Multi-class Motor Imagery EEG Signal Classification And Application Based On Deep Learnin

The brain-computer interface system based on motor imagery(MI)plays a certain role in motor function recovery of stroke patients.Therefore,the high-precision classification of motor imagery electroencephalogram(EEG)signal is of great significance.With the increase of motor imagery types and subjects,the classification performance of the algorithm needs to be improved.This thesis studies the classification of MI-EEG signals based on deep learning method.The main work of this thesis is as follows:1.Summarize the classification algorithm of MI-EEG signal based on deep learning.The research on the classification of MI-EEG signal using deep learning methods at home and abroad in recent years is summarized.The main contents are as follows: The forms of input data mainly include original data,images and extracted features.Data augmentation methods mainly include sliding window,adding noise,adversarial generative network,segmentation and reorganization,and neural network models mainly include CNN and RNN,and the mixture of the two.2.Propose an end-to-end classification model based on deep learning.First of all,this thesis proposes a new sample representation method based on the characteristics of multi-channel time series signals of EEG signal,which converts multi-channel signals into single-channel processing.This representation method can increase the number of input samples of the neural network while reducing the amount of calculation.Then,a multi-layer one-dimensional convolutional neural network(CNN)is designed,which can automatically learn the hidden time-frequency information related to different motor imagery tasks in the EEG signal,without the need to manually design feature extraction methods,reducing the uneven classification results caused by the algorithm on different participant datasets caused by human experience.Finally,considering that motor imagery is a continuous state in time,it is proposed to add a Gated Recurrent Unit(GRU)variant of the Recurrent Neural Network(RNN)on the basis of a single CNN model,cascade CNN-GRU and parallel CNN-GRU hybrid network models are designed respectively.On the basis of extracting time-frequency information,the hybrid network model increases the time-dependent information existing between sampling points and improves the classification performance.Experimental results show that the performance of the proposed methods in this thesis is better than that of advanced algorithms on public dataset.In addition,the hybrid network model has stronger antinoise ability,the effect on actual data is better than the single CNN model,and the cascade CNN-GRU model performs most prominently.3.Complete the actual data collection and control application.An audio-guided motor imagery experiment paradigm is designed,and the MI-EEG signal of 14 subjects are collected as the actual dataset.The proposed algorithm is used to classify the actual collected dataset,and the highest accuracy is 90%.Finally,the four-classification algorithm of MI-EEG signals is applied to the control of smart car,and the output of the classification result is transformed into commands for controlling the front,back,left,and right movement of the smart car.