The Research On Artifacts Removal And Classification Method Of Three Kinds Of Motion Execution FNIRS Signals

The brain computer interface(BCI)technology converts the signals collected from the brain into instructions through certain steps for controlling external devices,and then directly controls the external devices without passing through the peripheral nervous system and muscle tissue.The research of BCI technology is of great significance to people who have lost the ability to exercise due to diseases or accidents,and can provide such people with tools for communication and interaction with the outside world.In addition,BCI technology is also widely used in military,entertainment,transportation and other fields.Due to the advantages of functional near-infrared spectroscopy(f NIRS)technology such as portability,non-invasion,and low price,it is gradually being valued in the BCI field.However,the existence of motion artifacts in the f NIRS signal will cause the signal processing results to deviate from the real results,so the research of motion artifact removal algorithms is of great significance.The thesis focuses on the removal of motion artifacts in f NIRS signals and the classification of f NIRS signals for three types of motions,and the following work is mainly done:(1)Aiming at the problem that a single method cannot effectively remove the three common motion artifacts in the f NIRS signal at the same time,the tMedMor algorithm which combined mathematical morphology(Mor)with improved targeted median filtering(t Med)is proposed to remove three kinds of motion artifacts,including spikes,baseline shifts,and slow drifts.Since the Mor method introduces a baseline shift when removing two adjacent spikes,the t Med algorithm is used before the Mor algorithm to remove the spikes in advance,and then use the Mor method to remove the baseline shifts and slow drifts.Comparing with several existing algorithms on simulation data and experimental data,the tMedMor algorithm performs well on the four evaluation indicators and can be used in the preprocessing stage of the f NIRS signal.(2)After removing motion artifacts using tMedMor algorithm,the classification methods of three types of motions execution f NIRS signal(right hand finger tapping(RHT),left hand finger tapping(LHT),and foot tapping(FT))are researched.In terms of channel selection,in view of the problem that the redundant or irrelevant channels in the f NIRS signal will affect the classification performance,the significant difference based method is used for channel selection.First,the general linear model(GLM)is used to fit the oxy-hemoglobin(Hb O)and deoxy-hemoglobin(Hb R)signals,and then the obtained parameters are used to perform paired-sample t test.The channels with significant differences between different actions are selected for feature extraction and classification.During classification,Relief-F algorithm is used for feature selection,and support vector machine(SVM)is used for the classifier.Compared with the result of selecting active channels for feature extraction,the results show that the channel selection method based on significant differences can achieve higher accuracy when selecting fewer features.After channel selection,the number of channels is reduced by 45%,which greatly reduces the feature dimension,avoiding over-fitting problems.In addition,the removal of irrelevant channels can reduce the development cost of BCI application equipment.(3)In terms of feature type selection and combination during classification,aiming at the problem of redundancy when there are too many types of features,seven types of univariate features and one type of multivariate features are extracted and the features are selected from the two dimensions of feature importance and relevance.The features with high accuracy and low relevance are combined as the input of the classifier.The Relief-F algorithm is used for feature selection.The classifier uses SVM.Comparing the classification results with the results of non-feature type selection,the results show that,the classification performance after feature type selection is better than the result before feature type selection.It shows that using the feature type selection method before the Relief-F algorithm,the accuracy of the classifier can be kept stable while reducing the feature dimension.