The Motor Imagery EEG Classification Based On Deep Learning

The research on electroencephalograph(EEG) classification has a great contribution to the development of brain computer interface(BCI) research.The application of BCI depends on the accuracy and robustness of EEG classification. However, EEG is easily disturbed by noise and other signal sources(Such as: EMG, EOG, ECG, etc). The accuracy of EEG classifier is difficult to improve and the ability of generalization is poor. So it is necessary to study how to select the features of EEG accuretly, and how to gain the classifier with a better generalization ability. In recent years, deep learning as a multi-level neural network model has been applied effectively in image classification and speech recognition field. Thus how to apply the method of deep learning to EEG classification become the hotspot in BCI research.In this article, we compare traditional classifier with deep learning methods when applying them to classify motor imagery EEG.Our research aims at analysising the useage of sparse auto encoder(SAE) and convolutional neuro network(CNN) model and the difference between traditional classifier and deep learning methods. Our work mainly divided into three parts:During the first part, we mainly works on signal preprocess,we sort EEG data based on the statistical parameters,using filters to remove noise and baseline drift, intercept specific band from original data.Then we use Independent Component Analysis(ICA) method to remove the electro-oculogram(EOG) interference. In the second part we select the features of EEG by applying wavelet transform on preprocessed data. Then we puts the former result as features into the support vector machine(SVM) and softmax neural network to obtain a training model. The third part aims at applying deep learning methods to EEG classification. Using CNN and SAE models compare with traditional classifier then discuss their applicability.We found that softmax artifical neuro network(ANN) cost more time than SVM when achieving the same classification accuracy. However, softmax method is more suitable on the multi-labeled data, and has better proformance during EEG classification task. Compared with traditional classfier,CNN and SAE can directly obtain a model by integrating sparse coding method without manual feature extraction process. However,CNN and SAE require longer training time and computers with heigher performance than traditional methods.