由于运动想象脑机接口(MI-BCI)范式不需要视觉刺激,应用MI-BCI范式在提高人机交互系统舒适度方面具有重要意义.为实现辅助设备的异步控制,提高模型的鲁棒性,减少通道使用数量以降低BCI系统输入的复杂性,提出一种基于通道组合(channel combination,CC)-数据对齐(euclidean space data alignment,EA)-多尺度全局卷积神经网络(multiscale global convolutional neural network,MGCNN)的运动想象脑电分类方法.通过引入大脑静息状态下的脑电信号,扩展MI-BCI输出指令集;利用CC将22通道脑电数据重构为左右对称通道加中间通道的3通道形式,重构后的数据经过EA方法规范后作为网络输入;构建多尺度卷积模块与全局卷积模块,并行提取脑电信号的局部特征和ERS/ERD全局特征;利用迁移学习提升模型的解码能力.结果表明:该方法在BCI Competition Ⅳ 2a数据集上达到了99.28％的平均准确率和0.99的Kappa值,提高了运动想象脑电分类精度,为在线异步运动想象脑机接口的应用与发展作出了贡献.
Research on MI-EEG classification based on channel combination-data alignment-multiscale global CNN
Without visual stimuli, the motor imagery brain-computer interface ( MI-BCI ) paradigm plays a crucial role in enhancing the comfort level of brain-computer interaction systems. To achieve asynchronous control of auxiliary equipment, improve model robustness, and reduce the number of channels for decreased complexity in BCI system inputs, a motor imagery EEG classification method based on a channel combination ( CC )-euclidean space data alignment ( EA )-multiscale global convolutional neural network(MGCNN) is proposed in this paper. The output instruction set of the MI-BCI paradigm is expanded by incorporating EEG signals from the resting state of the brain. The 22-channel EEG data is reconstructed into a 3-channel form consisting of left and right symmetric channels and an intermediate channel using CC. The reconstructed data are normalized by the EA method and used as network input. A multiscale convolution module and two global convolution modules are built to extract local features and ERS/ERD global features from the EEG signals in parallel. The decoding ability of the model is improved by using transfer learning methods. Our experimental results on the BCI Competition Ⅳ 2a dataset show the proposed model achieves an average classification accuracy of 99.28％ and a kappa value of 0.99 , improves the accuracy of motor imagery EEG classification and contributes to the application and development of online asynchronous motor imagery brain-computer interfaces.
motor imagerybrain-computer interfacechannel combinationconvolutional neural networkdata alignment
国家科技期刊平台
NSTL
万方数据
