Research On Electroencephalogram Signal Based On Generative Adversarial Networks

The brain is the most advanced part of the nervous system.Exploring the mystery of the brain is a great challenge for human beings in the new century.The BCI becomes an intermediary of communication between people with nerve or muscle damage and nature.With the in-depth study of brain-computer interface system,brain-computer interface has been widely used in sports rehabilitation,neural intervention,smart home,environmental interaction,game entertainment and many other fields.Motor imagery-based brain-computer interface system can help stroke patients to carry out rehabilitation training and recover motor function.Research of this system cannot do without brain data.But at present,electroencephalography data are scarce due to the tedious experiment of collecting EEG data,and the subjects need to keep a long time of attention and wear uncomfortable EEG acquisition equipment.Although the number of stroke patients is large,the data of patients are scattered in different laboratories.It’s hard to achieve data sharing.In order to further study Motor imagery-based brain-computer interface system,it is necessary to increase the number of stroke patients’ data.This thesis conducts research based on the signal processing module of brain-computer interface.The data set of stroke patients is expanded.This study mainly considers the generation adversarial networks,and the cycle-consistent adversarial networks is used to generate the electroencephalography data of stroke patients to expand the data set of stroke patients and realize the classification.The main contents are:1.In this paper,a method based on feature extraction is used to transform one-dimensional brain data to two-dimensional EEG-topography based on feature extraction.On the basis of feature extraction,the brain data is transformed into two-dimensional color EEG-topography.The main feature extraction methods are rotation invariant local binary pattern and modified S-transform.The EEG-topography obtained by this method not only retains the electrode position information,but also contains the feature information.It is more beneficial to the research of left-and-right hand motor imagery task classification algorithm.The result of this scheme is satisfactory.2.We propose a brain data generation method based on an improved cycle-consistent adversarial networks.This method improves the original cycle-consistent adversarial networks,which is mainly reflected in the improvement of the converter network structure in the generator.The improved generation network has less parameters,which greatly reduces the complexity of the algorithm and makes the network easier to train.At the same time,the possibility of over fitting is reduced.Through the comparison of the final generation results,it is proved that the improved cycle-consistent adversarial networks can obtain better generation results.3.In this paper,convolution neural network is used to analyze end-to-end brain data.The main purpose of this method is applying convolution neural networks to feature extraction and classification of motor imagery-based brain-computer interface system.This method also analyzes the effectiveness of spectrum generation from another perspective.By inputting different proportions of original data and generated data,the feasibility of brain data generation method of improved cycle-consistent adversarial networks is judged by comparing classification results.This scheme obtains good classification results and can prove the effectiveness of the generated EEG-topographies of stroke patients.The research work of this paper is conducive to further promote the development of motor imagery-based brain-computer interface systems,and promote the application of deep learning in the field of brain-computer interface.