CRRT：基于改进CNN和ResRNN-Transformer的EEG数据分类网络

CRRT: EEG Data Classification Network Based on Improved CNN and ResRNN-Transformer

关键词:  脑电图  参数共享  全局注意力机制  动态位置编码

Key words:  electroencephalogram  parameter sharing  global attention mechanism  dynamic positional encoding

摘要:

为解决脑电图(Electroencephalogram，EEG)数据复杂且易受噪声干扰、传统分析方法难以高效处理的问题，本研究提出一种基于动态位置编码的CRRT模型。该模型通过残差循环神经网络(Residual Recurrent Neural Network，ResRNN)捕捉EEG数据的时序依赖性，将ResRNN的输出作为动态位置编码输入Transformer，提升其对EEG数据全局信息的建模能力；同时，通过共享参数降低模型参数量，提升EEG数据分类的精度。为验证所提出模型的有效性，在4个公开的EEG数据集上进行实验，并将CRRT模型与其他的先进方法进行对比，评估其在不同数据集上的性能表现。实验结果表明，CRRT模型在运动识别分类数据集BCI Competition Ⅳ Dataset 2a和BCI Competition Ⅳ Dataset 2b上的平均准确率分别达81.09%(Kappa值为0.745 3)、87.66%(Kappa值为0.725 3)，在SJTU情感脑电数据集(SEED数据集)上的分类准确率为97.31%(Kappa值为0.957 4)，在DEAP数据集上的效价分类准确率和唤醒分类准确率分别为99.37%、99.39%。上述结果表明，CRRT模型在EEG数据分类任务中的准确性和效率得到显著提高，为脑科学研究提供了有力支持。

Abstract:

Electroencephalogram (EEG) data is complex and susceptible to noise interference,and traditional analysis methods are difficult to process efficiently.In order to solve this problem,this study proposes a CRRT model based on dynamic position encoding.The model captures the temporal dependence of EEG data through Residual Recurrent Neural Network (ResRNN).The output of ResRNN is input into Transformer as dynamic position encoding to improve its ability to model the global information of EEG data.At the same time,the number of model parameters is reduced by sharing parameters to improve the accuracy of EEG data classification.In order to verify the effectiveness of the proposed model,experiments were conducted on four public EEG datasets,and the CRRT model was compared with other advanced methods to evaluate its performance on different datasets.The experimental results show that the average accuracy rates on BCI Competition Ⅳ Dataset 2a and BCI Competition Ⅳ Dataset 2b are 81.09% (Kappa value is 0.745 3) and 87.66% (Kappa value is 0.725 3),respectively.The classification accuracy is 97.31% (Kappa value is 0.957 4) on the SJTU emotional EEG dataset (SEED dataset).On the DEAP dataset,the classification accuracy of valence and arousal are 99.37% and 99.39%,respectively.The above results show that the accuracy and efficiency of CRRT classification model in EEG data classification tasks have been significantly improved,which provides a strong support for brain science research.