Motor Imagery EEG Recognition Based On Conditional Emd And Improved CNN

Brain Computer Interface(BCI)is a kind of technology that does not rely on other tissues of the body,and directly converts the Electroencephalography(EEG)signal recorded from human scalp into computer control commands to control external devices.It has application prospects in medical rehabilitation,life,military and other fields.The difficulty of this technology is how to accurately identify the user’s intention from the EEG signal.Since the EEG signal is non-linear,unstable and high-dimensional,its modeling and analysis is particularly difficult,and the collected EEG signal usually contains noise interference such as electrooculogram.In this paper,the Empirical Mode Decomposition(EMD)algorithm which can adaptively decompose the signal is selected to remove the noise of EEG signal.Considering the strong feature extraction ability of Convolutional Neural Network(CNN),it is selected as the feature extraction and classification method of EEG signal.In order to solve the problem that traditional EMD algorithm selects Intrinsic Mode Function(IMF)based on the experience of researchers,the conditional empirical mode decomposition algorithm is proposed to remove the noise in EEG signal,in which the correlation coefficient between each IMF and the original signal is used as the first condition,and the relative energy occupancy rates between the IMFs are the second condition.In addition,to make full use of the effective features of multi-channel EEG signal,a serial-parallel convolutional neural network is proposed to recognize EEG signal.In this network,two convolution kernels are respectively used to extract the features within each channel and the features between channels.The experimental results show that,the proposed method can achieve higher recognition accuracy,compared with other common algorithms.Considering that EEG signal is a kind of sequence signal,one-dimensional convolution operation which is more suitable for extracting its features,and according to the characteristics of EEG signal,this paper proposes an EEG signal combination method to overcome the problem that one-dimensional convolution operation can not extract the correlation information between EEG signal channels.EEG signals have local detailed features and relatively overall features.Therefore,in this paper,the idea of multiple scales of convolution kernels instead of a single convolution kernel is used to improve CNN.And then a one-dimensional multi-scale convolutional neural network is proposed to recognize EEG signals.In addition,the conditional empirical mode decomposition algorithm is also used to filter the EEG signal.The experimental results show that the proposed method can get the highest recognition accuracy,compared with other common algorithms.Finally,based on the proposed algorithm,an intelligent wheelchair BCI system is designed.The online experimental results show that the proposed scheme is a feasible design scheme of intelligent wheelchair BCI system.