【作者】 郭晓静；

【导师】 吴小培；

【作者基本信息】 安徽大学 ， 计算机应用技术， 2010， 博士

【摘要】 近年来,脑-机接口(Brain-Computer Interface:BCI)作为一种新型的人机交互技术,已成为当今脑科学和生物医学工程领域的热点课题。基于EEG信号的BCI是指在人脑和计算机或者其它电子设备之间建立的一种直接信息交流和控制通道。随着认知神经科学、计算机科学与信号处理技术的飞速发展,脑-机接口技术正在逐步向实用化方法发展。独立分量分析(Independent Component Analysis:ICA)是20世纪90年代中期发展起来的一种多维统计分析方法。近年来ICA在脑-机接口中的应用一直倍受关注,但是从目前已报道的研究中可以看出：ICA方法主要还是被作为一种预处理手段,用于消除原始脑电信号中的干扰成分,进而改善脑-机通信的正确识别率。而且在很多情况下均采用的是批处理ICA进行的脑-机接口离线分析。考虑到脑电信号具有很强的非平稳性,本论文的研究工作主要是围绕适合非平稳信号的在线ICA算法及其在脑-机接口中的应用问题展开的。论文的主要创新点在于：提出了将基于滑动窗的在线ICA算法应用于信号包络检测的新思路,并成功地将这种思路应用于BCI系统的信号特征提取和分类识别。通过对大量实测数据的分类测试,结果验证了所提方法的有效性。围绕这一创新点,本文在以下几个方面做了一些有特色的工作：1.在对ICA基本理论进行分析研究的基础上,重点对Herault-Jutten算法、Cichocki-Unbehauen算法、扩展Infomax算法和峭度极大ICA算法展开了研究,并对这几种算法的批处理和基于单个样本更新的在线形式进行了探讨和性能对比分析,为算法的改进提供了指导方向。2.研究了改进的基于滑动窗的在线ICA算法,并分别在非时变混合模型和时变混合模型两种情况下,对滑动窗在线算法的盲分离性能进行了研究分析。仿真实验结果说明滑动窗在线ICA算法能够在一定程度上兼顾批处理算法的稳定性和基于单个样本更新在线算法的自适应跟踪性能,对时变混合情况也能够获得较理想的盲分离结果。并提出了将滑动窗在线ICA算法应用于信号包络检测的新思路。仿真实验结果说明：利用在线算法得到的动态混合矩阵系数对信号进行包络提取是完全可行的。3.由于EEG信号的强非平稳性,提出了将适合非平稳性信号分析的Hilbert-Huang变换(HHT)方法引入脑-机接口的思想。通过实测数据分析得到：利用HHT算法中的EMD分解方法可以成功地实现对SSVEP信号的预处理和特征提取。此外,在基于HHT的左右手运动想象分析过程中,提出了一种基于运动想象¨节律包络特性的分类识别思想。实验结果验证了该思路是可行的,并为在线ICA算法在脑-机接口中的应用提供了指导思想。4.提出了将在线ICA算法应用于脑-机接口的新思路。大量的实测数据分类结果证明：利用在线ICA算法的包络提取性能,可以成功地实现对左右手运动想象脑电的动态分析和识别。为了验证分类结果的有效性,论文同时采用基于AR+BP、AR+LDA、AR+SVM和二阶矩能量的四种分类方法进行对比分析,结果表明基于在线ICA的分类效果非常理想。而且需要说明的是：在基于在线ICA算法的分类过程中并没有训练的环节,因此这种不经过训练就能获得较好分类结果的在线方法将会对BCI系统的简化提供很大的帮助。5.对在线ICA算法在α波检测中的应用进行了研究,结果进一步验证了在线ICA算法对于脑电节律波进行实时包络提取的可行性。6.成功地开展了对基于左右手运动想象、SSVEP和α波检测三类BCI系统的自主实验设计和数据采集工作,并获得了大量有效的自主实验数据,为本实验室的后续研究工作提供了数据来源。更多还原

【Abstract】 In recent years, as a new human-computer interaction technology, Brain-Computer Interface (BCI) has been a hotspot in brain science and biomedical engineering. EEG-based BCI can provide a direct communication and control channel for sending messages and instructions from brain to external computers or other electronic devices. With the rapid development of cognitive brain science, neuroscience, computer science and signal processing, BCI is working up to a practical technology.Independent Component Analysis (ICA) is a multi-dimensional statistical analysis method developed in the mid 90’s of the 20th century. Recently, the ICA-based BCI research has been attracted more and more researchers’attention. However, from the reported researches, it can be seen that ICA method is mainly used as a preprocessing step for EEG artifact removal and pattern enhancement so as to improve the communication accuracy between brain and computer. And in many cases, only batch ICA as off-line methods are often used for the analysis of BCI system.Considering the strong nonstationarity of EEG, the mainly research work of this thesis is about online ICA algorithm which can be used to analyze the nonstationary signals and its application on BCI. The main innovations of thesis are as follows:put forward a new idea of applying online ICA algorithm based on sliding window method to the real-time envelope extraction of signals. And successfully utilize the thought for the feature extraction and classification of BCI system. By using the method to classify a large number of real-world data, the experiment results prove that the proposed approach is effective for BCI research. Based on the innovations, some distinctive research works which have been finished are as follows:1. Based on studying the basic theory of ICA, the thesis focuses on the four ICA algorithms, including Herault-Jutten algorithm, Cichocki-Unbehauen algorithm, extended Infomax algorithm and the kurtosis-based ICA algorithm. Then the four block algorithms and their corresponding online algorithms updated with a new input sample are explored, and also the performances of block and online algorithms are comparatively analyzed in order to provide us an idea that is how to improve the current algorithms.2. Research on an improved online algorithm based on sliding window ICA. And under the two cases of non-time-varying and time-varying mixing model, the blind separation peformances of the improved online algorithm are respectively explored. Simulation results show that the improved approach not only partly has the stability of the batch algorithm, but also to some extent, it owns the adaptive tracking performance as the online algorithm updated with a new input sample. And for time-varying mixing model, the separation results by the method are also satisfied. In addition, the thesis proposed a novel thought of applying the online silding window ICA algorithm to the signal envelope detection. And simulation results illustrate that using dynamic mixing matrix coefficients got from improved algorithm can do well on the envelope extraction of signal.3. As the strong non-stationarity of EEG, the thesis proposes that using Hilbert-Huang transform (HHT) on BCI research. By analyzing the real-life data, the results show that utilizing the EMD method within HHT theory can successfully perform the pre-processing and feature extraction of SSVEP signal. Furthermore, during the HHT-based analysis of left and right hand motor imagery, a new idea is proposed that of emploring theμrhythm envelope as features to identify and classify the motor imagery EEG. Experimental results show that the proposed idea is feasible and can provide us a direction of applying online ICA algorithm on BCI.4. Propose to apply online ICA algorithm on BCI. Based on the classification results from large quantities of measured data, it can be identified that using online ICA algorithm to extract theμrhythm envelope can successfully classify the left and right hand motor imagery. To further verify the validity of the classification results, the thesis also uses the other four methods to identify and classify motor imagery EEG, such as AR and BP, AR and LDA, AR and SVM, and energy method based on second moment. By comparing of the results from defferent approachs, it shows that the online ICA algorithm can do well on the classification of motor imagery EEG. Moreover, it must be mentioned here that there is no training process when using online ICA algorithm to classify. Therefore, the method which has good classification peformence without training process will very helpful to simplify the BCI system.5. Research the detection of a wave by online ICA algorithm. Experimental results further confirm that using online ICA algorithms for envelope detection of EEG rhythm wave is effective.6. Proform by ourselves the experiment design and data collection successfully for three BCI systems, including left and right hand motor imagery based BCI, SSVEP-based BCI and a wave based BCI. Then a great deal of effective data is recorded and it will help us to further study on BCI.更多还原