Analysis Of Motor Imagery Eeg Signals Based On Deep Neural Network

The brain computer interface(BCI)system is a new type of human-computer interaction mode,which is directly established between the brain and external devices.It doesn't rely on the peripheral nervous system or muscle tissue and enables direct information exchange between the brain and the outside world.The BCI system has a broad application prospect,its key technology is to classify electroencephalogram(EEG)signals and convert the classification results into commands for controlling external devices.Nowadays,deep neural networks(DNN)have been widely used in many fields.It also provides a very effective analyzing tool for the research of EEG signals.This paper focuses on the analysis of EEG signals in the motor-imagery BCI(MI-BCI),and the purpose is to improve the accuracy and robustness of motor-imagery EEG(MI-EEG)signal recognition by using DNN algorithm.Based on the work of predecessors,The mainly completed work of this paper is as follows:1.First of all,the characteristics and categories of EEG signals as well as the composition and classification of BCI systems are introduced.The domestic and international research status of MI-BCI and related analysis algorithms is summarized.The related theoretical basis of DNN algorithm is explained.Based on this,the feasibility of processing EEG signals with DNN algorithm is analyzed.2.Aiming at the low signal-to-noise ratio and strong randomness of the EEG signals as well as the shortcomings of the traditional MI-EEG signal analysis methods.A novel end-to-end analysis model which combining temporal-spatial convolutional neural network(TSCNN)and stacked autoencoder(SAE)is proposed.What's more,this paper adopts a training mode of "global-individual classifier" to help improve the classification performance of TSCNN-SAE model.The model validity was verified by using three BCI competition datasets.The experimental results indicate that TSCNN-SAE modelshows universality among different subjects and achieves an average classification accuracy of 82.7% and 90.6% in two 2-class tasks,which are higher than the TSCNN and SAE models trained separately,the championship algorithm of corresponding BCI competition and other representative algorithms.In addition,the model achieved an average recognition rate of 83.8% in the4-class tasks,which is close to the championship algorithm of the corresponding BCI competition,but there is still a large gap compared with the excellent algorithms in recent years.3.In the case of a small sample size,the DNN model with high parameter complexity in 4-class MI-EEG signal recognition tasks cannot be fully trained,which results in poor classification performance.Regarding the issue above,this paper proposes a 4-class MI-EEG signal analysis model which is based on one-versus-the-rest common spatial pattern(OVR-CSP)algorithm and TSCNN.In addition,this model adopts a data augmentation scheme and tries to introduce majority voting algorithm for comparative experiments.The experimental results show that CSP-TSCNN model simplifies the structural parameters of the convolutional network while expanding the training sample size,and the model performs better before introducing the majority voting algorithm.This model finally achieves an average recognition rate of 90.1%,which is a significant improvement over the TSCNN-SAE model and better than the typical algorithms in recent years.