毕业生知识库

       脑机接口是研究大脑结构和功能的重要手段之一，可以通过脑电信号将人 类大脑与外部计算机设备相连接，从而解析大脑活动、推测大脑意图。基于运动 想象的脑机接口是一种十分重要的设备，它可以帮助使用者通过想象身体某一 部位的运动来实现对外部设备的控制，在医疗、娱乐领域都有着巨大的应用价值 和发展潜力，正逐渐吸引大家的关注。

      运动想象脑机接口是一系列软件、硬件的结合，完整的工作流程涉及多个环 节，其中最为重要的部分是脑电信号的特征提取与分类。脑电信号分类准确与 否直接影响着脑机接口的性能，故本文针对这一问题，在受试者相关模式脑电 信号分类算法研究的基础上，展开对受试者无关模式脑电信号分类算法的研究， 并实际采集运动想象脑电数据对所提出的算法进行验证，具体研究内容如下： 

     1. 由于脑电信号具有空间域及时间域特征，本文提出基于 Transformer 的模 型 CCTNet 来对受试者相关下的脑电信号进行分类。CCTNet 使用共空间模式、 多层卷积神经网络和 Transformer 逐步提取脑电信号的空间域特征和时间域特 征。在四分类运动想象脑电数据集 BCIC IV 2a 上，CCTNet 对 9 位受试者的分类 准确率最高达到 93.1%，与其他两种常用的网络结构相比，平均分类准确率高出 3.81% 和 6.13%。 

      2. 根据不同受试者在采集脑电信号时，电极在大脑上的位置保持一致这一 特点，本文提出基于图卷积神经网络的模型 GCCTN 来对受试者无关下的脑电信 号进行分类。GCCTN 将脑电数据按照电极的位置抽象成为一幅图后，首先通过图 卷积神经网络提取不同电极之间的关系，然后利用卷积神经网络和 Transformer 提取空间域和时间域特征。在四分类运动想象脑电数据集 BCIC IV 2a 上，GCCTN 对 9 位受试者的分类准确率最高达到 67.53%，平均准确率与其他方法相比提高 了 3.14%-33.46%，准确率标准差只有 1.01%。 

     3. 本文使用 8 通道脑电信号采集设备 OpenBCI，采集真实环境下 3 位受试 者的运动想象脑电数据，通过滤波去除脑电信号中的伪影，并在此数据集上验证 了所提模型 CCTNet 和 GCCTN 的有效性。此外本文利用离线训练好的 CCTNet 模型，基于多线程和环形缓冲区开发了实时脑机接口系统，并将其用于游戏的控 制。

请输入中文摘要

      Brain computer interface is one of the important means to study the structure and function of the brain. It can connect the human brain with external computer devices through EEG signals, so as to analyze the brain activity and predict the brain intention. Brain computer interface based on motor imagery is a very important device, which can help users to realize the control of external devices by imagining the movement of a certain part of the body. It has great application value and development potential in medical and entertainment fields, and is gradually attracting people’s attention. 

      The motor imagery brain computer interface is the combination of a series of software and hardware. The complete workflow involves many links, among which the most important part is the feature extraction and classification of EEG signals. The accuracy of EEG classification directly affects the performance of brain computer interface. Therefore, in view of this problem, based on the study of EEG classification algorithm of subjects-dependent mode, this thesis carries out research on EEG classification algorithm of subjects-independent mode, and verifies the proposed algorithm by actually collecting motor imagery EEG data. Specific research contents are as follows: 

      1. Due to the spatial and temporal features of EEG, this thesis proposes a model CCTNet based on Transformer to classify EEG signals under subject-dependent mode. CCTNet uses common spatial patterns, multi-layer convolutional neural networks and Transformer to extract spatial and temporal features of EEG signals step by step. CCTNet achieves a maximum classification accuracy of 93.1% for 9 subjects on the fourclassification motor imagery EEG dataset BCIC IV 2a, which is 3.81% and 6.13% higher than the other two commonly used network structures. 

      2. According to the characteristics that electrodes in the brain remain the same when collecting EEG signals from different subjects, this thesis proposes a model GCCTN based on the graph convolutional network to classify the EEG signals under subjectdependent mode. After abstracting EEG data into a graph according to electrode positions, GCCTN firstly extracts the relationship between different electrodes through the graph convolutional network, and then extracts spatial domain and time domain features using convolutional neural networks and Transformer. On the four-classification motor imagery EEG dataset BCIC IV 2a, GCCTN has the highest classification accuracy of 67.53% for 9 subjects, the average accuracy is 3.14%-33.46% higher than other methods, and the standard deviation of accuracy is only 1.01%. 

      3. In this thesis, 8-channel EEG acquisition equipment OpenBCI is used to collect motor imagery EEG data of 3 subjects in real environment, and the artifacts in EEG are removed by filtering, and the effectiveness of the proposed models CCTNet and GCCTN are verified on this dataset. In addition, a real-time brain computer interface system based on multithreading and ring buffer is developed by using the off-line trained CCTNet model, and it is used for game control.

输入英文摘要