Classification Method Of Motion Imagination EEG Signals Based On Wavelet Transform And Improved CNN

Brain-computer interface technology as a new technology involving clinical medicine,neurology,mechanics and computer science and other disciplines,with the help of this technology can achieve real-time interconnection between the human brain and external devices.Currently,countries around the world are very optimistic about its future,which is believed to bring new changes in medical,defense and daily entertainment,and the classification of motion imagery through EEG signals is one of the key technologies.To address the problems that the current traditional methods of motor imagery EEG signals are complicated and the classification accuracy has not yet reached to meet the theoretical and technical requirements,this thesis carries out the research on the classification methods of motor imagery EEG signals based on wavelet transform and deep learning.The main work of this thesis is as follows:In order to improve the signal-to-noise ratio of EEG signals and subsequently improve the accuracy of motor imagery classification,this thesis first uses the ICA algorithm to remove noise artifacts and filter the signals by FIR band-pass filter,and then generates the EEG signals into characteristic time-frequency maps by wavelet transform.At the same time,in order to get the time-frequency map with stronger feature expression ability,this thesis selects the better time-frequency map by experimenting the three parameters of the time-frequency map: wavelet basis function,number of channels and time period with the help of classical machine learning SVM model.Later,based on the selection of suitable time-frequency maps,a study on the classification of motor imagery EEG signals based on deep learning was conducted.First,a simple CNN model with only two convolutional layers was built for motor imagery classification experiments,and after reasonable adjustment of its parameters,the model achieved better experimental results than the SVM model,proving that deep learning has some advantages in motor imagery EEG signal classification research.Compared with traditional machine learning,deep learning can automatically extract features with the help of convolutional kernels,which is especially advantageous when processing images.This thesis then improves the CNN model by expanding the scale and replacing some of the convolutional layers with depth-separable convolutions to reduce the model parameters and improve the model performance,and also introduces dense connection modules and dilation convolutions to enhance the feature utilization and accelerate the model convergence depth,so that the model performance can be improved while reducing certain computational cost,thus keeping the model lightweight.In summary,the feature time-frequency map-based DDCNN model constructed in this thesis achieved an average accuracy of 93.72% in dataset A,and the best Kappa coefficient in dataset B was also higher than most methods.The results indicate that the time-frequency map-based DDCNN motor imagery EEG signal classification model proposed in this thesis achieves light weight and ensures high classification accuracy at the same time.It has some application value.