多变量脑电信号分析及其在BCI中的应用研究

发布时间：2018-06-02 06:47

本文选题：脑-机接口 + 运动想象　；参考：《杭州电子科技大学》2016年硕士论文

【摘要】：脑-机接口技术直接从人体的思维源头——大脑出发,建立起连接大脑和计算机或其他设备间的“交流桥梁”,颠覆了原有的依赖外周神经和肌肉组织的沟通方式。由于该技术在医学、娱乐、智能生活等多个领域具有广泛的应用前景,所以成为脑科学研究的热点之一。本文以多变量运动想象脑电信号为研究对象,对预处理、特征提取、模式分类等信号处理过程进行研究,并应用于电动假肢的脑电控制。本文主要完成的研究工作以及取得的成果,如下:(1)预处理:为了减少噪声信号的干扰,本文采用了两种预处理方法:巴特沃斯滤波器和自适应小波阈值消噪法。分别在2008年公开竞赛数据集上进行实验,结果说明前者可以获取与运动想象节律信号相关的频段信息,后者可以有效地降低噪声干扰以提高信噪比。(2)特征提取:针对选取IMFs分量依赖于经验的问题,本文提出一种基于噪声辅助多变量经验模式分解(NA-MEMD)和互信息的有用IMF识别方法,用于脑电特征提取。首先,使用NA-MEMD算法对多通道信号进行分解得到多尺度IMF分量。然后,采用互信息法分别计算各尺度上信号与其IMF分量、噪声与其IMF分量、信号IMF分量与噪声IMF分量之间的相关性,计算相应的敏感因子以筛选出包含有用信息的IMF分量,将它们叠加起来得到各通道重构信号,采用共同空间模式算法提取重构信号的特征。该算法自适应选取了与脑电信号相关的有用信息,提高了特征区分度。通过与其他选取方法对比,该算法有效性得到验证。(3)模式分类:针对传统的高斯过程采用共轭梯度法确定超参数时对初值有较强依赖性且易陷入局部最优的问题,本文提出了一种基于人工蜂群优化的高斯过程分类方法,用于脑电信号的模式识别。首先,构建高斯过程模型,选择合适的核函数且确定待优化的参数。然后,选取识别错误率的倒数为适应度函数,使用人工蜂群算法搜索寻找出限定范围内可以取得最优准确率的超参数。最后,采用参数优化后的高斯过程分类器对样本分类,并通过实验证明本文算法的有效性。(4)多变量运动想象脑电在电动假肢控制上的探究:首先介绍了实验背景,接着设计了总体控制方案,然后设计了实验范式并对脑电信号进行采集。随后采用自适应小波阈值消噪法对脑电信号进行预处理,使用NA-MEMD和互信息方法提取脑电信号的特征,运用基于人工蜂群优化的高斯过程分类器对脑电特征进行分类。最后将分类结果映射成对应的控制命令,驱动电动假肢完成握拳和展拳动作。
[Abstract]:The brain-computer interface technology starts directly from the brain, the source of human thinking, and establishes a "bridge of communication" between the brain and the computer or other devices, which subverts the way of communication that relies on peripheral nerve and muscle tissue. Because of its wide application prospect in medicine, entertainment, intelligent life and so on, this technology has become one of the hotspots in brain science research. In this paper, the signal processing processes such as preprocessing, feature extraction and pattern classification are studied, and applied to EEG control of electric prosthesis. In order to reduce the interference of noise signal, two preprocessing methods are adopted: Butterworth filter and adaptive wavelet threshold denoising method. Experiments were carried out on the open competition data set in 2008. The results show that the former can obtain the frequency band information related to the motion imagination rhythm signal. The latter can effectively reduce noise interference to improve signal-to-noise ratio (SNR) feature extraction. Aiming at the problem of selecting IMFs components depending on experience, this paper proposes a useful IMF recognition method based on noise-assisted multivariable empirical mode decomposition and mutual information. It is used for EEG feature extraction. Firstly, NA-MEMD algorithm is used to decompose multi-channel signals to obtain multi-scale IMF components. Then, the correlation between the signal and its IMF component, the noise and its IMF component, the signal IMF component and the noise IMF component are calculated by mutual information method, and the corresponding sensitivity factors are calculated to screen out the IMF components containing useful information. The reconstructed signals of each channel are obtained by stacking them together, and the features of the reconstructed signals are extracted by the common spatial pattern algorithm. The algorithm adaptively selects useful information related to EEG signals and improves the classification of features. By comparing with other selection methods, the validity of the algorithm is verified. The proposed method can be applied to the traditional Gao Si process where the conjugate gradient method is used to determine the superparameters, which is highly dependent on the initial value and easily falls into the local optimal condition. In this paper, a Gao Si process classification method based on artificial bee colony optimization is proposed for pattern recognition of EEG signals. Firstly, the Gao Si process model is constructed, the appropriate kernel function is selected and the parameters to be optimized are determined. Then, the inverse of the recognition error rate is selected as the fitness function, and the artificial bee colony algorithm is used to search for the super-parameters which can obtain the best accuracy in the limited range. Finally, the Gao Si process classifier with optimized parameters is used to classify the samples, and the experimental results show that the algorithm is effective and the multivariable motion imagination EEG is applied to the control of electric prosthesis. Firstly, the background of the experiment is introduced. Then the overall control scheme is designed, and then the experimental paradigm is designed and EEG signals are collected. Then the adaptive wavelet threshold de-noising method is used to pre-process EEG signals, NA-MEMD and mutual information methods are used to extract the features of EEG signals, and Gao Si process classifiers based on artificial bee colony optimization are used to classify EEG features. Finally, the classification results are mapped to the corresponding control commands, driving the electric prosthesis to complete the grip and swing.
【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】：2016
【分类号】：TN911.6;R318

【参考文献】