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

As a new human-computer interaction technology,brain-computer interface(BCI)is a multidisciplinary research field involving neuroscience,signal processing,and pattern recognition.The BCI technology was originally developed to assist disabled individuals with physical disorders in rehabilitation training.Nowadays,it has broad prospective applications in assisted medical diagnosis,industrial control,and entertainments.The BCI system is organized in a variety of categories according to the EEG generation mechanism,among which the BCI system based on Motor Imagery(MI)is considered to be the most promising one.Traditional classification methods are mostly divided into two independent stages: feature extraction and signal classification.Feature extraction algorithms are used to extract characteristics of EEG in time domain,frequency domain and spatial domain.Then,machine learning models are established to map these characteristics to the target category.The best classifier is finally chosen among several models.Traditional methods require manual selection of features.However,the classification accuracy based on such methods is hardly to be improved due to the lack of prior knowledge and the simple types of feature,accompanied with the large differences among individuals.In order to solve the above problems,a wavelet time-frequency image is proposed as the input of the classifier,preserving the time-frequency information and the relative position information of the electrodes.The convolutional neural network is used as the classifier for time-frequency images,making full use of the nonlinear representation and feature extraction capabilities of neural networks.Under the support by the national natural science foundation of China,the present research focuses on the motor imagery EEG signals.A motor imagery EEG classification method based on wavelet time-frequency image and convolutional neural network is proposed.Distinguished from the feature in traditional vector form,the time-frequency image is generated based on the wavelet transform and multi-resolution analysis theory.The time-frequency images of multiple channels are combined as the input of convolutional neural network.According to the scale of data sets,a neural network with 2 convolution layers is designed,and two convolution kernels of different sizes,1D-Filter and 2D-Filter,are adopted.On BCI 2003 data set Ⅲ,the results based on wavelet time-frequency images of different channels,different wavelets,and the wavelet with different parameters are compared,the best classification accuracy is improved to 92.75%.By comparing the results of BCI 2008 data set 2a with other algorithms,the average accuracy of the 9 subjects is 81% and the standard deviation is reduced to 7%.The 1D-Filter has the higher accuracy than 2D-Filter,but 2D-Filter has the lower deviation.The results obtained on the data set from motor imagery experiment are consistent with the conclusions reached by the former groups of data,the average classification accuracy is increased by 9% in comparison with the traditional CSP algorithm.These results illustrate that the proposed method,as a new form of motor imagery EEG classification method,has higher classification accuracy and a smaller individual deviation,providing a new idea and method for the research of EEG classification.