基于通道选择的多尺度Inception网络的脑电信号分类研究

Research on EEG signal classification based on channel selectionand multi⁃scale Inception network

基于运动想象脑电信号的脑机接口系统有可能在大脑和外部设备之间创建通信通道。然而,特征提取的局限性、通道选择的复杂性和被试者之间的可变性使得脑电信号分类模型难以有效泛化。在这项研究中,文中提出一种端到端的深度学习模型,该模型使用并行多尺度Inception卷积神经网络在6个通道选择区域中进行多分类运动想象任务。为了解决被试者间可变性,实验进行了跨被试和跨被试微调两种评估场景。在BCI竞赛IV 2a数据集上的实验和测试结果表明:ROI F达到了98.49%的最高分类精度,比最低准确率高17.26%;且跨被试微调场景分类性能优于被试内和跨被试场景,分类准确率分别提高了1.82%和1.69%。此外,并行多尺度Inception卷积神经网络模型的平均分类准确率比单尺度Inception CNN模型高5.17%。总之,文中提出一种基于通道选择的端到端的脑电信号分类框架,可以促进高性能和稳健的脑机接口系统的开发。

A brain⁃computer interface(BCI)system based on motor imagery electroencephalography(MI⁃EEG)has the potential to establish communication pathways between the brain and external devices.However,limitations in feature extraction,complexity of channel selection and variability among subjects make it challenging for EEG signal classification models to generalize effectively.In view of this,an end⁃to⁃end deep learning model that employs a parallel multi⁃scale Inception convolutional neural network(MSICNN)in six channel selection regions of interest(ROIs)for multi⁃class MI classification.To mitigate subject variability,cross⁃subject and cross⁃subject fine⁃tuning evaluation scenarios are performed.Experimental results on the BCI competition IV 2a dataset demonstrate that ROI F achieves the highest classification accuracy of 98.49%,which is 17.26%higher than the lowest accuracy.Furthermore,the classification results of the cross⁃subject fine⁃tuning scenario outperform those of within⁃subject and cross⁃subject scenarios by 1.82%and 1.69%,respectively.Additionally,the average classification accuracy of the parallel multi⁃scale Inception CNN model is 5.17%higher than that of the one⁃scale Inception CNN model.In summary,it is an end⁃to⁃end EEG signal classification framework based on channel selection,which can facilitate the development of high⁃performance and robust BCI systems.