user data, so that other user data can meet the requirements of the training target user model. the effective use of other users' data is another effective method of existing data in the database, filtering and the weighted against imagination in the process of EEG signals, this paper established a pure data driven end-to-end convolution Neural Network (Convolutional Neural Network, CNN) model, and on the basis of the model established a set of existing data in the database and weighted method of screening. By analyzing how well other people's data fit into the target individual model, we can remove samples that contribute negatively to the target model. Then, in the training process of CNN network, we give each training sample a weight, and update the weight and CNN network parameters synchronously in the network training process. The experimental results show that our CNN network can reach the accuracy of the current mainstream feature extraction method FBCSP when only single data is used. When using multi-person data and performing data cleaning and adaptive sample weighting, our CNN network can obtain better accuracy than FBSCP. Finally, an online control system of the manipulator based on the imagination of left and right hand motion is established. The problem of EEG real-time control was evaluated by controlling the movement of the ball in the virtual interface. On the basis of analyzing the results of the ball control experiment, we selected the optimal feature extraction and classification algorithm, and compared the difference in evaluation indexes between the offline recognition experiment and the online control experiment. Finally, we use the selected model to control the manipulator to complete the task of moving to the target position.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/25945
Collection机器人学研究室
Recommended Citation
GB/T 7714
邹宜君. 基于运动想象的脑机接口及其应用研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2019.
Files in This Item:
File Name/Size DocType Version Access License
基于运动想象的脑机接口及其应用研究.pd(3261KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[邹宜君]'s Articles
Baidu academic
Similar articles in Baidu academic
[邹宜君]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[邹宜君]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.