【作者】 邹清；

【导师】 汤井田；

【作者基本信息】 中南大学 ， 生物医学工程， 2008， 硕士

【摘要】 脑—机接口因其广阔的应用价值和前景成为近年来脑科学、康复工程、自动控制、军事领域和生物医学工程等领域的热门研究课题。脑电信号的处理过程是研究过程中的重点和难点。本文将脑电信号中事件去同步化/相同步化现象作为特征信息,深入讨论了基于AR模型的自适应算法(AAR)和多变量参数AAR模型算法(MVAAR)在脑电信号特征提取中的应用。介绍多种对模型系数进行估计的方法,采用卡尔曼滤波方法和快速QR分解分别对AAR、MVAAR模型进行系数估计,以最大化显现脑电信号中的特征信息。采用线性分析、基于马氏距离分类和留一法三种分类器分别进行任务识别。引入了互信息,kappa值,ROC曲线下面积值的概念对分类效果进行性能评价。从实验结果上看,MVAAR算法比AAR算法达到了更高的分类正确率。AAR模型很好地描述了EEG信号的非平稳随机特征,MVAAR算法识别法主观性较小,阶次一般选取也比较低,数据仿真吻合度高,实现多导联数据的输入,具有更强的通用性。传统的线性分类、基于马氏距离的二次分类,留一法分类都达到了很好的效果,但也各有优缺点。LDA和MDA算法都是只由数据的均值和协方差决定的,当两类的协方差矩阵差别较大时,LDA方法则会表现出较大的偏差,而MDA方法则会表现出较好的结果。留一法的原理简单,容易实现,但如果当实验数据庞大时,计算量和计算时间将会是我们必须考虑的问题。不同对象因为个体的区别和测试反馈时间段的不同,对其使用同一组算法分类得到的效果也有差异。更多还原

【Abstract】 Brain-computer interface (BCI) has wide prospects for brain science, rehabilitation engineering, automatic control, military and biomedical engineering in recent years. It becomes the hot research topic in many areas. Signal processing of EEG is important and difficult.The paper discuss the methods based on the AR model of adaptive algorithm and multi-variable AAR model algorithm to extract feature information which is related with ERD/ERS in EEG Methods to estimate coefficients of the model are introduced. The paper adopts kalman filtering and QR decomposition to estimate the coefficients of AAR and MVAAR models respectively to maximize the real information of EEG Three ways as linear analysis (LDA), classification based on the Mahalanobis distance (MDA) and the method named "leave-one-trial out" are used to classify different tasks. The concepts as mutual information, kappa coefficient, and values of area under the ROC curve (AUC) are introduced to estimate the performance of classification.From the results, we can see MVAAR algorithm made higher accuracy than AAR. MVAAR algorithm reached the higher correct rate than AAR algorithm. AAR model describe non-stationary features of EEG very well. MVAAR algorithm which orders are relatively low need less subjectivity but has high simulation. It realizes multi-channel data input and is more practical. The traditional LDA, MDA and "leave-one-trial out" also reached good results. Although LDA and MDA algorithms are depend on the mean and covariance of the data, when the covariance of two types is great, MDA will perform better than LDA. The principle of "leave-one-trial out" is simple and easy to achieve, but if the experimental data is huge, computation and calculation time will be the problem that we need to consider. For characteristic of person and the test, using one algorithm to do classification for different subject made different results.更多还原