基于CNN时-空卷积优化的EM-EEG识别方法研究

Research on EM-EEG recognition method based on CNNtime-space convolution optimization

针对当前情绪脑电信号(emotion electroencephalogram,EM-EEG)识别研究中时间域信息的时间尺度难以把握和空间域信息易被忽视致使辨识率停滞不前,以及采集EM-EEG时通道过多导致信息冗余和信息处理成本增加等问题,提出了基于CNN的时-空卷积优化融合网络进行EM-EEG识别研究。该融合网络由提取EM-EEG时域信息的长卷积(long convolution,L-Conv)CNN和提取EM-EEG空域信息的CNN并联组成,在CNN模型时-空优化中使用粒子群算法(particle swarm optimization,PSO)对时域CNN中的L-Conv尺度进行了优化,并使用短时功率谱(short time power spectrum,STPS)的相关分析方法进行空域CNN模型通道数目优化,深层且有效地提取了EEG中的时间域和空间域特征。结果表明,提出的时-空卷积优化融合CNN在SEED IV数据集上对平和、悲伤、恐惧、高兴4种情绪最终准确率可以达到90.13%,相比传统单一CNN的识别准确率提高了4.76%,并且通道数目由62路降低至33路,缩减了46.77%,证实了本方法的可行性。

In view of the current emotional electroencephalogram(EM-EEG)identification research on time scales is difficult to grasp the time domain information and the spatial domain information is easy to ignore the recognition rate is stagnant,and collect the EM-EEG with too many channels in the excessive information redundancy and increasing cost of information processing problems,it puts forward the space-time convolution based on CNN optimization study on EM-EEG identification fusion network.The fusion network is composed of a parallel long convolution(L-Conv)CNN that extracts EM-EEG time domain information and a CNN that extracts EM-EEG spatial information.Particle swarm optimization(PSO)is used in the time-space optimization of the CNN model.The L-Conv scale in CNN has been optimized,and use the short time power spectrum(STPS)correlation analysis method of the spatial CNN channel number optimization model,temporal and spatial domain features in EEG are extracted deeply and effectively.The results show that the proposed optimization of space-time convolution integration CNN on SEED IV data set for peace,sadness,fear,happy four final accuracy can reach 90.13%,compared with the traditional single CNN recognition accuracy rate increased by 4.76%,and channel number from 62 to 33 road,shrank by 46.77%,confirmed the feasibility of this method.