基于时空特征卷积神经网络的运动想象脑电信号识别方法

RECOGNITION METHOD OF MOTOR IMAGINATION EEG SIGNAL BASED ON TEMPORAL-SPATIAL FEATURES CONVOLUTION NEURAL NETWORK

为了正确提取脑电信号的特征信息以提高大脑运动想象的识别准确率,提出一种基于时空特征的卷积神经网络(TSCNN)方法。通过一个时间卷积神经网络和一个空间卷积神网络,自主学习最优的时空滤波器参数,之后再通过卷积神经网络进一步提取不同尺度的脑电信号特征信息,完成运动想象分类识别,并在小样本情况下通过信号分段、膨胀卷积等策略进一步优化训练效率和识别性能。该方法在BCI Competition IV Dataset 2a数据集上取得78.8%的准确率和0.72的kappa系数,相比其他方法可以在不进行预处理及额外特征提取的情况下,取得更好的识别效果。

In order to extract the feature information of EEG signal correctly to improve the recognition accuracy of brain motion imagination,a temporal-spatial features convolutional neural network(TSCNN)method is proposed.The optimal parameters of the time-space filter was learned through a time convolutional neural network and a space convolutional neural network,and then this method completed the recognition of motion imagination with another convolution neural network to extract further different scales of EEG signal feature information.Training efficiency and recognition performance was optimized in the case of small samples through signal segmentation,expansion convolution and other strategies.The method was tested on BCI competition IV dataset 2 A,and the accuracy was 78.8%and kappa coefficient was 0.72.Compared with other methods,the TSCNN method can achieve better recognition effect without preprocessing and additional feature extraction.