基于变分模态分解与深度信念网络的运动想象分类识别研究

传统人工确定最优时段及最优频段的方法会造成信息遗漏进而导致运动想象识别率的降低, 因此基于脑电信号的运动想象分类研究成为了脑-机接口研究领域的难点问题。针对该问题,变分模态分解和深度信念网络被应用于运动想象分类。对脑电信号进行变分模态分解得到窄带分量,利用希尔伯特变换提取边际谱、特征频带下的瞬时能谱以及时-频联合特征; 特征融合后采用深度信念网络对高维特征降维并实现运动想象模式的识别, 避免了人工确定想象最优时段及最优频段造成的信息遗漏。实验结果表明,利用变分模态分解与深度信念网络自动提取最优时段及最优频段特征的方法有效提升了运动想象识别率。

The Classification of EEG Induced by Motor Imagery Based on Variational Mode Decomposition and Deep Belief Network

The traditional method of manually determining the optimal periods and frequency bands resulted in the omission of information and the reduction of the recognition rate of motor imagery (MI). Therefore, MI became a challenging issue in brain-computer interface (BCI). Aiming at this issue, variational mode decomposition (VMD) and deep belief network (DBN) were applied to the classification of MI. VMD was proposed to decompose the electroencephalograph (EEG) into multiple narrow band components, then the marginal spectrum, the instantaneous energy spectrum and the joint time-frequency features were extracted by Hilbert transform, then these features were fused. The DBN was proposed to reduce the dimensions of fused high-dimensional features to recognize the pattern of MI, which avoided the omission of information caused by choosing the optimal periods and frequency bands manually. The results showed that the recognition accuracy of MI was improved effectively by the proposed method based on VMD and DBN to automatically extract the optimal period and frequency bands .