A Deep Learning-based Method For Recognition Of Motor Imagery EEG

Brain-computer interface(BCI)is a new type of human-computer integration technology with multi-disciplinary integration.At present,it is common to use EEG signals generated by specific brain activities to control external electronic devices.The basic principle and process of EEG signals are firstly collected by electrodes,and then after a series of preprocessing such as amplification,denoising and filtering,feature extraction and classification algorithm are finally decoded to control external devices.It is the key of brain-computer interface technology to extract the characteristics of EEG signals effectively and classify them accurately.It is very important to solve the problem of EEG classification in the field of brain-computer interface by using deep learning method.The research content of this thesis is mainly divided into the following two parts:(1)In view of the non-stationarity and obvious time-frequency characteristics of the motor imagery EEG signals,this thesis proposes a method for the recognition of the motor imagery EEG signals based on the S transform time-frequency image and combined with the convolutional neural network and the extreme learning machine.In the data set of BCI competition,the S transform time-frequency images of EEG signals are first obtained,and then the characteristic frequency band was extracted from the time-frequency images for combination.Finally,the combined image was used as the input of neural network to realize the recognition of motor imagery EEG signals.The S transform effectively overcomes the deficiency of Short-time Fourier transform and effectively improves the generalization performance of CNN by combining CNN with ELM.The experimental results show that the combination of the convolutional neural network and the extreme learning machine has a high recognition accuracy for the time-frequency images of left and right hand and right hand and foot motor imagery EEG signals,and has a good generalization performance across the subjects.In the classification experiment of multi-class motor imagery,the experimental model appears overfitting,which indicates that this method is not suitable for the classification task of multi-class motor imagery.(2)Compared with image and voice signals,expensive EEG signal acquisition equipment makes it difficult to collect large-scale EEG signal data.Therefore,a feasible method is to manually produce data to make up for the small amount of original data.Therefore,this thesis proposes a data augmentation method for left-hand and right-hand motor imagery EEG based on Conditional Image Synthesis with Auxiliary Classifier GANs,which effectively expands the number of time-frequency characteristic images on the basis of the existing small data sets,improves the classification performance of the Classifier,and overcomes the lack of data.