基于运动想象脑电信号非线性特性分析的脑—机接口研究

发布时间：2018-09-03 13:35

【摘要】：脑-机接口（Brain-Computer Interface，BCI）是一种帮助人们利用他们的大脑控制和使用外部设备的一种通信系统，在此过程中不需要外周神经和肌肉的参与。BCI是一门涉及神经科学、信号处理、计算机科学等多个领域的交叉学科。近20年来，已成为国际智能科学领域的一个研究热点。BCI研究的核心就是如何将用户的脑电信号转换成外部设备的控制信号。所以BCI研究最重要的工作就是要寻找合适的信号处理和转换方法，使得人脑的意识特征信号能够快速、准确地被计算机识别。一般来说，BCI系统可以看成是一个模式识别系统。一个BCI系统是否成功主要取决于两个方面的因素：①获取的特征能够区分不同的意识任务；②分类算法准确有效。所以如何建立准确可靠的特征提取模型和设计高效的分类算法是目前研究的主要难点。目前，在基于运动想象的脑-机接口研究中，对EEG进行特征提取和分类往往都建立在EEG信号是线性的这一假设的基础之上。然而，大量研究表明，EEG信号是非线性的，采用线性方法来对EEG信号进行处理，会导致其非线性特征丢失，从而减弱这些特征在区分不同意识任务时的性能。所以，本论文在针对EEG信号的非线性特性研究的基础上，根据目前特征提取和分类算法中存在的问题，提出了新的基于EEG非线性特性的特征提取算法，并通过仿真实验证明了其可行性。论文的主要研究内容包括以下几个方面： ①对EEG动力学模型的非线性特性进行分析。通过相空间重构技术，对求解得到的EEG信号进行了重构。得出了he的吸引子随着参数p ee和pe i变化的规律。从而证实了大脑中存在混沌这一观点。对实际测得的EEG信号进行了非线性特性的研究。计算了脑-机接口竞赛提供的两个基于运动想象的数据集中EEG样本的最大Lyapunov指数，计算结果表明，几乎所有的标准数据集当中的EEG样本的最大Lyapunov指数均大于零。进一步证实了大脑中存在混沌的论点，因而可以使用非线性分析方法来对EEG信号进行分析。 ②分别计算了两个标准数据集样本的几种常见的混沌特征量，即最大Lyapunov指数、关联维数和近似熵，并分别使用最大Lyapunov指数、关联维数和近似熵作为运动想象的特征进行分类。结果表明，直接使用最大Lyapunov指数和关联维数作为运动想象任务的特征，不能很好的区分各种运动想象任务。而近似熵是衡量时间序列中产生新模式概率大小的一种度量，它更适合表示不同的意识任务。在对近似熵特征进行分析的基础上，提出了一种基于时间窗的近似熵特征提取和分类算法。该算法模拟在线脑-机接口的情况，在每个时间窗内对意识任务进行分类，，实验结果表明，分类器能较好的区分左右手运动想象任务。 ③提出了一套基于相空间重构的特征提取方法。从理论上证明了相空间重构函数具有滤波功能，并能够对EEG信号进行相位和幅度调节，从而使相空间的特征更能区别不同的脑电任务。基于相空间的特征提取方法保留了传统的线性特征提取方法的优势，又使获取的特征具有相空间的信息，因而提高了分类器的分类性能。本文使用了2003和2005两届脑机接口竞赛提供的数据进行了仿真，并采用了和竞赛相同的评价标准：互信息和最大互信息峭度。实验结果表明，该方法是一种极具竞争力的特征提取方法。采用相空间特征的Fisher分类器在Graz2003数据集取得了最大互信息值0.67，这是目前报道的最好结果。在对Graz2005数据集进行仿真的结果表明，相空间特征同样具有很好的效能，在平均最大互信息峭度和分类正确率的评价标准下均取得了很好的成绩。 ④针对共空间模式（Common spatial pattern，CSP）在解决多分类问题中的组合方式问题，提出了一种基于CSP和Fisher线性分类器的二叉树组合方式（BCSP）。在该方式下，Fisher线性分类器和CSP以二叉树的方式进行排列。任务的分类采用二叉查找的方式进行。在BCSP中，使用的CSP滤波器和Fisher分类器的数目比传统的“一对它”方式更少。而且在N分类过程中，对CSP投影矩阵和分类器的计算也能保持在最多log2N级别，大大提高了分类的效率，提高了分类的准确率。 ⑤在线BCI游戏平台的开发与实现。在对研究结果进行总结的基础上，设计了一套基于Neuroscan的在线BCI游戏系统。该系统可以通过脑电来进行Hangman游戏操作。该系统集成了训练模块，测试模块和游戏模块。能够完成从训练到实际操作的一整套功能。系统使用了C3、C4和O1通道来记录EEG信号，其中C3和C4通道的EEG信号用来提取左右手运动想象任务的脑电特征，O1通道的波作为确认信号。该系统采用了基于相空间重构的特征提取算法和Fisher线性分类器。对6个用户进行实验的结果表明，相空间特征均提高了运动想象任务的识别率，从而证明了相空间特征的有效性。文章的最后对所有的研究工作进行了总结，指出了论文主要研究工作的内容和取得的成果，并对下一步的工作进行了展望。
[Abstract]:Brain-Computer Interface (BCI) is a communication system that helps people use their brains to control and use external devices without the involvement of peripheral nerves and muscles. BCI is an interdisciplinary subject involving neuroscience, signal processing, computer science and many other fields. In the past 20 years, it has become a success. The core of BCI research is how to convert the EEG signals of users into the control signals of external devices. So the most important task of BCI research is to find suitable signal processing and conversion methods, so that the consciousness characteristic signals of human brain can be recognized quickly and accurately by computer. Generally speaking, a BCI system can be regarded as a pattern recognition system. The success of a BCI system depends on two factors: 1) the acquired features can distinguish different conscious tasks; 2) the classification algorithm is accurate and effective. The main difficulties of previous research.
At present, feature extraction and classification of EEG are often based on the assumption that EEG signals are linear in the study of brain-computer interface based on motor imagery. In this paper, a new feature extraction algorithm based on EEG nonlinear characteristics is proposed on the basis of the nonlinear characteristics of EEG signals and the problems existing in the current feature extraction and classification algorithms. The following aspects should be studied:
(1) The nonlinear characteristics of the EEG dynamic model are analyzed. The EEG signals are reconstructed by the phase space reconstruction technique. The law of the attractor of he changing with the parameters p EE and PE I is obtained. The chaos in the brain is confirmed. The nonlinear characteristics of the EEG signals are studied. The maximum Lyapunov exponents of two EEG samples from the brain-computer interface contest based on motion imagery are calculated. The results show that the maximum Lyapunov exponents of EEG samples from almost all the standard data sets are greater than zero. This further confirms the argument that chaos exists in the brain, so the nonlinear analysis method can be used. The EEG signal is analyzed by the method.
(2) Several common chaotic features of two standard datasets, namely, the maximum Lyapunov exponent, correlation dimension and approximate entropy, are calculated, and the maximum Lyapunov exponent, correlation dimension and approximate entropy are used to classify the motion imagery. Approximate entropy is a measure of the probability of generating new patterns in time series, and it is more suitable for representing different conscious tasks. Based on the analysis of the characteristics of approximate entropy, an approximate entropy feature extraction and classification based on time window is proposed. The algorithm simulates the on-line brain-computer interface and classifies conscious tasks in each time window. The experimental results show that the classifier can distinguish left and right hand motion imagery tasks better.
(3) A method of feature extraction based on phase space reconstruction is proposed. It is proved theoretically that the phase space reconstruction function has the function of filtering and can adjust the phase and amplitude of EEG signals, so that the phase space features can distinguish different EEG tasks better. In this paper, we use the data provided by the 2003 and 2005 BCI contests to simulate and adopt the same evaluation criteria as the contest: mutual information and maximum mutual information kurtosis. Fisher classifier based on phase space features achieves a maximum mutual information value of 0.67 in Graz 2003 data set, which is the best result reported so far. Simulation results on Graz 2005 data set show that phase space features also have good performance in average maximum mutual information kurtosis and score. Good results have been achieved under the accuracy rate of class accuracy.
(4) To solve the problem of combination of common spatial pattern (CSP) in multi-class classification, a binary tree combination method (BCSP) based on CSP and Fisher linear classifier is proposed. In this way, Fisher linear classifier and CSP are arranged in binary tree. The classification of tasks adopts binary search method. In BCSP, the number of CSP filters and Fisher classifiers used is less than the traditional "one-to-one" method, and the calculation of CSP projection matrix and classifier can also be maintained at the maximum log2N level in the N-classification process, which greatly improves the classification efficiency and classification accuracy.
_The development and implementation of online BCI game platform.On the basis of summarizing the research results,an online BCI game system based on Neuroscan is designed.The system can perform Hangman game operation by EEG.The system integrates training module,testing module and game module.It can complete one from training to practical operation. The system uses C3, C4 and O1 channels to record EEG signals. The EEG signals of C3 and C4 channels are used to extract EEG features of left and right hand motion imagery tasks, and the waves of O1 channels are used as confirmation signals. The results show that both the phase space features improve the recognition rate of the motion imagery task, thus proving the validity of the phase space features.
At the end of the paper, all the research work is summarized, and the main research contents and achievements are pointed out, and the future work is prospected.
【学位授予单位】：重庆大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TN911.7

【参考文献】