The Research Of EOG Signal Enhancement Algorithm Based On Independent Component Analysis

With the rapid development of wearable devices and Internet technologies,the detection and application of bioelectrical signals of human beings play a very important role in people’s daily life.Compared with other bioelectrical signals,Electro-oculogram(EOG)has the advantages of low cost,high accuracy,light weight,long-time recording and convenient wearing.Meanwhile,EOG signals can reflect the patterns of eye movement under different activities.Therefore,EOG-based Human Activity Recognition(HAR)has become a new research hotspot.In the real scenarios,however,the acquisition of EOG signals will inevitably be interfered by the ambient noise.As a result,the quality of EOG signals will decrease,which leads to a challenge of recognizing them.In order to reduce the impact of noise signals on the EOG-HAR system and improve the recognition accuracy,this thesis presents a study of EOG signal enhancement algorithm based on independent component analysis.The details are depicted as follows:(1)Introduced an EOG signal enhancement algorithm in the instantaneous mixed model.Firstly,we described several common ICA algorithms about the instantaneous mixed model.On this basis,we propsed an EOG signal enhancement algorithm based on instantaneous ICA aming to impriove the quality of EOG signal under the "reading"state.In label environment,the average recognition ratio of the proposed algorithm reaches 95.5%,which obtains the relative increasement of 3.39%,5.0%and 2.7%compared with raw signals,band-pass filter and PCA.Experimental results verify the effectiveness of the proposed enhancement algorithm.(2)Proposed an enhancement algorithm of EOG signal based on convolutive mixed method.According to the theorem that observation data model described by convolutional mixture in real environment has universal applicability,we also proposed an EOG signal enhancement algorithm based on the convolutive ICA model in this section.To start with,raw time-domain EOG signals are converted into frequency-domain using the Short Time Fourier Transform(STFT).Then a complex Independent Component Analysis(ICA)algorithm was applied to separate the saccade sources and noises in the frequency-domain.Finally,the Power Spectral Density(PSD)of separated saccade sources was computed as feature parameter and fed into the classifier.The average recognition ratios of enhanced EOG signals achieve 95.60%and 97.30%under between-subjects test and within-subjects test.Compared with bandpass filtering,wavelet denoising,extended Infomax algorithm,real JADE algorithm,and classic JADE algorithm,the recognition ratios obtain the relative increment of 4.45%,3.44%,2.78%,2.76%,and 2.80%(within-subjects test)and 4.88%,3.32%,2.12%,2.13%,and 2.16%(between-subjects test).The experimental results reveal that the proposed algorithm presents a robust classification performance in saccadic EOG signals recognition.(3)Proposed a new frequency-domain constraint DOA algorithm to solve the permutation ambiguity in the convolutive ICA model.As for the ICA model,the permutation ambiguity is one of the inherent indeterminacies in the BSS problem.Especially for the convolutive model,the ICA algorithm will be independently performed in each frequency bin,the order of recovered signals in each frequency bin must be aligned so that the reconstructed signals in the time-domain will not be mixed with other sources.Therefore,the solution of solving permutation ambiguity becomes significantly important for the convolutional ICA model.To solve this problem,this thesis developed a constraint direction of arrival algorithm to solve the problem of permutation ambiguity for 6-channel convolutional model.It first empirically initializes angle of different source signals as a constraint after blind separation,then compares the output angle of the DOA with the initialized angle in each frequency bin.On this basis,we adjusted the channel order so as to make the new output in accordance with different sources respectively.The experimental results show that the average recognition ratio of between-subjects test and within-subjects test increased by 95.25%and 97.30%respectively after being sorted by the permutation algorithm proposed in this thesis.Experimental results reveal that performance obtains the relative increasement of 3.03%(between-subjects test)and 4.26%(within-subjects test)compared with the noisy EOG signal before permutation,which verify the effectiveness of the proposed permutation algorithm.