| 引用本文: | 张锦,刘熔,田森,等.面向癫痫脑电的简化深度学习模型.[J].国防科技大学学报,2020,42(6):106-111.[点击复制] |
| ZHANG Jin,LIU Rong,TIAN Sen,et al.Simplified deep learning model for epilepsy electroencephalogram[J].Journal of National University of Defense Technology,2020,42(6):106-111[点击复制] |
|
| |
|
|
| 本文已被：浏览 6328次 下载 5490次 |
| 面向癫痫脑电的简化深度学习模型 |
| 张锦1,刘熔1,田森2,陈胜1,魏建好3 |
|
(1. 湖南师范大学 信息科学与工程学院, 湖南 长沙 410081;2. 湖南师范大学 数学与统计学院, 湖南 长沙 410081;3. 湖南大学 信息科学与工程学院, 湖南 长沙 410082)
|
| 摘要: |
| 针对脑电信号随机性强、动态变化迅速等特点,提出了一种简化深度学习模型研究癫痫脑电识别问题。提出的模型以一维卷积神经网络为基础,在结构方面简化了卷积层、池化层等以提高模型效率,在整体框架方面应用了Keras框架,在训练优化算法方面采用RMSProp算法作为模型优化算法,通过预定义的目标函数来进行损失估计,模型设计上加入了批标准化层和全局均值池化层。基于所提模型,从三个方面研究了癫痫脑电识别问题,即:利用经验模态分解,分别选取前三阶、前五阶、前七阶、前八阶的本征模态函数分量,在简化模型上进行对比分析；利用提出模型所具备的深度学习特点,直接识别原始脑电信号而无须特征提取环节；增加了三种不同方法分别提取7类特征,对相同的脑电数据进行对比分析。性能分析结果表明:对于五类不同的脑电信号,前三阶的本征模态函数分量的识别率达到92.1%,比其他几种处理方式识别率高；前八阶的本征模态分量识别率不及原始信号,表明人工数据处理时会给数据带来噪声; 所提出的简化深度学习模型能高效处理癫痫脑电识别问题,具备较高效率和较好性能。 |
| 关键词: 癫痫脑电 卷积神经网络 Keras框架 经验模态分解 |
| DOI：10.11887/j.cn.202006013 |
| 投稿日期：2019-09-24 |
| 基金项目:国家教育部产学合作协同育人基金资助项目(201702001043,201801037136,201901051021); 湖南省教育厅创新平台开放基金资助项目(15K082); 湖南省研究生教改基金资助项目(JG2018A012); 湖南省交通运输厅科技进步与创新计划资助项目(201927) |
|
| Simplified deep learning model for epilepsy electroencephalogram |
| ZHANG Jin1, LIU Rong1, TIAN Sen2, CHEN Sheng1, WEI Jianhao3 |
|
（1. College of Information Science and Engineering, Hunan Normal University, Changsha 410081, China; 2. College of Mathematics and Statistics, Hunan Normal University, Changsha 410081, China;3. College of Information Science and Engineering, Hunan University, Changsha 410082, China）
|
| Abstract: |
| A simplified deep learning model was proposed to solve the problem of recognition based on the strong randomness and rapid dynamic change of EEG(epilepsy electroencephalogram) signals. The proposed model utilizes one-dimensional convolutional neural network, which simplifies the convolutional layers and pooling layers to improve the efficiency. Based on the overall Keras framework, the RMSProp algorithm was used for the model in the training process, and the algorithm estimated the loss through a predefined objective function. The model design incorporated a batch normalization layer and a global mean pooling layer. The EEG recognition was researched from two aspects based on the proposed model:with empirical mode decomposition, the first three orders, the first five orders, the first seven orders, and the first eight orders of intrinsic mode functions were selected for comparative analysis on the simplified model. Because of deep learning characteristics, the proposed model can directly recognize the original EEG signals without feature extraction. After extracting 7 types of features,it adds three different methods to compare the accuracy. The experimental results show that:the recognition rate of the first three orders of intrinsic mode function reaches the level of 92.1% for the five different types of EEG signals, which is higher than that of other features. The first eight orders′ recognition rate is lower than the original signal, which indicates that data preprocessing will lead to the noise. The proposed simplified deep learning model can effectively deal with the epileptic EEG recognition problem with higher efficiency and better performance. |
| Keywords: epilepsy electroencephalogram convolutional neural network Keras framework empirical mode decomposition |
|
|
|
|
|
