基于运动想象的脑电特征提取及特征迁移方法研究

发布时间：2018-03-13 13:22

本文选题：脑-机接口　切入点：运动想象　出处：《哈尔滨工业大学》2017年硕士论文　论文类型：学位论文

【摘要】：运动想象脑-机接口技术不依赖人的外周神经和肌肉组织,直接实现人脑对外部设备的控制,它可以帮助有运动障碍的患者,更好地与外界进行信息交流,在军事、航天、医疗和虚拟现实等领域有巨大的应用价值。脑电信号具有非平稳性,而传统运动想象技术在应用前需要标注大量的训练样本,并采用多通道采集的方式,这大大限制了其应用范围。本文在传统脑电信号处理方法的基础上,将迁移学习的思想应用于运动想象的分类,减少训练样本和测试样本的分布差异,以提高分类准确率。此外,针对运动想象技术对运算实时性要求高的问题,研究通道选择优化方法,在保证分类正确率损失有限的条件下,减少分析脑电信号的通道数量,以提高运动想象脑-机接口技术的实时性。本文具体研究工作如下。基于运动想象生理基础,研究运动想象脑电信号预处理方法。利用AR模型对运动想象脑电信号频谱分析,得出信号有效的频带范围8-30Hz,为滤波器通带频率的选择提供分析依据;并分析公共平均参考法(CAR)空间滤波增加不同思维脑电信号空间分布差异的优势,为获得高信噪比的脑电信号奠定基础。研究基于小波包变换的特征提取方法,选择小波包分解后特定子节点的小波系数,并提取能量特征,利用支持向量机,识别两种类型的运动想象任务,得出平均分类正确率为79.4%。在此基础上,研究通道选择的优化方法,基于Relief-F算法计算通道权重,在对分类效果影响有限的条件下,减少分析脑电信号的通道数量,有助于减少计算量,提高运动想象脑-机接口实时性。研究基于最小化MMD的迁移学习算法,并将算法应用于运动想象的分类。结果表明,该方法有助于提高实验者一段时间内运动想象的分类正确率,且能够使一个实验者训练的分类模型更加适用于另一个实验者的测试。证明了迁移学习算法比传统的分类方法有更好的适应性。结合以上研究,设计基于运动想象迁移学习实验。针对真实的脑电信号含有的伪迹问题,研究小波分析眼电伪迹滤除的方法,并探讨迁移学习在线实现方案。结果表明迁移学习应用于不同实验者,相比传统的分类方法,分类正确率提高了约10%,且提升效果优于其对同一实验者的应用。本文设计不但可以进行传统运动想象离线采集,还可以为在线测试及迁移学习在运动想象领域的进一步研究提供借鉴。
[Abstract]:The brain-computer interface technology does not rely on human peripheral nerve and muscle tissue, and directly realizes the human brain's control of external equipment. It can help patients with motor disorders to exchange information with the outside world better, in military, aerospace, There is great application value in the fields of medical treatment and virtual reality. The EEG signal is non-stationary, but the traditional motion imagination technology needs to mark a large number of training samples before application, and adopt the method of multi-channel acquisition. On the basis of the traditional EEG processing method, this paper applies the idea of transfer learning to the classification of motion imagination to reduce the difference between the distribution of training samples and test samples. In order to improve the classification accuracy, in order to improve the accuracy of classification, aiming at the problem that the motion imagination technology requires high real-time operation, the optimization method of channel selection is studied to reduce the number of channels for analyzing EEG signals under the condition that the loss of classification accuracy is limited. In order to improve the real-time performance of motion imagination brain-computer interface technology, the research work in this paper is as follows. Based on the physiological basis of motion imagination, the preprocessing method of motion imagination EEG signal is studied. The spectrum analysis of motion imagination EEG signal is made by using AR model. The effective frequency band range of the signal is 8-30Hz, which provides an analytical basis for the selection of the passband frequency of the filter, and analyzes the advantages of the common average reference method (CAR) spatial filtering to increase the spatial distribution differences of EEG signals of different thinking. In order to obtain EEG signals with high SNR, this paper studies the feature extraction method based on wavelet packet transform, selects wavelet coefficients of specific sub-nodes after wavelet packet decomposition, extracts energy features, and uses support vector machine (SVM). Based on the recognition of two types of motion imagination tasks, the average classification accuracy is 79.4%. On this basis, the optimization method of channel selection is studied, and the channel weight is calculated based on Relief-F algorithm, under the condition that the influence on classification effect is limited. Reducing the number of channels for analyzing EEG signals is helpful to reduce the computational complexity and improve the real-time performance of the motion imagination brain-computer interface. A migration learning algorithm based on minimized MMD is studied and applied to the classification of motion imagination. This method is helpful to improve the classification accuracy of the motion imagination of the experimenters for a period of time. Moreover, the classification model which can be trained by one experimenter is more suitable for the test of another experimenter. It is proved that the transfer learning algorithm has better adaptability than the traditional classification method. To solve the problem of real EEG artifacts, a wavelet analysis method of Eye-electric artifact filtering is designed. The results show that the transfer learning is applied to different experimenters, compared with the traditional classification method. The classification accuracy rate is increased by about 10%, and the effect is better than its application to the same experimenter. The design of this paper can not only carry on the traditional motion imagination off-line collection, It can also provide reference for the further research of online testing and transfer learning in the field of sports imagination.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R318;TN911.7

【相似文献】