【作者】 张莉；

【导师】 何为；

【作者基本信息】 重庆大学 ， 电气工程， 2008， 博士

【摘要】 基于脑电信号(EEG)的脑-机接口(BCI)是当今BCI研究的主流。其中利用自发脑电信号,通过识别特定意识任务实现控制的BCI系统不需要外部刺激装置和长时间的训练,具有很大的应用前景和价值。本论文主要从EEG信号的预处理,特征提取方法和分类器的设计等方面对此类BCI系统进行了较为深入的研究,主要工作和成果有:采用有效的盲源分离(BSS)方法分别对EEG信号中的工频、眼电(EOG)和肌电(EMG)伪迹进行了分离和去除。独立分量分析(ICA)能够取得较为理想的去除EOG伪迹的效果,得到了广泛的应用,因此在本文中也采用了这种方法去除EEG中的EOG伪迹。此外,工频噪声也利用了ICA方法来去除。在直接利用ICA无法分离出6导EEG信号中的工频噪声的情况下,人为构造了两路工频信号,引入ICA的输入作为参考信号,从而成功的分离和去除了工频噪声。在以往的基于意识任务BCI的研究中,均没有对EMG伪迹进行处理,本文提出了一种基于典型相关分析(CCA)和低通滤波的BSS方法来分离和去除EMG伪迹。采集到的EEG信号可以看作是不相关的EEG源和EMG伪迹源的瞬时混合,因此,CCA能够将真实的EEG信号和EMG伪迹分离开。相较于EEG而言,自相关性较弱的EMG伪迹出现在最小自相关的CCA分量中。但是这些分量也包含了EEG信息,因此在对这些分量进行低通滤波处理以后,再重构出EEG信号。提出的这种新方法在去除EMG伪迹的同时能够有效的保留EEG信息,是BCI系统中一种更为理想的去除EMG伪迹的EEG信号预处理方法。本文研究了时-频特征提取方法和时-频-空特征提取方法。频域分析方法提取EEG的节律特征及频谱特征是目前最主要的方法,然而这种方法是建立在被分析的EEG信号是平稳的假设上。时频分析方法是适合于处理EEG信号这种典型非平稳信号的有效工具。本文将线性时频变换和双线性时频变换应用于不同意识任务EEG信号的特征提取。短时傅里叶变换(STFT)虽然不存在交叉项干扰,但是其时频聚集性不好。作者将STFT与AR模型相结合,在STFT的求解过程中采用AR谱代替傅里叶谱,这种改进的STFT获得了比较好的时频聚集性,能够更有效的提取EEG特征。Wigner-Ville分布(WVD)虽然时频聚集性好,但又存在交叉项干扰。平滑伪Wigner-Ville分布(SPWVD)是WVD的一种改进,达到了抑制交叉项的目的,因此能够更准确的提取EEG特征。时-频-空特征提取方法考虑了从多通道EEG信号(多变量EEG信号)提取空域特征,是一种提取EEG特征的新思路。由于直接对多变量EEG信号处理时,计算量大,得到的特征数多,该方法首先对多变量信号进行空域解相关处理,从而降低了计算量和特征数,提高了这种方法的实用性。设计了两种分类器:基于Fisher辨别分析(FDA)和马氏距离(MD)的分类器和基于最小二乘支持向量机(LS_SVM)的分类器。FDA是一类较为简单的线性判别函数,该线性分类器容易实现,计算量小,能够满足BCI系统对实时性的要求。基于结构风险最小化原则的SVM是近些年来发展起来的一种新型模式识别方法,克服了传统方法的过学习、高维数、局部最小等问题,具有很强的推广能力,但是其计算速度较慢。因此,本文利用了SVM的改进方法LS_SVM来设计分类器,LS_SVM将SVM中的二次规划问题转化为线性方程求解,提高了计算速度,更适合于在BCI系统中充当分类器。此外,本文也对利用SVM进行多分类的方法进行了研究,设计了基于LS_SVM的多分类器。本文设计了三组实验,每组实验均对8个受试者的意识任务进行了二分类和多分类。第一组实验用于验证本文所给出的特征提取方法的有效性,并比较了它们的性能。采用基于AR模型的STFT和SPWVD的特征提取方法取得了最好的分类效果。不过基于AR模型的STFT更加简单,因此是一种更具优势的方法。第二组实验对设计的两种分类器进行了验证和比较。在二分类时,LS_SVM分类器取得了更好的分类效果,而在三分类,四分类和五分类时,FDA+MD分类器在正确率和计算速度上都更具优势。本论文提出了一个新的观点,认为头皮方式采集的EEG信号的高频成分也包含了与意识状态相关的信息,增加从高频范围提取的EEG特征有助于提高基于意识任务BCI系统的分类效果。本文在实验三中对这个观点进行了验证,将包含高频带(40-100Hz)特征的分类结果与没有包含高频带特征的分类结果进行了比较,实验结果显示包含高频带特征的分类正确率显著高于不包含高频带特征的分类正确率,证明了提出的这个新的观点是正确的。更多还原

【Abstract】 The majority of research on human brain-computer interface (BCI) has been performed using electroencephalographic (EEG) recordings. The BCI systems based on spontaneous EEG and distinguishable mental tasks would be a promising system for this kind of BCIs is no need for stimulation equipment and long-term training. The study of the mental task based BCI in this paper mainly included EEG signal preprocessing, EEG feature extraction and classification. The main work and contributions are as follows:Effective blind source separation (BSS) methods were used to remove power line noise, electrooculography (EOG) artifacts and electromyography (EMG) artifacts from EEG signals. EOG artifact removal was performed with independent component analysis (ICA) approach that is effective and widely used in removing EOG artifacts from EEG. Besides, Power line noise was also removed by ICA method. When the six-channel EEG signals were fed as the input of ICA, there were no independent components found to be only related to the power line noises. As we know the noise frequency, we created two-channel power line signals artificially and added them to the input of ICA. In this way, the power line noises were separated from the EEG signals successfully. No studies of the mental task based BCI removed EMG artifacts before. In this paper, a new method for EMG artifact removal in EEG was presented, based on canonical correlation analysis (CCA) and low-pass filtering as a BSS technique. The raw EEG can be seemed as the mixture of EEG and EMG sources that are uncorrelated, thus, CCA can be used to separate the EEG signal and EMG artifacts. EMG artifacts have a low autocorrelation, therefore they were present in the lowest autocorrelated CCA components. However, these components were also found to contain brain activity. Therefore, low-pass filters were used to remove the EMG artifacts in those corresponding CCA components. The EEG signals were reconstructed with the brain activity related CCA components along with the low-pass filter processed CCA components. This new method is able to remove the EMG artifacts as well as keeping the related EEG as intact as possible and is shown to be effective in eliminating EMG contamination.Time-frequency and time-frequency-space feature extraction methods were proposed in this paper. Rhythm and spectra features extracted with frequency domain methods are mainly used in the mental task based BCI. But these methods are based on the hypothesis of stationarity of analyzed EEG signals. For EEG are typical nonstationary, the time-frequency methods are more suitable to analyze EEG. Linear time-frequency analysis and bilinear time-frequency analysis were applied to extract features from the spontaneous EEG during different mental tasks. Short-time Fourier transform (STFT) has no cross-terms,however it’s time-frequency concentration is not satisfied. In the solving procecess of STFT, auto-regressive (AR) model was used to replace FFT for spectral estimation. In this way, the time-frequency concentration became better. The STFT based on AR model was more effective for feature extraction. Wigner-Ville distribution (WVD) has good time-frequency concentration, but the cross-term is disturbed. Smoothed pseudo WVD (SPWVD), the improved WVD method, can suppress the cross-terms, thus it can extract EEG feature more accurately. Time-frequency-space method adding space domain features from the multi-channel EEG signals is a valuable idea. For the high dimension features extracted from the multivariable EEG and heavy calculation using this method, spatial decorrelation was performed to reduce the dimension of feature vector and calculation, which improved practicality. Two kinds of classifiers, namely the classifier based on Fisher discriminant analysis (FDA) and Mahalanobis distance (MD) and the classifier based on least squares support vector machine (LS_SVM), were developed in this paper. FDA+MD classifier with low computational complexity is simple and fit for on line application. SVM based on structural risk minimization is a new method for pattern recognition. This algorithm solves practical problems such as over learning, high dimension and local minimum in traditional methods and has very well generalization ability, but the calculation time is long. For this reason, we used LS_SVM for classification because it is more efficient and fit for BCI after transforming the quadratic programming problem into linear equation. Besides, multiple classifiers based on LS_SVM were also proposed.Three experiments were conducted in this study. In each experiment, two-class and multi-class classifications were performed using the EEG data from eight subjects. The effectiveness of the proposed feature extraction methods was studied in the first experiment. The results indicated that the best classification performances were achieved by STFT based on AR model and SPWVD methods. But the method of STFT based on AR model is simpler. Comparison between the two kinds of classifiers was carried out in the second experiment. It is evident that LS_SVM classifiers gave better classification accuracies in two-class classifications and FDA+MD classifiers had advantages on classification accuracy and calculating speed in three-class, four-class and five-class classifications. We proposed a new point of view that the high frequency band of scalp EEG might contain valuable information that may contribute to more accurate mental task classification. This view was verified in the third experiment. We compared the classification results obtained using the features from the high frequency band (40-100 Hz) together with those from the lower frequency bands with the classification results obtained using the features only from the lower frequency bands. Significantly higher classification accuracies were obtained by adding the high frequency band features compared to using the low frequency bands alone, which demonstrated that the information in high frequency components from scalp-recorded EEG is valuable for the mental task based BCI.更多还原