基于通道组合-数据对齐-多尺度全局CNN的MI-EEG分类

Research on MI-EEG classification based on channel combination-data alignment-multiscale global CNN

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。

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.