| Brain Computer Interface(BCI)technology is an interdisciplinary subject involving neuroscience,signal detection,pattern recognition and other fields.Due to the different mechanisms of EEG signal generation,brain-computer interface systems have also been differentiated into different research directions,among which brain-computer interface technology based on motor imagery has also received extensive attention.More and more algorithms and models are applied in the field of brain-computer interface,but how to improve the classification accuracy is still a difficult problem.Aiming at the non-stationarity of motor imagery EEG signals,the time localization of excitation and the distribution of frequency bands,this paper proposes a binary deep learning framework based on motor imagery EEG signals.The framework starts from the physiological characteristics of motor imagery EEG signals,and makes appropriate improvements from two aspects of feature extraction and pattern recognition,respectively.In the aspect of feature extraction,an Image Subtraction(IS)method is proposed according to the physiological characteristics of motor imagery EEG signals,which enhances the feature difference in the form of data input.First,the original EEG signal is preprocessed,and the signal is smoother by filtering out some noise,and then the time domain-frequency domain information of the EEG signal is fully extracted by wavelet transform,and the time series signal of the EEG is converted into a time-frequency image.According to the logical symmetry of the C3 and C4 channels,the time-frequency images of the EEG signal are subtracted,and the subtracted difference image is used as the input of the classifier,which not only reduces the redundancy of the input data,The characteristic difference of the input data increases again.In the aspect of pattern recognition,the difference images of EEG signals are processed by adding an attention mechanism to the classifier.First,a shallow convolutional neural network(CNN)is established as the basic classifier,and the information on the time-frequency image is automatically extracted by using the characteristics of the convolutional neural network.The attention module adaptively extracts the time position and frequency distribution information of EEG signal occurrence,reduces irrelevant noise interference,and increases the robustness of the model.Finally,the performance of the framework is evaluated on BCI Competition II dataset III and BCI Competition IV dataset 2b.The average accuracy is 79.6%,and the average kappa value is 0.592.The experimental results verify the feasibility of this framework to improve the performance of motor imagery EEG signal classification,and provide a reference for further research on motor imagery EEG signal classification. |
