计算机应用 ›› 2021, Vol. 41 ›› Issue (4): 1042-1048.DOI: 10.11772/j.issn.1001-9081.2020081300

所属专题: CCF第35届中国计算机应用大会(CCF NCCA 2020)

• CCF第35届中国计算机应用大会(CCF NCCA 2020) • 上一篇    下一篇

基于深度卷积网络的运动想象脑电信号模式识别

霍首君, 郝琰, 石慧宇, 董艳清, 曹锐   

  1. 太原理工大学 软件学院, 山西 晋中 030600
  • 收稿日期:2020-08-25 修回日期:2020-09-23 出版日期:2021-04-10 发布日期:2020-11-05
  • 通讯作者: 曹锐
  • 作者简介:霍首君(1994—),男,山西运城人,硕士研究生,主要研究方向:深度学习、脑机接口;郝琰(1996—),男,山西太原人,硕士研究生,主要研究方向:脑科学、人工智能;石慧宇(1996—),男,山西大同人,硕士研究生,主要研究方向:脑科学、人工智能;董艳清(1998—),女,山西运城人,硕士研究生,主要研究方向:脑科学、人工智能;曹锐(1982—),男,山西太原人,副教授,博士,CCF会员,主要研究方向:脑科学、人工智能、虚拟现实。
  • 基金资助:
    国家自然科学基金资助项目(61672374);山西自然科学基金资助项目(201901D111093)。

Pattern recognition of motor imagery EEG based on deep convolutional network

HUO Shoujun, HAO Yan, SHI Huiyu, DONG Yanqing, CAO Rui   

  1. College of Software, Taiyuan University of Technology, Jinzhong Shanxi 030600, China
  • Received:2020-08-25 Revised:2020-09-23 Online:2021-04-10 Published:2020-11-05
  • Supported by:
    This work is partially supported by the National Natural Science Foundation of China (61672374), the Natural Science Foundation of Shanxi Province (201901D111093).

摘要: 针对运动想象脑电信号(MI-EEG)分类准确率普遍偏较低的问题,引入基于深度框架的卷积神经网络模型(CNN)。首先,使用短时傅里叶变换(STFT)和连续小波变换(CWT)得到两种不同解析度下的时频信息;然后将其与电极通道位置信息相结合并以三维张量的形式作为CNN的输入;其次,设计了两种基于不同卷积策略的网络模型MixedCNN和StepByStepCNN来分别对两种形式的输入进行特征提取和分类识别;最后,针对因训练集样本过少而易发生的过拟合问题,引入mixup数据增强策略。在BCI Competition Ⅱ dataset Ⅲ数据集上的实验结果表明,CWT得到的样本集通过mixup数据增强后送入MixedCNN网络训练出的模型的识别准确率最高(93.57%),相较于另外四种分析方法:公共空间模式(CSP)+支持向量机(SVM)、自适应自回归模型(AAR)+线性判别分析(LDA)、离散小波变换(DWT)+长短期记忆网络(LSTM)、STFT+堆栈自编码器(SAE)分别提高了19.1%、20.2%、11.7%和2.3%。所提方法可以为MI-EGG分类任务提供参考。

关键词: 脑机接口, 运动想象, 时频分析, 卷积神经网络, 数据增强, 深度学习, 脑电信号, 模式识别

Abstract: Concerning the low classification accuracy of Motor Imagery ElectroEncephaloGram(MI-EEG), a new Convolutional Neural Network(CNN) model based on deep framework was introduced. Firstly, the time-frequency information under two resolutions was obtained by using Short-Time Fourier Transform(STFT) and Continuous Wavelet Transform(CWT). Then, it was combined with the channel position information and used as the inputs of the CNN in the form of three-dimensional tensor. Secondly, two network models based on different convolution strategies, namely MixedCNN and StepByStepCNN, were designed to perform feature extraction and classification recognition of the two types of inputs. Finally, in order to solve the problem of overfitting due to insufficient training samples, the mixup data augmentation strategy was introduced. Experimental results on BCI Competition Ⅱ dataset Ⅲ showed that the model performed highest accuracy by training the CWT samples reconstructed by mixup data augmentation on MixedCNN(93.57%), which was 19.1%, 20.2%, 11.7% and 2.3% higher than those of the other four analysis methods including Common Spatial Pattern(CSP) + Support Vector Machine(SVM), Adaptive Autoregressive Model(AAR) + Linear Discriminant Analysis(LDA), Discrete Wavelet Transform(DWT) + Long Short-Term Memory(LSTM), STFT + Stacked AutoEncoder(SAE). The proposed method can provide a reference for MI-EEG classification tasks.

Key words: Brain-Computer Interface (BCI), motor imagery, time-frequency analysis, Convolutional Neural Network (CNN), data augmentation, deep learning, ElectroEncephaloGram (EEG), pattern recognition

中图分类号: