基于GAF和混合模型的运动想象分类研究

Research on Motor Imagery Classification Based on GAF and Hybrid Model

针对运动想象脑−机接口的分类识别问题,提出了一种结合格拉姆角场理论、卷积神经网络(Convolutional Neural Networks,CNN)和长短期记忆网络(Long Short-Term Memory,LSTM)的新模型。首先,分别使用格拉姆角场中的格拉姆角和场与格拉姆角差场算法将一维运动想象脑电信号表示为二维图像;然后,设计针对性的浅层CNN和LSTM相结合的模型来识别该图像特征,从而完成运动想象分类。在BCI Competition IV 2a公开数据集上就运动想象任务进行了四分类验证。实验结果表明,在单被试和多被试的情况下,GASF-CNN-LSTM模型和GADF-CNN-LSTM模型相比其他模型性能提升显著,准确率均达87.66%以上,最高准确率可达99.09%。且针对运动功能障碍患者数据也能表现出良好的性能。对运动想象脑电信号的时间依赖性和对应特征的图像生成表征方法进行了探讨,为运动想象脑电信号特征挖掘提供了新思路。

As a paradigm of brain-computer interface,motor imagery has a broad application prospect in the field of medical rehabilitation.Due to the non-stationarity and low signal-to-noise ratio of Electroencephalograph(EEG)signals,how to effectively extract the features of motor imagery signals and achieve accurate recognition is a key issue in the motor imagery brain-computer interface technology.Aiming at the classification and recognition problem of motor imagery brain-computer interface,this paper proposes a new method combining Gramian Angular Field(GAF)theory,Convolutional Neural Networks,and Long Short-Term Memory(LSTM).First of all,The Gramian Angular Summation Field(GASF)and the Gramian Angular Difference Field(GADF)in GAF are used respectively.GADF algorithm represents one-dimensional motor imagery EEG signals into two-dimensional images.Then,a targeted shallow Convolutional Neural Network(CNN)model is designed to realize the recognition of the image features to complete the motor imagery classification.A 4-class validation on the BCI Competition IV 2a public dataset is performed on the motor imagery task.The experimental results indicate that,in both singlesubject and multi-subject scenarios,the GASF-CNN-LSTM and GADF-CNN-LSTM models exhibit significant performance improvements compared to other state-of-the-art models.Their accuracies surpass 87.66%,with the highest accuracy reaching 99.09%.Moreover,these models demonstrate strong performance when handling data from patients with motor functional disorders,further confirming the effectiveness of the models.In this paper,the time dependence and the image generation and representation technology of the corresponding features of the motor image EEG are discussed,which provides a new idea for the feature mining of the motion image EEG.