The Research Of EEG Classification And Recognition Based On Spatio-temporal Neural Network

Recently,brain science and neuroscience are developing rapidly.As a great challenge of brain science,it is substantial to promote the ability of feature learning based on EEG signals.Classification and recognition of EEG are important research basis for diagnosising various brain diseases such as epilepsy,anxiety,insomnia and senile dementia,as well as human behavior states such as brain-computer interface,emotion classification,fatigue detection,lie detection,etc.The research contents mainly focus on classification of motor imagery EEG,assessment of cognitive load based on EEG,epilepsy prediction and detection based on EEG,emotion recognition based on EEG.They have become a recognized mainstream research direction in the field of EEG analysis.In particular,deep learning models have made great breakthroughs in the fields of computer vision,intelligent medical diagnosis,which have attracted several attention of scholars.The deep learning architecture have been applied to various tasks of EEG classification,achieving productive results.However,it reveals the lack of sufficient ability in effectively modeling various high-dimensional features in EEG.Furthermore,the research of promotion in feature expression ability still remain significant potential.Therefore,it is of great significance to explore the analysis of EEG based on deep learning network,further resulting in great medical and commercial value.In this paper,EEG signals are utilized for the corresponding research of brain,through which four diverse classification tasks of EEG are conducted.We aim at optimizing deep learning models for EEG classification tasks,further settling the drawbacks existing in present models.Specifically,we deal with the weak ability of feature expression,poor robustness of models and difficulty in distinguishing features within the intra-and inter-class,resulting in the significant enhancement of recognition accuracy.Firstly,we adopt an improved preprocessing method and further propose a multi-scale convolution bidirectional gated recurrent units(IncepCNN-BGRU),effectively dealing with the low signal-to-noise ratio(SNR)of motor imagery and difficulty in extracting abstract and detailed features.In detail,we utilize wavelet transform and cubic spline interpolation to convert the multi-channel EEG into a series of spectrograms.Then we combine a modified convolutional neural network(CNN)and bidirectional gated loop unit(BGRU)for the construction of hybrid deep model.It reveals efficient capacity in learning spatial-temporal representations in EEG images.The result shows that our method achieve great performance in recognition accuracy compared with existing models.Secondly,we propose a convolutional bidirectional neural Turing Machine assisted with ternary-task learning framework(TT-CBNTM),efficiently extracting multidimensional robust features from the limited EEG database during the task of brain cognitive load assessment.Our model owns the ability to deal with spatio-temporal informations existing in EEG images through the ingenious combination of CNN and NTM,thus effectively promoting the multi-dimensional feature expression ability of the our model.Furthermore,we introduce the ternary-task learning framework to the proposed architecture.Through two auxiliary tasks of identity identification and recognition,we can effectively reduce the diversity of intra-class and promote capability in feature modeling of inter-class.It demonstrates that TT-CBNTM based on ternary-task regularization mechanism can achieved excellent results in the evaluating of cognitive workload.Thirdly,we propose a contractive spike-and-slab convolutional deep belief network assisting with dual-task framework,resulting in a ssEEGNet for the prediction and diagnosis of EEG-based epilepsy.It demonstrates great ability in learning covariance information high-order features existing in epilepsy EEG.Our model can learn high-order representations and capture spectral and spatial information existing in EEG spectral images.In addition,the task of identity identification can consider more about the contextual information in EEG frames during the fine-tuning stage,further reducing variation within intra-class and facilitating the training.The experimental results show that ssEEGNet has achieved remarkable results in all indicators.Lastly,we propose a novel self-attention spike-and-slab convolutional long short term memory,efficiently promoting the ability in modeling the high-order spatial features and temporal features of emotional EEG images effectively.Our model introduces spike-and-slab variables into the convolutional operation,further enhancing the capability of feature mining without the increase of network parameters.Furthermore,the model adopts a LSTM that integrates the attention mechanism,which can recode the samples according to the distribution of important information inside the samples,thus further enhancing the network’s ability to model the temporal features.In addition,for the purpose of preventing over-fitting,we adopt a multi-task learning framework,in which the main task is emotion classification,aiming to correctly identify the emotion category of the given EEG.The auxiliary tasks are the authentication and recognition of eeg spectral images respectively.The auxiliary tasks are identification and recognition of EEG spectral images respectively,further effectively enhancing the robustness of our model.Experimental results show that the proposed model achieves excellent accuracy in emotional EEG classification tasks.