【摘要】 针对脑机接口系统中P300电位识别正确率不高的问题,提出一种基于改进卷积神经网络的P300事件相关电位分类识别方法。通过将传统卷积神经网络中第二个串行连接的卷积层改为3个并行连接的卷积层,可加大网络宽度,提升网络对P300信号特征提取的能力;将提取的特征经全互连层组合后,采用sigmoid函数构建P300事件相关电位分类器。针对脑机接口竞赛数据中靶刺激与非靶刺激数据量不平衡的问题,采用过抽样方式,对含有P300事件相关电位的脑电数据做部分平均来增加数据量,其训练集和测试集样本量分别为25 500和18 000。采用Adam优化方法,有监督地训练这种改进的卷积神经网络。结果表明,相比传统的卷积神经网络,该方法在实验次数大于11次时,字符识别正确率均高于95%,这对于脑机接口的应用具有重要的意义。更多还原

【Abstract】 To improve the recognition rate of P300 potentials in the brain computer interface system, a novel method based on the improved convolutional neural network was proposed. By transforming the second serially connected convolutional layer of a traditional convolution neural network to three parallel connected convolutional layers, the method widens the network to enhance the ability of feature extraction in the proposed network. Combining the extracted features with the fully connected layers and sigmoid function, a P300 visual evoked potential classifier was constructed. Targeting to the problem of unbalanced data volume between target and non-target stimulus data in BCI competition, this paper adopted an oversampling method. To increase the amount of data, this paper partially averaged the EEG data that contains P300 visual evoked potentials. The training set and test set sample sizes were 25500 and 18000, respectively. Adam optimization method was adopted to train the improved convolutional neural network supervisely. The analysis results showed that the proposed network achieved an accuracy of higher than 95% when the number of experiments was over 11 times, which is of great significance for the application of brain-computer interface.更多还原