基于WCSP-LSTM的脑电运动想象意图识别研究

Research on EEG Motor Imagery Intention Recognition Based on WCSP-LSTM

基于皮层脑电图的脑机接口系统中,由于皮层脑电信号采集难度大数据量少导致提取特征和分类准确率不理想。为此,提出重叠式加窗的共空间模式和长短期记忆(Window CSP-LSTM,WCSP-LSTM)深度神经网络结合对病例的皮层脑电进行深层次特征扩充和提取,利用全连接进行运动想象意图识别,同时用相位锁定值来构建脑功能网络分析样本的脑功能机制。所提出的算法最高识别准确率达到100%,十折交叉验证平均准确率为93.423%。通过脑网络分析表明,在执行运动想象过程中病例的顶-枕叶连接更紧密。结果表明,WCSP-LSTM算法在基于脑机接口系统的癫痫康复方面具有巨大的潜力。

In the brain-computer interface system based on electrocorticography, the feature extraction and classification accuracy are not ideal due to the difficulty of collecting electrocorticography signals and the small amount of data. To this end, a co-spatial pattern of overlapping windowing and long short-term memory(Window CSP-LSTM, WCSP-LSTM) deep neural network are proposed to expand and extract deep features from the cortical EEG of cases, and use full connections for motor imagery intent recognition, while using the phase locking value to construct the brain function mechanism of the brain function network analysis samples. The highest recognition accuracy of the proposed algorithm is 100%, and the average accuracy of ten-fold cross-validation is 93.423%. Brain network analysis showed that the parietal-occipital connections of the cases were tighter during the execution of motor imagery. The results show that the WCSP-LSTM algorithm has great potential in epilepsy rehabilitation based on brain-computer interface systems.