Research On Classification Method Of Motor Imagery EEG Signals Based On Feature Expression And Separable Convolution

As one of the important paradigms of EEG signals,motor imagery has broad application prospects in medical rehabilitation,military transportation,games and entertainment,etc.Therefore,it is of great significance to carry out research on the identification of motor imagery EEG signals.When the traditional methods are used to extract the EEG signal features of motor imagery,most studies only consider the information of a single dimension or two dimensions in the time domain,frequency domain and spatial domain information,and rarely explore the internal relationship between the dimensions,and it is difficult to achieve a comprehensive extracting EEG signal features;when using deep neural networks for motor imagery EEG classification,most studies do not take into account the relationship between efficiency and accuracy.Because the deep networks are difficult to deploy in mobile devices,and the lightweight networks do not fully extract EEG Deep features of the signal.Therefore,this paper focuses on the above problems,and the main contents are as follows:(1)Aiming at the problem that the traditional feature extraction method is single and the features are poorly recognizable independently,an EEG signal classification method MD-LSTM based on multi-domain feature expression is proposed.The method uses discrete wavelet transform to reorganize the data to remove high-frequency noise and low-frequency interference;uses empirical mode decomposition to slice the reorganized signal,selects appropriate intrinsic mode functions to extract time-frequency-energy features;uses common spatial pattern method to extract spatial features.Finally,the local features are combined,and LSTM is used to filter out the most discriminative global features.Experiments are performed on the BCIC IV dataset 2a and 2b datasets,and the classification accuracy reaches 77.91% and 88.09%,respectively.The experimental results show that this method can effectively extract and fuse the features of different domains compared with other feature extraction methods.(2)Aiming at the problems of single convolution scale and difficulty in deep feature extraction in light-weight motor imagery network,a motor imagery EEG classification method Ghost-DSNet is proposed based on ghost module and depthwise separable convolution.The method uses multi-scale convolution to automatically extract motor imagery EEG features at different time scales,replaces traditional convolution with ghost modules to reduce the amount of model parameters;uses the superposition of separable convolution units to extract deep-level features;and finally uses full connection layers for classification.Experiments on BCIC IV dataset 2a and 2b datasets,the classification accuracy is 5.84% and 6.13% higher than the baseline model.In addition,using a few key channels to improve the method,a new motor imagery EEG signal processing method MGhost-DSNet is designed,and the classification accuracy reaches 78.58% and 89.07%.The experimental results show that the Ghost-DSNet method achieves a good balance between the amount of parameters and the accuracy.The MGhost-DSNet method uses a smaller number of channels,which expands the lightweight advantage of the original method and improves the classification accuracy.