Research On Motor Imagination EEG Recognition Based On AlexNet

Brain-computer Interface（BCI）plays an increasingly important role in areas such as rehabilitation,educational care and military warfare by providing the brain without the help of the body’s normal nerve and muscle systems,providing a direct link to control of external devices.However,the E-signal has the characteristics of low signal-to-noise ratio and strong randomness,which is more difficult to identify the E-signal,and it is more difficult to identify with the increase of classification.Many traditional non-end-to-end ECE signal classification methods split feature extraction and feature classification into two independent stages of weak correlation,including feature extraction algorithm and classification algorithm.These methods require artificial selection of characteristics,lack of sufficient prior knowledge,feature type is relatively single,resulting in the identification accuracy is difficult to further improve.In view of the problems and challenges in the current ECE research,this thesis further explores the study of ECE identification of motion imagination by using end-to-end ECE recognition method based on deep learning.The main research work of this thesis includes the following three aspects:（1）In order to overcome the problem of neural network overfitting caused by the small number of training samples,the time frequency domain conversion of EEG signal is carried out by using the short-term Fourier algorithm.Then the Gaussian white noise is added to the frequency domain signal,and then converted back to the time domain by the inverse short-term Fourier algorithm,which produces more training data,and finally solves the problem of insufficient generalization ability in model training.（2）In this thesis,a method of electroencephaly of shallow neural network based on Lennet is proposed.In order to solve the problem that the characteristic distribution changes easily when the data passes through the middle layer of the neural network,the technique of batch normalization（BN）is added to the network model,and the influence of the batch normalization layer on the accuracy of model recognition is explored on the same data set.The experimental results show that the data of the middle layer is beneficial to the training of the network after the data normalization of the middle layer is used,which improves the recognition effect of the network and lays the foundation for the convolutional network structure of more layers.（3）In order to make full use of the space-time information contained in the original ECE signal,a deep neural network ECI identification model based on Alex Net is proposed.On the basis of the previous experiment,the LRN in the network model is replaced with BN,which prevents the network model from overfitting during training and thus accelerates convergence.In order to reduce parameters and speed up model training,the full connection layer is replaced with the global average pooling layer.In the comparative experiment with the model recognition effect based on Lennet,it is proved that the deep convolutional neural network can extract more space-time features from EEG signal and obtain high recognition accuracy.