摘要
为了在脑电信号鲁棒特征学习中提取更多脑电抽象和深层特征,本文在卷积长短时记忆网络的基础上提出一种深度学习混合网络。采用快速傅里叶变换将多通道的脑电信号转换为一系列具有空域、时域、频域相关信息的频谱图;将改进的卷积神经网络和神经图灵机组合搭建完成深度学习混合模型卷积神经图灵机C-NTM;通过认知工作负载脑电的分类任务对改进的模型进行评估。实验结果表明:本文所提模型在相应的数据库上取得了94.5%的准确率,优于目前在脑电分类任务中效果最好的模型。该模型能够有效地学习不同受试者之间和同一受试者不同状态时的脑电特征,实现更好的脑电鲁棒特征学习。
In order to solve the problem of extracting more abstract and deep features from EEG,the difficulty of learning robust features from an electroencephalogram (EEG) must involve how to extract more abstract representations and deep features.To solve this problem,an improved deep-learning hybrid network was proposed on the basis of Convolutional-Long Short-Term Memory/1D-Conv(C-LSTM/1D-Conv).First,the fast Fourier transform (FFT) and relevant pretreatment methods were adopted to convert multichannel EEG signals into a series of spectrograms with spatial domain,time domain,and frequency domain-related information.We then constructed a deep-learning hybrid model,C-NTM,which associates a modified Convolutional Neural Network (CNN) with a Neural Turing Machine (NTM).Finally,a cognitive workload EEG classification task was carried out to evaluate our improved model.The experimental results indicated that the proposed approach obtained a classification accuracy of 94.5%,which had a more significant effectiveness compared with state-of-the-art models.The model demonstrated that it could effectively learn EEG features between different subjects and in different states of the same subject,and achieve better EEG robust feature learning.
作者
毕晓君
乔伟征
BI Xiaojun;QIAO Weizheng(College of Information Engineering,Minzu University of China,Beijing 100081,China;College of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China)
出处
《哈尔滨工程大学学报》
EI
CAS
CSCD
北大核心
2019年第9期1642-1649,共8页
Journal of Harbin Engineering University
基金
国家自然科学基金项目(61175126)
国家国际科技合作专项项目(2015DFG12150)
关键词
脑电信号
鲁棒特征
深度学习
卷积神经网络
神经图灵机
频谱图
卷积神经图灵机
认知负载
electroencephalogram (EEG)
robust feature
deep learning
convolutional neural network (CNN)
neural turing machine (NTM)
spectrogram
C-NTM
cognitive workload