Studies For MI-EEG Decoding Based On Representation Of Discriminative Features

Brain Computer Interfaces are applied to provide a direct communication between the human brain and external devices which can serve patients with impaired nerve conduction.Viewed as the unique voluntary paradigm,motion imagery EEG are now facing several difficulties in decoding such as low recognition rate due to insufficient analysis about representation among features,failing to forming effective global feature from abundant imformation.Unfortunately the task to extract and represent feature reasonablely and effectively is still challenging.To address these problems,this thesis focus on appropriate selection of subject-specific temporal segments and frequency bands.In addition,Considering the scarcity of training data in the decoding,we proposed a discrete sequence generation method for MI-EEG The main content and relevant contributions in this thesis are highlighted as follows:(1)We proposed a discrete sequence generation method.This method uses Monte Carlo search to complete the discrete sequences at different times.The next action about synthesis depend on the reward which is assessed through the completed discrete sequences.After certain samples formed,the identification result from the discriminator of GAN about these fake samples will guide the generator to update the parameters for better generation.The experimental results show that this method can generate EEG data which is similar to the real EEG data and provide enough training data for decoding.(2)The features extract by discrete wavelet transform(DWT)from different spectral bands are condensed by convolution neural network in temporal and spatial domain.The condensed imformation formed a new inter-feature combination through long short-term memery model.Then the independent class probabilities are infered according to the features of different frequency bands and final recognition result is obtained by fusion those independent class probabilities.This method focuses the representation of correlated features.The experimental results show that the proposed decoding model has higher recognition accuracy and better model robustness.(3)We introduce a new grouped channels expression structure.This structure will help the convolution neural network we used for channel information fusion and compression of time domain.The purpose of this method is to find out the feature subspace which can express identification information from multi-channel and improve the recognition efficiency based on multi-channel correlation analysis.The experimental results show that our approach outperforms a series of baselines and state-of-theart methods.In the analysis of channel features,we find out the correlation between different local low-level features and global and high-level features,which will lay a foundation for our further studies of MI-EEG decoding based on channel selection.In conclusion,the methods were proposed in this thesis to address the problems of automatic acquisition and reasonable expression of MI-EEG data feature.The validity and reliability of these methods were verified by conducting experiments against several existing methods.The analysis and modeling about sequential discriminative features provided a certain originality in the fields of signal processing and pattern recognition.Meanwhile,the proposed methods provide potential ability in complex MI-EEG data analysis.