【摘要】 采用3D卷积神经网络模型,对脑电信号进行解码研究,旨在挖掘其深层的特征表达,以提高脑-机接口系统的性能.实验在获取"模拟阅读"脑-机接口系统的多维脑电信号后,将原始的通道特征构建成"脑电视频"的格式.其构造方法为:将通道按实际空间排布为二维矩阵,这样某时刻的多通道采样点在空间上形成一个"视频帧",这些空间信息在连续时间帧上的堆叠,形成"脑电视频".这种自然表达信息的方法,不仅包含大脑的空间分布信息,还反映了时间信息的关联,丰富了数据所包含的事件相关信息.借鉴图像领域特征学习的"局部感受野"和"权值共享"思想,搭建了自主学习脑电信号特征的3D卷积神经网络模型,将已打标签的脑电视频数据对模型进行训练,之后对测试集进行测试.与经典的卷积神经网络和传统的最佳单通道算法相比,分类正确率有了进一步的提高.实验表明,基于脑电视频的3D卷积神经网络能够更有效地学习脑电特征,改善了模拟阅读脑-机接口系统的性能.更多还原

【Abstract】 The 3 D convolutional neural network(CNN) model is used to decode the EEG signals, aiming at digging its deep feature representation to improve the performance of the brain-computer interface system. After obtaining the multi-dimensional EEG signal of the "Imitating-Reading" brain computer interface(IR-BCI) system, the original channel features are constructed into the format of "EEG video". The construction method is used as follows: the channels are arranged into a two-dimensional matrix by the actual space, so that at a certain moment, the multi-channel sampling points form a "video frame" in space. Those space information, stacked on the continuous time frame becomes "EEG video". This natural way of expressing information not only includes the spatial distribution information of the brain, but also the association of time information, which enriches the event-related information contained in the data. Based on the idea of "local receptive field" and "shared weights" in feature learning of image field, a 3 D convolutional neural network model is built, It can learn EEG signal characteristics automatically, train models by the tagged EEG video data and then test the test set. Compared with the classical CNN and the traditional best single-channel algorithm, its accuracy rate of classification has been further improved. Experiments show that the 3 DCNN based on EEG video can learn EEG features more effectively and improve the performance of the IR-BCI system.更多还原