Research Of P300Processing Algorithm Based On Independent Component Analysis

Brain computer interface (BCI) is a new information transmission channel between human being and external devices, it need not depend on muscle tissue and peripheral nerve. As a new type of interactive way, BCI has been more and more widly used in helping disabled people, controling equipment and even entertainment.P300is an endogenous event-related potentials, which is one of the most commonly used electroencephalogram (EEG) signals in BCI applications for it’s excellent characteristics of highly stability and needless of training. Such as the P300speller, which was based on oddball experiment paradigm and proposed by Farwel and Donchin, the sampling EEG signals usually has huge amount of data and low Signal Noise Ratio (SNR), various noises and interference would be added to the EEG recordings, such as ocular artifacts、ECG and EMG therefore, the step of preprocessing and feature extraction in BCI system is particular important.Based on the existing achievement, a series of research on automatic ocular artifact removal, dynamic electrode combination selection and dynamic feature extraction algorithms has been done and verified offline in the thesis. The main work and achievements are as follows:(1) While sampling EEG signals, various noises and interference would be added to the EEG recordings, especially ocular artifact which usually has high amplitude. Aiming at the problems such as the overestimation of ocular artifact, the need of human intervention, and the difficulty for online application in traditional blind source separation (BSS) based methods for artifacts removal, a totally automatic method for removing ocular artifacts was proposed. ICA was used to separate EEG signals to obtain the independent components. With the criterion of correlation coefficient, different time windows according to vertical electrooculogram (VEOG) and horizontal electrooculogram (HEOG)’s respective characteristics were used to find the maximum correlation components combination, by which the time intervals that have blink or eye movement activities could be calibrated. Then, the calibrated time intervals were set to0and the EEG signals were reconstructed. With experiments of P300signals processing, this method was proved to be effective and practical in removing ocular artifact automatically and overcoming the drawbacks above. Compare with related literatures, the experiment results showed that the proposed method increased the average correlation coefficient between the reconstructed EEG signals and the original EEG signals from0.8513and0.9006to0.9237respectively, while the mean square error was decreased by19.3%and16.6%, contributing to online application.(2) With the aim to solve the problems in brain-computer interface, such as huge amounts of processing EEG data and ignorance of EEG’s variance from person to person in traditional EEG preprocessing algorithms, after researched the relative literatures about EEG preprocessing algorithm, a dynamic electrode combination selection method based on ICA has been proposed. ICA was used to separate EEG signals to obtain the independent components, calibrate the independent components which has P300with the criterion of peak value and temporal energy entropy between250-400ms; Then, find the electrode that has large weights of those calibrated P300component’s space projection, these electrodes formed the electrode combination we need. With experiments of BCI Competition III P300data processing, the proposed method was proved to be effective and practical in finding different combination of electrodes for different subjects and overcoming the drawbacks aforementioned. Compared with prefixed4channels, the proposed method can improve classification accuracy greatly; While comparing with prefixed8channels method, it has reduced43.75%of the total processing EEG data in BCI while remain almost equal classification accuracy, the proposed electrodes combination selection algorithm was processed offline, so it’s would not increase the online time consuming. (3) With the aim to solve the problems in brain-computer interface, such as huge amounts of processing EEG data, single feature extraction and ignorance of EEG’s variance from person to person in EEG processing algorithms, after researched the relative literatures about EEG feature extraction algorithm, a dynamic feature extraction method based on ICA and Wavelet Transform is proposed. ICA was used to separate EEG signals to obtain the independent components, considering the prior knowledge about P300, select the independent components which has P300and reconstruct EEG signals to get the feature enhanced EEG; Then use wavelet transform to decompose the feature enhanced EEG, find the electrode combination that has high classification divisibility according to the feature of approximate coefficients、detail coefficients with the criterion of fisher distance; Combined with the time domain features of the selected electrodes combination by (2), use these three electrode combination’s different feature as the feature vector and classify. With experiments of BCI Competition Ⅲ P300data processing, the proposed method was proved to be effective and practical in finding different feature for different subjects, improving the classification accuracy, and overcoming the drawbacks aforementioned, contributing to online application.