基于降噪自编码神经网络的事件相关电位脑电信号分析方法

An event related potential electroencephalogram signal analysis method based on denoising auto-encoder neural network

中文关键词  脑电信号  降噪自编码  神经网络  事件相关电位

英文关键词  electroencephalogram (EEG)  denoising auto-encoder  neural network  event related potential

中文摘要

      本文提出一种基于降噪自编码神经网络事件相关电位分析方法, 首先建立三层神经网络结构, 利用降噪自编码对神经网络进行初始化, 实现了降噪自编码深度学习模型的无监督学习. 从无标签数据中自动学习数据特征, 通过优化模型训练得到的权值作为神经网络初始化参数. 其次, 经过有标签的样本进行网络参数的微调即可完成对神经网络的训练, 该方法有效解决了神经网络训练中因随机选择初始化参数, 而导致网络易陷入局部极小的缺陷. 最后, 利用上述神经网络对第三届脑机接口竞赛数据集Data set II （事件相关电位脑电信号）进行分类分析. 实验结果表明: 利用降噪自编码迭代2500 次训练神经网络模型, 在受试者A 和受试者B 样本数据叠加5 次, 10 次, 15 次三种情况下获得的分类准确率分别为: 73.4%, 87.4% 和97.2%. 该最高准确率优于其它分类方法, 比竞赛第一名联合SVM分类器(Ensemble of SVMs, ESVM)提高了0.7%, 为事件相关电位脑电信号提供了一种深度学习分析方法.

英文摘要

      In this paper, an algorithm based on denoising autoencoder neural network for event related potential analysis was proposed. Firstly, we establish a three layer neural network structure which is initialized by the denoising autoencoder. By using the unsupervised learning, the denoising autoencoder deep learning model is implemented. The weights obtained from the optimizing model are used as initialization parameters of the neural network by automatically learning of data characteristics from unlabeled data. Secondly, the training of the neural network can be completed through the fine-tuning of the network parameters with labeled data. This method effectively solves the problem of easy falling into local minimum for the neural network, which may be caused by random initialization. Thirdly, the proposed neural network was used in the competition III Data set II for classification analysis. Experimental results show that by using the denoising autoencoder neural network model under the training iterative of 2500, the average accuracy of 73.4%, 87.4%, and 97.2% was obtained between subject A and subject B in three conditions which are the data is superimposed for 5, 10 and 15 times respectively. These significant results show that our framework demonstrated superior performance in the higher classification than other methods (97.2% in comparison the highest accuracy 96.5%). In summary, we provided a denoising autoencoder neural network, which can learn more robust features from training data automatically. This deep learning model would be a new method for event-related potential EEG signals analysis.