Classification For Motor Imagery Based On Deep Learning

Brain-Computer Interface(BCI)is a complex system with software and hardware that directly connects the human brain and external devices without dependency on human peripheral nerve and muscle tissue.It is of great application value in medical treatment,education,security,and other fields.Motor imagery(MI)signal is a kind of spontaneous electroencephalogram(EEG).MI-based BCI classifies the EEG signal and then converts it into communication or control signals to the external device.However,there are two problems in traditional classification algorithms for MI: inadequate utilization of multi-domain information in EEG and difficulty in constructing generalized crosssubject models.On the one hand,most of the algorithms build a model based on the feature from a particular domain,such as spatial domain,frequency domain,and time domain,which miss some information about EEG more or less,so these algorithms have low accuracy and generation ability.Therefore,making full use of features from multiple domains is a problem in EEG classification research.On the other hand,EEG signals vary widely among different individuals.When the user is new to BCI,it requires a long calibration time to get enough train data with the label for the classification model.Therefore,how to build a reliable model with less label data is an essential problem in EEG classification research.To solve the problems above,this paper proposes some methods using ensemble learning and transfer learning based on deep learning.The research of this paper is as follows:(1)To solve the problem of modeling in the multiple domains,we propose the ensemble model based on factorization machine(FM).The model consists of three base learners and a secondary learner.Firstly,The base learners build models from spatial domain,time-frequency domain,and time domain separately and then the extract the deep feature of the train data;secondly,mask the deep feature at a certain rate and get a spare feature,which feeds to the secondary learner;lastly,train secondary learner based on FM with the spare feature.The experiment shows that the ensemble model improves accuracy,and FM is better than other ensemble methods.(2)To solve the problem of modeling with multiple domains,this paper proposes a deep learning model,which is based on convolution neural networks(CNN)and long short-term memory(LSTM).Firstly,We filter the signal in frequency and space;secondly,use the wavelet transform to get the spectrogram concatenated with multiple channel’s;lastly,train the deep neural networks with the spectrogram.The model achieved 85.3% and 74.6% accuracy on BCI Competition II-3 and BCI Competitions IV-2a,respectively.(3)To reduce the label data for model training,this paper proposes a method based on transfer learning.The method use labeled data from other source subject and unlabeled data from the target subject,and only needs a small amount of labeled data from the target subject.The method firstly finds the most similar source subject to the target subject with some pretrain model.Then finetune the pretrain of the source subject with the little label of the target subject;at last,improve the performance with unlabeled data by two losses: Jenson–Shannon divergence loss and centering loss,which align the distribution of label data and unlabeled data.The experiment shows that the model’s performance gets close to those trained by extensive label data.