Research On Algorithms Of Generative Adversarial Network In EEG Signals

In the world,stroke is not only the second leading cause of death,but also one of the leading causes of disability.Stroke is usually accompanied by motor dysfunction caused by motor nerve injury.The motor imagery brain-computer interface(MI-BCI)system is currently one of the most advanced rehabilitation technologies,and it can be used to restore the motor function of stroke patients.The deep learning algorithm in the MI-BCI system requires lots of training samples,but the electroencephalogram(EEG)data of stroke patients is relatively scarce.Therefore,artificial generation of stroke data is a promising strategy for further development of stroke clinical rehabilitation research.This thesis mainly studies the problem of generating medical EEG data by deep convolution generative adversarial network(DCGAN),and deepens the research on the credibility of artificial EEG data according to referring to previous studies.Generating effective EEG data is a very important concept in this study,which can complement the gap in the field of medical EEG data generation.The main contents are as follows:1.A method based on optimal modified S-transform(MST)features is proposed to convert one-dimensional EEG data into two-dimensional EEG spectrograms.First,the interference in EEG signal is removed by preprocessing method.Then,the MST is used to extract the features in the time-frequency domain and calculate the power spectrum of the feature data.Finally,one-dimensional EEG data is converted into two-dimensional EEG spectrograms using EEG2 Image technology.2.A DCGAN-based EEG data generation method is proposed.In the field of medical image dataset expansion,the most commonly used model is the DCGAN model.DCGAN is the first model to apply micro-stride convolutional neural network architecture to the generator.DCGAN is the first model to apply the micro-step convolutional neural network architecture to generators.Micro-step convolution can effectively visualize the features extracted by convolutional neural network.The micro-step convolution operation can improve the quality of artificial images to a certain extent.Compared with the traditional GAN model,the DCGAN model is more stable,and the generated images are more diverse,which can better meet the needs of researchers.3.Evaluation metrics for the raw and artificial EEG data is proposed.First,the raw EEG data is evaluated from two aspects,which are statistical analysis and classification performance based on fused features.Then,the artificial EEG data is evaluated through three ways: the classification performance based on convolutional neural network,analysis of Fréchet inception distance score and event-related desynchronization/event-related synchronization phenomenon.The research work of this paper can improve the problem of insufficient medical data and lay a foundation for the further development of clinical rehabilitation.