多类运动想象脑电模式识别及其在电动轮椅控制上的应用

发布时间：2018-09-13 16:47

【摘要】：脑-机接口是一种不依赖外周神经和肌肉组织的参与，在大脑和计算机或其他电子设备之间直接建立交流和控制通路的技术，从而将大脑信号解读成相应的命令来实现与外部世界的交流与控制。该技术不仅具有重要的理论研究价值，还具有实际的应用前景，已成为生物医学工程领域的研究热点之一，而基于运动想象脑电信号的研究是脑-机接口的一个重要分支。本文的研究建立在国家自然科学基金资助项目(61201302)的要求上，从课题的研究背景及意义出发，介绍了脑电信号的特点，分析了运动想象脑电信号的预处理、特征提取、模式分类的方法。本文进一步介绍了本文脑电信号的采集装置及方案，并通过其对脑电信号进行处理，然后将特定的运动想象任务转化为与其相对应的控制命令，最后输入到电动轮椅上，控制其完成特定的运动。本文完成了以下研究工作，并取得了一些研究成果： (1)脑电信号预处理阶段：本文采用优化的广义权重估计算法对脑电信号进行预处理，它可以在一定程度上消除与运动无关的信号，同时也可以增强运动想象的信噪比，从而为脑电信号的特征提取和模式分类提供好的基础条件。 (2)脑电信号特征提取阶段：从常规运动想象激活的局部脑区分析的角度出发，考虑到运动想象脑电信号中存在很多与运动想象无关的频率信号，而共空间模式特征提取方法缺少对频率信息的处理，本文提出了一种双树复小波与共空间模式相结合的特征提取方法。该方法首先选取特定通道的脑电信号，然后利用双树复小波多尺度分解，获取适当的频段，接着将各频段的信号联合起来输入到空间滤波器中，从而得到所需的特征向量。此外，从复杂脑功能网络的角度出发，又提出了一种基于脑功能网络邻接矩阵分解的新方法。该方法首先采用多通道运动想象脑电信号构建脑功能网络，然后对相应的邻接矩阵进行奇异值分解，依据矩阵奇异值特征向量定义了脑电的特征参数，最后将其组合为特征向量。 (3)脑电信号模式分类阶段：为了提高BCI系统中分类精度和分类速度，提出了一种基于深度自编码降维的主轴动态核聚类分类方法。首先，为了降低特征向量之间的相关性和计算的复杂度，引入深度自编码方法将特征向量进行降维处理，然后利用主轴动态核聚类分类进行分类识别。另外，由于支持向量机能够解决小样本估计、非线性、非平稳信号的分类问题，所以本文设计了基于多核学习支持向量机的多类分类器，可使分布复杂的数据信息在高维的特征空间中得到更充分的体现，在减少支持向量数目的同时提高分类精度。 (4)电动轮椅控制实验：首先设计四类运动想象的实验范式，并采集相应的脑电信号，然后将利用优化的广义权重估计算法实现盲源分离，接着采用双树复小波-共空间模式相结合的方法提取出脑电特征向量，进而使用多核学习支持向量机多类分类器对所得特征向量进行分类识别，，最后将识别结果转化为控制命令控制电动轮椅运动，三名受试者的平均正确率分别为66.78%，76.58%和72.53%。
[Abstract]:Brain-computer interface (BCI) is a technology that directly establishes communication and control channels between brain and computer or other electronic devices without the involvement of peripheral nerves and muscle tissues, thus interpreting brain signals into corresponding commands to achieve communication and control with the outside world. This technology is not only of great theoretical value, but also of great theoretical value. It has practical application prospects and has become one of the hotspots in the field of biomedical engineering. The study of EEG based on motor imagery is an important branch of brain-computer interface.
This paper is based on the requirement of the project supported by the National Natural Science Foundation of China (61201302). Starting from the research background and significance of the subject, the characteristics of EEG signals are introduced, and the methods of pretreatment, feature extraction and pattern classification of motor imagery EEG signals are analyzed. This paper completes the following research work and achieves some research results:1.
(1) EEG signal preprocessing stage: This paper uses the optimized generalized weight estimation algorithm to preprocess the EEG signal, which can eliminate the motion-independent signal to a certain extent, but also can enhance the signal-to-noise ratio of motor imagination, thus providing a good basis for feature extraction and pattern classification of EEG signal.
(2) Feature extraction of EEG signals: From the point of view of the analysis of local brain regions activated by conventional motor imagery, considering that there are many frequencies unrelated to motor imagery in motor imagery EEG signals, and the common spatial pattern feature extraction method lacks the processing of frequency information, this paper proposes a dual-tree complex wavelet and common space. First, the EEG signals of a specific channel are selected, and then the appropriate frequency bands are obtained by the dual-tree complex wavelet multi-scale decomposition. Then, the signals of each frequency band are jointly input into the spatial filter to obtain the desired eigenvectors. A new method based on the adjacency matrix decomposition of the brain functional network is proposed. Firstly, the brain functional network is constructed by using multi-channel motor imagery EEG signals, and then the corresponding adjacency matrix is singular value decomposition (SVD). According to the singular value eigenvector of the matrix, the characteristic parameters of the EEG are defined and combined into the eigenvector.
(3) EEG pattern classification stage: In order to improve the classification accuracy and speed in BCI system, a dynamic clustering classification method based on deep self-coding dimensionality reduction is proposed. In addition, support vector machine can solve the problem of small sample estimation, non-linear, non-stationary signal classification, so this paper designs a multi-class classifier based on multi-kernel learning support vector machine, which can make the distributed complex data information get more in the high-dimensional feature space. It fully reflects that the classification accuracy can be improved while reducing the number of support vectors.
(4) Electric wheelchair control experiment: Firstly, four kinds of experimental paradigms of motion imagery are designed, and corresponding EEG signals are collected. Then the optimized generalized weight estimation algorithm is used to realize blind source separation. Then, the EEG feature vectors are extracted by the method of dual-tree complex wavelet-common-space pattern, and then the multi-kernel learning support vector is used. Finally, the recognition results are converted into control commands to control the motion of the electric wheelchair. The average accuracy of the three subjects is 66.78%, 76.58% and 72.53% respectively.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN911.7

【参考文献】