基于卷积神经网络的运动想象脑电信号特征提取与分类

发布时间：2018-05-30 19:21

本文选题：脑机接口技术 + 公共空间模式　；参考：《郑州大学》2017年硕士论文

【摘要】：脑机接口技术(Brain Computer Interface,BCI)是指不依靠脑神经和肌肉组织,通过计算机在大脑和外部环境之间构建一种新的信息传输回路,可以直接实现大脑与外界环境的信息交流,是一种新型的人机交互方式。它能为思维正常但患有肌肉损伤、肌肉功能障碍的群体提供一种与外界交流的新途径,并且在军事应用、娱乐等领域也有着较大的应用前景。脑机接口系统的关键技术是通过对脑电信号进行处理,从中提取出能表征受试者的思维活动的脑电成分,作为分类器的输入,将分类结果转换成控制命令输出,实现对外部设备的控制。本文以提高四类运动想象脑电信号的识别率为目标,结合运动想象脑电信号的特性,对脑电信号的预处理、特征提取以及分类等过程进行了分析研究,主要工作如下:(1)首先介绍了脑电信号的分类及特点,对运动想象脑电信号的特性进行了分析。结合小波包变换(Wavelet Packet Transform,WPT)与快速独立分量分析(Fast Independent Component Analysis,FastICA)两种算法,对脑电信号进行预处理,滤除了脑电信号中的高频噪声与伪迹,便于后续的特征提取和分类。(2)当输入样本量较少,采用卷积神经网络(Convolutional Neural Network,CNN)对脑电信号进行分类时,网络权值得不到充分训练,导致分类效果差。针对以上问题,本文创新性地将公共空间模式(Common Spatial Pattern,CSP)与CNN相结合,对多类运动想象脑电信号进行两级特征提取并分类,与传统脑电分类算法相比,分类正确率有较大提高。(3)较小的脑电信号样本量导致无法充分训练CNN网络权值,而不同实验过程中网络权值的初始值对分类结果影响较大。针对这个问题,本文首先利用遗传(Genetic Algorithm,GA)算法对CNN权值进行预训练,在解空间内进行权值全局寻优;其次,利用误差反向传播(Back Propagation,BP)算法对最优网络权值进行局部修正。实验结果表明,本文所提出的方法能够在小样本量的情况下,达到较好的CNN网络训练效果,比传统利用BP算法对CNN进行权值训练得到的分类结果更加稳定。
[Abstract]:Brain-computer interface technology brain Computer Interface (brain-computer interface) refers to the construction of a new information transmission circuit between the brain and the external environment without relying on the brain nerve and muscle tissue, which can directly realize the information exchange between the brain and the external environment. It is a new type of human-computer interaction. It can provide a new way to communicate with the outside world for the group with normal thinking but suffering from muscle injury and muscle dysfunction, and also has a great application prospect in military application, entertainment and other fields. The key technology of the BCI system is to extract the EEG components which can represent the mental activity of the subjects by processing the EEG signals, which can be used as the input of the classifier to convert the classification results into the output of the control commands. To achieve the control of external equipment. In order to improve the recognition rate of four kinds of motion imaginary EEG signals, this paper analyzes and studies the processes of EEG preprocessing, feature extraction and classification combined with the characteristics of motion imaginary EEG signals. The main work is as follows: (1) first, the classification and characteristics of EEG are introduced, and the characteristics of motor imaginary EEG are analyzed. Combining wavelet Packet transform (WPT) with fast independent component analysis (Packet) and fast Independent Component analysis (Fast Independent Component Analysis) algorithm, the EEG signal is preprocessed to filter the high-frequency noise and artifact in the EEG signal. When the input sample size is small and the Convolutional Neural Network is used to classify EEG signals, the network weights are not sufficiently trained, which results in poor classification effect. Aiming at the above problems, this paper innovatively combines Common Spatial pattern with CNN to extract and classify the two-level feature of multi-class motion imaginary EEG signals, which is compared with the traditional EEG classification algorithm. The small sample size of EEG leads to the inability to fully train the weights of CNN networks, and the initial values of network weights in different experiments have a great influence on the classification results. To solve this problem, this paper first uses genetic algorithm GA (genetic algorithm) algorithm to pretrain the weight value of CNN, and makes a global optimization of the weight value in the solution space. Secondly, the weight value of the optimal network is locally modified by using the error back propagation (back propagation) algorithm. The experimental results show that the proposed method can achieve better training effect of CNN network under the condition of small sample size, and is more stable than the classification results obtained by traditional BP algorithm for CNN weight training.
【学位授予单位】：郑州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R318;TN911.7

【相似文献】