Research On Classification Of Motor Imagery EEG Signals Based On Ensemble Convolutional Neural Network

As an important research direction of brain science and brain information processing,the Brain-Computer Interface(Brain-Computer Interface,BCI)is a communication system that does not rely on the normal output pathway composed of peripheral nerves and muscles.At present,more and more attention has been paid to the research of BCI.By decoding EEG(Electroencephalogram,EEG)signals,BCI technology can help people with dyskinesia complete the control of external devices.Decoding EEG signals is the key step to realize BCI technology,and the reasonable analysis and extraction of EEG signal characteristics to classify is the current research hotspot in the field of EEG decoding.In traditional methods,the recognition of Motor Imagery EEG(Motor Imagery EEG,MIEEG)signal usually needs more preprocessing and complicated feature extraction steps,and requires certain prior knowledge,and can not guarantee that the selected feature is the optimal one.Different from the traditional manual feature extraction,Convolutional Neural Networks(Convolutional Neural Networks,CNN)have excellent performance of automatic feature extraction and classification recognition.Combined with the ensemble learning method,it can improve the accuracy and reliability of classification.The recognition performance and stability of the model in the EEG motor imagery classification task were studied.The specific research contents are as follows(1)An end-to-end one-dimensional convolutional neural network model is designed based on the temporal characteristics of EEG signals.The EEG data collected by multiple channels are integrated into one-dimensional data as the input of the neural network.Two onedimensional convolution layers are used to extract the time-domain features of EEG data.Finally,two fully connected layers are used to complete the classification of EEG data.The effectiveness of the model is verified by using the BCI competition IV 2b data set.The kappa coefficient calculated by the average recognition rate of multiple training results is 0.55,which is slightly lower than the result of the second place in the competition,but the kappa coefficient calculated by the highest recognition rate is 0.73,which is better than the result of the first place in the competition.The results show that the convolution neural network model is effective for EEG motor imagery recognition and classification,and has great potential to improve.(2)Based on the analysis of one-dimensional convolution kernel parameters in the convolution neural network model,the essence of convolution operation is to filter the signal.The convolution kernel parameters are the same as the filter coefficients.Usually,the filter coefficients are transformed by a discrete Fourier transform to check the frequency response of the filter,and the band-pass characteristics of the filter can be judged.The experimental results show that the convolution neural network model can autonomously learn the effective information of specific frequency bands,and has similar band-pass characteristics as the filter,and the selection of specific frequency bands for different subjects is slightly different.(3)In order to improve the stability of model recognition results,two ensemble convolution neural network models are designed by using the classic Ada Boost and Bagging methods in ensemble learning.The average recognition rate of the CNN-Ada Boost and CNN-bagging integrated model designed in this paper is 4.00% and 5.26% higher than that of a single CNN model,respectively.The results show that the CNN-Bagging integrated model has better recognition performance,and the kappa value of the model recognition results is better than that of the competition winners and other related research results,but it needs more training time,while the CNN-Ada Boost integrated model has faster training speed,which can save a lot of computing resources and achieve relatively excellent results.The results show that the ensemble method can effectively improve the stability of the model recognition results.This research provides a new idea and reference direction for the subsequent EEG recognition research based on neural networks.