| Motor imagery based Brain-Computer Interface(BCI)can establish the communication between human brain and external equipment,without assistance of peripheral nerve and muscle tissue.It can help patients with movement disorders,better communicate with the outside world,and is of great value in military,aerospace,medical,and virtual reality fields.However,due to the non-stationary nature of EEG signals,the traditional motor imagery technology needs to be labeled with a large number of training samples and adopt multi-channel acquisition methods,which greatly limits the application scope of Brain-Computer Interface technology in motor imagery.In this work,to reduce the difference in the distribution of training samples and testing samples and improve the classification accuracy,the idea of transfer learning is applied to motor imagery based on the traditional EEG signal processing methods.In addition,in order to reduce the computational complexity,channel selection methods are systematically investigated to reduce the number of channels with an acceptable loss of accuracy for EEG-based motor imagery recognition,and improve the real-time performance of the BCI technology.This paper launches the following research work.Firstly,the preprocessing method for EEG-based motor imagery is studied based on the physiological basis of motor imagery.The AR model is used to analyze the EEG spectrum of motor imagery,and the effective band range 8-30 Hz is obtained,which provides an analysis basis for the selection of filter pass-band frequency.In addition,the advantages of the common average reference(CAR)spatial filtering in increasing the spatial distribution difference of different imaginary types of EEG signals is analyzed.It provides a basis for obtaining high signal-to-noise ratio EEG signals.Secondly,feature extraction method based on wavelet packet transform is studied.The wavelet packet of EEG is decomposed to extract the wavelet coefficients,then the energy feature is calculated.By using support vector machines(SVM)to recognize two types of motor imagery tasks,the average classification results is 74.88%.Obviously,it proves the effectiveness of the feature extraction method.Furthermore,the channel selection methods were systematically investigated based on Relief-F algorithm to reduce the number of channels with an acceptable loss of accuracy for EEG-based motor imagery recognition,which helps to reduce the amount of calculation and improve the real-time performance of the system.In addition,transfer learning algorithm based on minimum maximum mean difference(MMD)is studied to apply in the classification of motor imagery.The results show that the method can improve the classification effect of a subject motor imagery for a period of time,and can make the classification model trained by one subject more applicable to another subject.It is proved that the transfer learning algorithm has better adaptability than the traditional classification method.Comprehension from the discussions above,the motor imagery experiments based on transfer learning are designed.Aiming at the artifact of real EEG signals,the method of wavelet analysis for the removal of ocular artifacts is studied and the online implementation scheme of transfer learning is discussed.The results showed that method based on transfer learning gains 10% performance improvement compared to the traditional machine learning algorithm.It is worth mentioning that the increase is greater than experiment for same subject.In conclusion,the design of this paper not only can be used for offline experiment of tradition motor imagery,but also can provide reference for further research in the field of online test of motor imagery based on transfer learning. |
PDF Full Text Request