Research And Implementation Of Brain-computer Interface System Based On Convolutional Neural Network

In recent years,with the rapid development of artificial intelligence and human-computer interaction,brain-computer interfaces are gradually becoming the focus of today's technology and business.This new type of human-computer interaction allows the human brain to communicate directly with external devices,achieves interactive control of external devices,bypasses tedious intermediate devices,is convenient,and has significant research and commercial value.In the brain-computer interface system,the classification of EEG signals is the core of brain-computer interface technology,which determines whether the brain-computer interface technology can be truly practical.At present,the EEG classification method based on feature extraction requires manual extraction of data features and optimization of the classifier parameters.Such a research method requires a large amount of prior knowledge and a unified measurement standard.The classifier's robust and accurate are also greatly influenced by the feature based method.However,Convolutional neural networks in the field of artificial intelligence can automatically extract data features and classifications.Therefore,it is meaningful to study how to use the convolutional neural network theory to automatically extract features from EEG data and construct a robust and accurate EEG classifier.In order to solve the problems above,this paper designed a brain-computer interface system based on convolutional neural network.The specific work is as follows:(1)This Paper construct a 5-layer neural network classifier that classifies the left and right hand motion imaging EEG signals.This classifier based on the convolutional neural network theory and use Tensor Flow,an open-source machine learning framework to do this work.The classifier can automatically extract features from the EEG data and automatically classify them.The universality and robustness of the classifier is stronger than that of feature-based EEG classifiers.After filtering and normalizing the EEG data,the classifier can achieve a classification accuracy of75.3% on the test set of the official data set Data Sets 2b of the Fourth Brain-Computer Interface Contest(BCI IV).(2)The LRN regularization and dropout methods are introduced to reduce the degree of over fitting of the classifier on the test set to improve the generalizationcapability of the classifier model more.(3)A set of brain-computer interface system based on convolutional neural network visual evoked type is designed and implemented.The system uses visual stimulation of the user's way,so that the experimenter generates sports imaging EEG signals,and then through the convolutional neural network classifier for identification and classification.The whole system is easy to operate,and the subjects can use it correctly after simple training.The recognition rate of the system reaches 76.25%,and the recognition delay is less than 1 second,which has certain application value.