CRRT：基于改进的CNN和ResRNN-Transformer的EEG多模态网络

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

关键词:  脑电图（EEG）  参数共享  全局注意力机制  动态位置编码

Key words:  Electroencephalography (EEG)  Parameter sharing  Global Attention Mechanism  Dynamic Positional Encoding

摘要:

脑电图（Electroencephalogram，EEG）作为一种非侵入性技术，广泛应用于临床诊断、认知科学研究和脑机接口（Brain-Computer Interface，BCI）等领域。针对EEG信号复杂且易受噪声干扰，传统分析方法难以高效处理的问题，提出了一种CRRT分类框架。该方法通过卷积神经网络CNN(Convolutional Neural Network)提取EEG信号的局部空间特征，利用残差循环神经网络ResRNN (Residual Recurrent Neural Network)捕捉EEG信号的时序依赖性，并将ResRNN的输出作为动态位置编码输入Transformer，提升其对EEG全局信息的建模能力。此外，框架通过共享参数设计降低了模型参数量，有效提升EEG信号分类的精度。在运动识别分类BCI Competition IV Dataset 2a和2b数据集上的平均准确率分别达81.09%（kappa值0.745 3）和87.66%（kappa值0.725 3）；在情绪识别分类SEED数据集上分类准确率为97.31%（kappa值0.957 4）；在情绪识别分类DEAP数据集上，效价和唤醒的分类准确率分别为99.37%和99.39%。实验结果表明，所提方法在EEG信号分析与分类任务中显著提高了准确性和效率，为脑科学研究提供了有力支持。

Abstract:

Electroencephalogram (EEG), as a non-invasive technology, is widely used in clinical diagnosis, cognitive science research, and brain-computer interfaces (BCI), among other fields. To address the challenges of EEG signal complexity and susceptibility to noise interference, which hinder efficient processing by traditional analysis methods, a CRRT classification framework is proposed. This method extracts local spatial features of EEG signals using Convolutional Neural Networks (CNN), captures the temporal dependencies of EEG signals with a Residual Recurrent Neural Network (ResRNN), and utilizes the output of ResRNN as dynamic positional encoding to input into a Transformer, enhancing its ability to model the global information of EEG signals. Additionally, the framework reduces the model's parameter count through a shared parameter design, effectively improving the classification accuracy of EEG signals. On the motor recognition classification task using the BCI Competition IV Dataset 2a and 2b, the average accuracies are 81.09% (kappa value 0.7453) and 87.66% (kappa value 0.7253), respectively; on the emotion recognition task using the SEED dataset, the classification accuracy is 97.31% (kappa value 0.9574); and on the DEAP dataset for emotion recognition, the classification accuracies for valence and arousal are 99.37% and 99.39%, respectively. Experimental results show that the proposed method significantly improves accuracy and efficiency in EEG signal analysis and classification tasks, providing strong support for brain science research.