Feature Extraction And Recognition Of EEG Based On Semi-supervised Deep Learning

Exploring the working mechanism of the brain has always been an important challenge for human beings.The human brain communicates with the outside world through peripheral nerve and muscle channels.The Brain Computer Interface(BCI)is designed to provide a bridge between the human brain and the external environment,but does not rely on the normal output path of the brain.Through the processing of the collected EEG signals,the different mind states are converted into different control commands to complete the control of the external devices.EEG signals are behavioral outputs that produce a series of changes in brain electrical rhythms due to stimulation of brain neurons,and are often used as theoretical basis for BCI systems based on motor imaging.In view of the problem that supervised learning can easily lead to the waste of unlabeled samples and manual feature extraction can easily lead to the loss of EEG information,the main direction of this paper is to adopt semi-supervised feature learning method.This paper focuses on the feature extraction and classification of EEG signals.Aiming at the time-varying feature loss of EEG signals and the the effectiveness of feature extraction of existing EEG-related deep neural networks,a new semi-supervised EEG recognition method based on sparse self-encoding and convolutional neural networks is proposed.Firstly,the sparse autoencoder is used to extract the sparse features of onedimensional EEG signals,and the sparse features are learned by iteration of error backpropagation method,which is used as the input of convolution neural network.Then,combined with the idea of dense convolutional neural network and multi-scale convolutional neural network,a multi-scale dense convolutional neural network is proposed to learn the high-level abstract features of EEG signals.This method has high recognition accuracy for motor imagery EEG signals.In addition,the traditional Gaussian Bernoulli restricted Boltzmann machine is the generation model of unsupervised training.In order to reduce the over-fitting risk of model training,this paper adds sparsity constraints to the training from the proposed energy function;Combined with convolutional neural network(CNN),a convolution deep belif network based on Gaussian Bernoulli restricted Boltzmann machine(GBCDBN)is proposed.According to the deep recurrent neural network,the residual block is added and the BN block is added in series after the feature is added,so that there is not much difference in the parameter magnitude of each layer,and the numerical interval is controlled,thereby improving the the stability of deep network training.