Research On Classification Of Motor Imagery EEG Signals By Combining Independent Component Analysis And Ensemble Learning

Brain-computer interface(BCI)technology establishes a direct communication path between the brain and external devices by decoding the electrical signals generated by the brain,which is a new type of human-computer interaction mode.BCI based on scalp electroencephalogram(EEG)is widely concerned by researchers because of its good safety and simple operation,but the signal-to-noise ratio and spatial resolution of the scalp EEG signals are not high.Independent component analysis(ICA)can improve the signal-to-noise ratio and spatial resolution of EEG signals.It is often used in the preprocessing and feature extraction stage of BCI system,and has great application potential in EEG signals analysis.In ICA-based motor imagery BCI system(ICA-MIBCI),the performance of ICA spatial filter directly determines the classification accuracy of ICA-MIBCI.However,the design and optimization process of ICA spatial filter is complex and time consuming,and the existing methods have poor practicability.In order to obtain an ICA-MIBCI system with good classification results in less time,in this thesis,ICA and ensemble learning methods are combined for the construction of classifier in ICA-MIBCI.At the same time,the performance differences of ensemble classifiers obtained by combining different spatial filters and ensemble learning methods are studied,and the online MIBCI system is developed based on the above research.Specific work of this thesis are as follows:(1)The generation mechanism and characteristics of EEG signals are studied,and the feature extraction methods commonly used in time domain,frequency domain and spatial domain,as well as the classical classification algorithm are introduced,and the event-related desynchronization/synchronization phenomenon in motor imagery EEG(MIEEG)signals is explained.(2)Two classification methods of MIEEG signals by combining ICA and Bagging,as well as ICA and Adaboost are proposed.First,a piece of EEG data is randomly selected from the original training set for ICA analysis,and the effectiveness of the obtained ICA spatial filter is judged.The effective ICA spatial filter is selected and combined with the simple classification criterion of variance comparison for the construction of base classifier(ICAVCr).Then,the ensemble learning methods Bagging and Adaboost are employed to combine multiple ICAVCr separately,and two ensemble classifiers are obtained.The experimental results show that compared with the single classifier constructed by commonly used ICA spatial filter optimization methods,the ensemble classifier have better classification performance and lower time cost in training phase.(3)The classification methods of MIEEG signals by combining ICA and ensemble learning,common spatial pattern and ensemble learning are compared and analyzed in session-tosession transfer and subject-to-subject transfer.The results show that the ICA spatial filter is more suitable for the construction of the base classifier in ensemble learning method.Especially in subject-to-subject transfer,the ensemble classifier obtained by combining ICA and Bagging(ICA?Bg)shows good robustness,which is suitable for building a practical MIBCI system.(4)On the basis of the above research,an online MIBCI system based on ICA?Bg algorithm using C++ language and Microsoft Foundation Classes is developed.In order to facilitate users to observe the results of the feedback,a simulation control system is developed using C#language and Unity3 D engine.The entire system uses TCP/IP protocol for data transmission,which has the characteristics of simple operation and easy to transplant.Two subjects performed a total of six online experiments,and all successfully controlled the virtual characters to the target point,which demonstrates the effectiveness of the online MIBCI system.