应用多尺度混合卷积网络的脑电信号特征提取与识别  

Electro-Encephalogram Feature Extraction and Recognition Using Multi-Scale Hybrid Neural Network

作　　者：王蒙昊 方慧娟[1,2] 龚亨翔 罗继亮 WANG Menghao;FANG Huijuan;GONG Hengxiang;LUO Jiliang(College of Information Science and Engineering,Huaqiao University,Xiamen 361021,China;Fujian Engineering Technology Research Center of Motor Control and System Optimal Schedule,Huaqiao University,Xiamen 361021,China)

机构地区：[1]华侨大学信息科学与工程学院,福建厦门361021 [2]华侨大学福建省电机控制与系统优化调度工程技术研究中心,福建厦门361021

出　　处：《华侨大学学报（自然科学版）》2023年第5期628-635,共8页Journal of Huaqiao University(Natural Science)

基　　金：国家自然科学基金资助项目(61973130)。

摘　　要：为了解决脑电信号特征提取能力不足导致的分类准确率不高的问题,提出一种全新的混合神经网络模型(EEG-MSTNet模型),实现脑电信号的时-频-空域特征提取和识别.首先,EEG-MSTNet模型采用一种适合脑电信号特点的多尺度卷积,提取4组不同大小卷积核的特征,并拼接在一起,从而增强对原始脑电信号的时频域提取能力.其次,通过通道注意力机制进一步提取信号的空间特征和高维时域特征,最终用于脑电信号识别.EEG-MSTNet模型在BCI Competition Ⅳ Dataset 2a数据集上进行测试,结果表明:EEG-MSTNet模型的每个模块都对分类准确率的提升做出了贡献,最高分类准确率为95.83%,平均准确率为83.52%,明显优于其他模型.In order to solve of low classification accuracy problem caused by insufficient feature extraction ability of electro-encephalogram signals,a novel hybrid neural network model(EEG-MSTNet model)is proposed to achieve time-frequency-spatial domain feature extraction and recognition of EEG signals.Firstly,EEG MSTNet model adopts a multi-scale convolution that is suitable for the characteristics of EEG signals,four sets of features of different sizes convolutional kernels are extracted,and they are concatenated together to enhance the time-frequency domain extraction ability of the original EEG signals.Secondly,the spatial features and high-dimensional temporal domain features of the signals are further extracted through the channel attention mechanism,and ultimately used for EEG signals recognition.The EEG-MSTNet model is tested on the BCI CompetitionⅣDataset 2a dataset,the results show that each module of the EEG-MSTNet model contri-to the improvement of classification accuracy,with a maximum classification accuracy of 95.83%and an average accuracy of 83.52%,which is significantly better than that of the other models.

关 键 词：卷积神经网络 通道注意力机制 脑电识别 特征提取 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]