摘要
脑机接口（Brain-computer interface, BCI）是通过分析反映大脑功能的脑电（Electroencephalograph, EEG）信号实现人与计算机直接通信的一种新型人机交互系统，已广泛应用于军事、助残康复、灾害救援、娱乐体验等多个领域。
在脑电信号分析研究中，事件相关电位（Event-Related Potential, ERP）通过多试次叠加而具有稳定的潜伏期、恒定的波形等优点，是脑电信号分类中的主要研究对象。但是实时BCI系统需要对单试次脑电信号进行分析，其最大难点在于单试次信号信噪比低，且性能与实验设计、预处理、分类方法等都有很大关系。对于不同类型的脑电数据，目前还没有比较系统全面的基于单试次的预处理和特征提取方法的对比分析研究。
针对以上问题，本文在包括四个稳态视觉诱发、三个运动想象和一个帕金森疾病检测的三种类型共八个EEG数据集上，使用现有的比较成熟的预处理和特征提取方法对每个数据集分别进行处理，根据单试次信号分类准确率对预处理和特征提取方式的性能进行了对比分析研究，总结了这三类脑电信号比较适合的单试次分类方式。
首先，在预处理方面，脑电数据的预处理与研究目的和实验设计有关，没有标准的流程，但是不同预处理步骤之间存在相互影响关系，其顺序对于分类结果也有影响。在加州大学圣地亚哥分校SCCN实验室提出的EEG信号预处理流程的基础上，本文提出了一种改进的噪声移除方法，首先对原始信号进行分段线性回归分析去除基线漂移，然后使用多窗谱进行工频噪声抑制，接着进行截止频率为0.1-1Hz的高通滤波，最后使用盲源分离进行生理伪迹移除。对稳态视觉诱发数据集一的一个被试和运动想象数据集三的五个被试，在SCCN实验室提出的和本文改进的预处理流程处理后的数据上，分别采用共同空间模式（common spatial pattern, CSP）与频率带的组合特征进行分类，四个被试在改进后的预处理流程上的分类准确率比SCCN实验室的预处理方法提高0.5%左右，验证了该预处理流程的有效性。
在特征提取方面，本文分别从时域、频域、空域和多域进行分析。实验结果表明，对于稳态视觉诱发类脑电信号，仅使用时域最具判别性时间窗的特征进行分类的准确率高于频域使用多窗谱估计提取的特征进行分类的准确率，在本文研究的四个稳态视觉诱发数据集上平均高4%；采用时频域特征进行分类的准确率比只用时域或频域的平均高1.4%，表明多域信息互补有助于提升单试次EEG信号的分类准确性。对于运动想象类脑电信号，本文提出了一种改进的根据得分函数选择滤波器的CSP方法，使用该方法提取空域特征进行分类可以获得不错的分类结果，在运动想象的三个数据集的分类准确率比使用传统CSP方法平均高1.2%；改进的CSP方法在使用之前需要进行线性滤波并选择时间间隔、带通滤波器的截止频率、以及CSP滤波器的子集等参数，实验结果表明，根据被试选择合适的参数之后，在本文研究的三个运动想象数据集上，有两个数据集的分类准确率比使用固定参数提高了3%左右，另外一个没有明显的提升，原因可能是该数据的通道数只有三个。疾病检测脑电是一种特殊的脑电类型，不同疾病表现的特征不一样，对于本文的帕金森疾病检测脑电数据，本文通过交叉验证选择判别性时间窗并根据被试选择合适频段提取多域特征进行分类，实验结果表明药物治疗组的平均分类准确率为73.8%，非药物治疗组的平均分类准确率为74.2%，比目前最好的根据先验确定时间窗和频段提取特征进行分类的准确率平均高0.4%；使用改进的CSP方法，在该数据集上实现了与时频域特征一致的结果。此外，本文还使用了双线性判别分析（Bilinear Discriminant Component Analysis, BDCA）的时空域方法，在稳态视觉诱发数据集二和运动想象数据集三上，分类准确率比现有最好的方法低1%左右。BDCA方法不需要过多的预处理，但是耗时较长，不适合实时BCI系统。对于分类器方法，在没有特殊说明的情况下，本文均使用支持向量机（Support Vector Machine, SVM）进行分类。当分类效果不好的时候，改用随机森林分类器，对于使用SVM分类效果不好的稳态视觉诱发数据集四，使用相同的特征类型，随机森林分类器的准确率比SVM高2.8%。
最后，本文设计了基于脑电的人脸识别实验，采集了九个正常被试人脸识别过程中的脑电信号并进行分析。首先，采用本文改进的预处理流程进行脑电信号预处理，接着将预处理后的数据分别从时域、频域和空域进行分析。在时域，先使用独立成分分析（Independent Components Analysis, ICA）增强时域信号，然后确定时域分析窗，该窗的确定采用根据先验知识选择和根据被试选择两种，实验结果表明，后者的平均分类准确率比前者高1%；在频域，提取每个被试最具判别性的频段，使用多窗谱提取每个频段的特征进行分类的平均准确率为68.8%，比只用时域特征分类的平均高8.4%，比采用时频域联合特征进行分类的平均低3.3%，证明了多域互补信息有助于单试次脑电信号的分类；在空域，先使用ICA在空域增强信号来提高信噪比，然后通过PCA降维，接着用本文改进的CSP方法提取特征进行分类，在该数据集的九个被试上，平均分类准确率达到了84.7%，比使用传统CSP分类的平均高5.7%，验证了改进CSP算法对单试次脑电分类的有效性。在此基础上，采用一种训练集扩展的方法，将被试一的测试样本作为每个被试的扩展训练样本，在扩展后的训练集上提取多域特征进行分类，九个被试的平均分类准确率提升了3%左右。
关 键 词：脑电， 多类， 单试次分析， 脑机接口
ABSTRACT
Brain-computer interface (BCI) is a new type of human-computer interaction system that realizes direct communication between humans and computers by analyzing the signals of Electroencephalograph (EEG) that reflects brain function. It has been widely used in military, disability rehabilitation, Disaster relief, entertainment experience and many other fields.
In the analysis of EEG signals, Event-Related Potential (ERP) has the advantages of stable latency and constant waveform through multiple trials. It is the main research object in EEG signal classification. However, the real-time BCI system needs to analyze the single trial EEG signal. The biggest difficulty is that the signal-to-noise ratio of single trial is low, and the performance has a lot to do with experimental design, preprocessing, and classification methods. For different types of EEG data, there is currently no systematically comparative analysis of comprehensive pre-processing and feature extraction methods based on single trials.
In response to the above problems, this paper uses three mature types on a total of eight EEG datasets including four steady-state visual evoked, three motor imaging and one Parkinson's disease detection. Each data set is processed separately used existing mature preprocessing and feature extraction methods, and the performance of preprocessing and feature extraction methods are compared and analyzed based on the accuracy of single-signal classification, and the three types of EEG signals suitable for single trial classification are summarized. First of all, in terms of preprocessing, the preprocessing of EEG data is related to the research purpose and experimental design. There is no standard process, but there are interactions between different preprocessing steps, and their order also affects the classification results. Based on the EEG signal pre-processing process proposed by the SCCN laboratory at the University of California, San Diego, this paper proposes an improved noise removal method. First, piecewise linear regression analysis is performed on the original signal to remove baseline drift, and then multi-window spectrum is used to power frequency noise suppress, followed by high-pass filtering with a cut-off frequency of 0.1-1Hz, and physiological artifacts are removed using a blind source separation method. For one subject in the steady-state visual evoked dataset one and five subjects in the motor imaging dataset three, based on the data proposed by the SCCN laboratory and processed by the improved preprocessing process in this paper, the combined features of the common spatial pattern and the frequency band were used for classification. The classification accuracy ratio of the four subjects on the improved preprocessing process was compared. The classification accuracy of the four subjects in the improved preprocessing process was improved by about 0.5% compared with the preprocessing method of the SCCN laboratory, which verified the preprocessing process Effectiveness.
In terms of feature extraction, this paper analyzes from time domain, frequency domain, space domain and multi-domain. The experimental results show that for steady-state visual evoked EEG signals, the accuracy of classification using only the features with the most discriminative time window in the time domain is higher than the accuracy of classification using features extracted from multi-window spectrum estimation in the frequency domain. The four steady-state visual evoked dataset studied in this paper, the average accuracy is 4%; the accuracy of classification using time-frequency domain features is 1.4% higher than using only time-domain or frequency-domain, it is shown that the multi-domain information complementation is helpful for the classification accuracy of single trial EEG signals. For the EEG signals of motor imagination, this paper proposes an improved CSP method based on selecting a filter based on a scoring function, using this method to extract spatial features for classification can obtain good classification results. The classification accuracy of the three data sets of motor imagination is 1.2% higher on average than using the traditional CSP method; The improved CSP method requires linear filtering and selection of the cutoff frequency, time interval, and a subset of the CSP filter before use. The experimental results show that after selecting the appropriate parameters according to the participants, on the three motion imagination datasets studied in this paper, the classification accuracy of two datasets was improved by about 3% compared with the use of fixed parameters, and the other one did not improve significantly. The reason may be that the data has only three channels. Disease detection EEG is a special type of EEG. The characteristics of different disease manifestations are different. For the EEG data of Parkinson's disease detection in this paper, this paper selects a discriminative time window through cross-validation and select multi-domain features for classification according to the appropriate frequency band selected by the participants, the experimental results show that the average classification accuracy of the drug treatment group is 73.8%, and the average classification accuracy of the non-drug treatment group is 74.2%, which the accuracy rate is 0.4% higher than the best classification based on the priori time window and frequency band extraction features; Using the improved CSP method, results consistent with the time-frequency domain characteristics are achieved on this dataset. In addition, this paper also uses the time-space domain method of Bilinear Discriminant Component Analysis (BDCA). On the steady-state visual evoked dataset 2 and the motion imaging dataset 3, the classification accuracy rate is better than the best available methods about 1% lower. The BDCA method does not require excessive preprocessing, but it takes a long time and is not suitable for real-time BCI systems. Regarding the classifier method, in the absence of special instructions, this paper uses support vector machine (SVM) for classification. When the classification effect is not good, the random forest classifier is used instead. For the steady-state visual evoked dataset 4 that uses the SVM classification effect poorly, using the same feature type, the accuracy of the random forest classifier is 2.8% higher than the SVM.
Finally, this paper designs an EEG-based face recognition experiment, collecting and analyzing the EEG signals in the face recognition process of nine normal subjects. First, the EEG signal pre-processing is carried out by using the improved pre-processing procedure in this paper, and then the pre-processed data is analyzed from time domain, frequency domain and space domain, respectively. In the time domain, independent component analysis (ICA) is used to enhance the time domain signal, and then the time domain analysis window is determined, the window is determined by using prior knowledge and by the subject. The experimental results show that the average classification accuracy of the latter is 1% higher than the former; In the frequency domain, the average discriminant frequency of each subject is extracted and using the multi-window spectrum to classify the features of each frequency band for classification, the average accuracy is 68.8%. The classification using only the time-domain features is 8.4% higher on average and 3.3% lower than the classification using the time-frequency domain combined features, which proves that multi-domain complementary information is helpful for the classification of EEG signals in a single trial; in the airspace, first use ICA to enhance the signal in the airspace to improve the signal-to-noise ratio, and then use PCA to reduce the dimension, and then use the improved CSP method to extract features for classification. On the nine subjects in this data set, the average classification accuracy reached 81.9%, which is 5.7% higher than the average using traditional CSP classification, which validates the effectiveness of the improved CSP algorithm for single trial EEG classification. On this basis, a method of training set expansion is adopted, and one test sample of each subject is used as the extended training sample of each participant. Multi-domain features are extracted and classified on the expanded training set, and nine participants the average classification accuracy improved by about 3%.
Keywords: electroencephalogram (EEG), multiclass, single trial analysis, Brain-computer interface (BCI)
值图 
值 
