A Study Of Blind Source Deconvolution Methods Of Multichannel EOG Signals

The eyes are one of the most important sense organs of human beings,and humans could interact with outside world or convey information through eye movements.Furthermore,the abnormalities of eye movements are the common clinical symptoms of some important diseases.The electrooculography(EOG)technique is widely employed on the EOG-based human-computer interface(HCI)system and fatigue driving for its non-invasion,low cost and so forth,compared with other recording techniques.The accurate classification of eye movements is one of the key points in improving the performance of the EOG-based HCI system.In addition,the saccadic eye movement works as the main research object in this study,since it has the high frequency of occurrence.To obtain high quality saccadic EOG signals,this thesis focuses on the blind source separation(BSS)problem of multichannel EOG signals and the major research object is the convolutional BSS problem.The detailed information of this thesis is as follows:(1)Proposing appropriate evaluation methods for the separation results of multichannel EOG signals.Some common evaluation indicators are not suitable for the EOG signals,since the EOG signals are always recorded through an indirect way and we can get no source signals as a reference on this condition.On this basis,two kinds of indicators are employed for the evaluation,in which one is the qualitative analysis,the other is the quantitative analysis method.The qualitative analysis compares the time-domain waveforms and frequency-domain spectrograms corresponding to the mixed and separated signals;the quantitative analysis contains three kinds of statistical indicators,which are the second-and fourth-order statistical dependencies,total square cross correlation and saccadic classification accuracy,respectively.The indicators mentioned above could be used to evaluate the degree of separation corresponding to different BSS methods.(2)This thesis studies the independent vector analysis(IVA)algorithm and its applications in the BSS problem of multichannel EOG signals.The EOG signals need to transfer through the skull,subcutaneous tissues and so on,which may exist a convolution.Therefore,the simple instantaneous mixture model may be hard to describe the generation process of multichannel EOG signals.In order to achieve better separation performance,this paper makes a comparative experiment between two different generation models of multichannel EOG signals.We further use the IVA algorithm as the BSS method corresponding to the convolutional mixture model,in order to solve the permutation ambiguity inherent in the ICA algorithm.The IVA algorithm could theoretically reduce the impact of permutation ambiguity inherent in ICA.For reducing the impact on the separation performance which are caused by the individual difference,this paper also makes a parameter selection of frame size.The experiment results show that the outputs of convolutional mixture corresponding to second-and forth-order statistical dependencies,and total square cross correlation are 0.04(second-order)/ 0.08(forth-order),and 1.25,respectively.They are lower than instantaneous mixture by 0.3(second-order)/ 0.17(fourth-order)and,0.32,respectively.In addition,in terms of the results of the classification accuracies,the results of convolutional mixture are 95.76 %(within-subject)and 94.08 %(between-subject),which were higher than those of the instantaneous mixture model by 1.69 % and 5.53 %,respectively.It indicates that the IVA could make the separation of multichannel mixed EOG signals,and the convolutional mixture model could achieve a better performance in solving the BSS problem compared with the instantaneous mixture model.(3)This thesis studies the improved BSS method based on IVA algorithm for multichannel EOG signals.In order to improve the separation performance of IVA algorithm for EOG signals,the non-negative matrix factorization(NMF)algorithm has been employed to decompose the spectrogram of multichannel EOG signals.With the help of NMF algorithm,the limitation of the IVA algorithm could be optimized and the separation performance of IVA algorithm for BSS problems of multichannel EOG signals could be improved.The selection of several parameters is also made in the experiment part to achieve a comparatively better separation performance.From the result of classification accuracies,the independent low rank matrix analysis(ILRMA)algorithm reaches 98.1 %,while the corresponding result of IVA algorithm only reaches 95.75 %.This result indicates that the employment of the NMF algorithm could improve the BSS separation performance of the IVA algorithm,in which the NMF algorithm is used to update the variance parameter related to the spectrogram in the iteration step.