基于深度卷积神经网络的运动想象分类及其在脑控外骨骼中的应用

发布时间：2018-05-10 21:47

本文选题：深度学习 + 卷积神经网络　；参考：《计算机学报》2017年06期

【摘要】：基于运动想象的脑机接口技术已经广泛的应用于康复外骨骼领域.由于脑电信号的信噪比低,使得脑机接口分类率很难提高.因此,有效的脑电特征提取与分类方法成为现在的研究热点.该文创新地采用基于深度学习理论的卷积神经网络对单次运动想象脑电信号进行特征提取和分类.首先,根据脑电信号时间和空间特征相结合的特性,针对性地设计了一个5层的CNN结构来进行运动想象分类;其次,基于想象左手运动和脚运动设计了运动想象实验范式,获得运动想象实验数据;再次,将该方法应用于公共数据集和实验数据集并建立分类模型,同时与其它3种方法(功率值+SVM、CSP+SVM和MRA+LDA)相比较;最后,将从实验数据集中获得的分类模型(具有最好分类表现)应用于上肢康复外骨骼的实时控制中,验证该文提出方法的可行性.实验结果表明,卷积神经网络方法可以提高分类识别率:卷积神经网络方法应用在公共数据集(90.75%±2.47%)和实验数据集(89.51%±2.95%)中的平均识别率均高于其它3种方法;在上肢康复外骨骼的实时控制中,也验证了CNN方法的可行性:所有被试的平均识别率为88.75%±3.42%.该文提出的方法可实现运动想象的精确识别,为脑机接口技术在康复外骨骼领域的应用提供了理论基础与技术支持.
[Abstract]:The brain-computer interface technology based on motion imagination has been widely used in the field of rehabilitation exoskeleton. Because of the low signal-to-noise ratio (SNR) of EEG signals, it is difficult to improve the classification rate of BCI. Therefore, the effective method of EEG feature extraction and classification has become a hot topic. In this paper, a convolution neural network based on depth learning theory is used to extract and classify the feature of a single motion imaginary EEG signal. Firstly, according to the characteristics of EEG time and space, a five-layer CNN structure is designed to classify motion imagination. Secondly, based on the imagination of left hand motion and foot movement, a motion imagination experimental paradigm is designed. The experimental data of motion imagination are obtained. Thirdly, the method is applied to the common data set and experimental data set, and the classification model is established. At the same time, it is compared with the other three methods (the power value SVMN CSP SVM and the MRA LDAs). Finally, the proposed method is applied to the common data set and the experimental data set. The classification model (with the best classification performance) obtained from the experimental data set is applied to the real-time control of exoskeleton in upper limb rehabilitation, and the feasibility of the proposed method is verified. The experimental results show that convolution neural network method can improve the classification recognition rate: the average recognition rate of convolution neural network method is higher than that of the other three methods in common data set (90.75% 卤2.47%) and experimental data set (89.51% 卤2.95%). In the real-time control of exoskeleton for upper limb rehabilitation, the feasibility of CNN method was also verified: the average recognition rate of all subjects was 88.75% 卤3.42. The method proposed in this paper can realize the accurate recognition of motion imagination and provide the theoretical basis and technical support for the application of brain-computer interface technology in the field of rehabilitation exoskeleton.
【作者单位】：浙江大学计算机科学与技术学院;
【基金】：国家自然科学基金(61303137) 中国博士后科学基金(2015M581935) 浙江省博士后科学基金(BSH1502116) 浙江省科技计划项目(2015C31051,2016C33139)资助~~
【分类号】：R49;TN911.7;TP183

【相似文献】