Research On Extraction And Classification Of Brain Activity Signals Based On Near Infrared Spectroscopy

Near-infrared Spectroscopy(NIRS)is an emerging brain function detection technology with balanced temporal and spatial resolution,suitable for brain science research in natural situations,and has begun to be used to explore the brain computer interface systems in recent years.In order to improve the accuracy of classification for different mental task based on NIRS signals,this paper studies the signal extraction and classification technology through the three processes of acquisition,correction and classification.Based on the multi-channel continuous-wave NIRS system,this paper proposes a dual-level light intensity excitation mode for signal acquisition,which improves the signal quality in channels with different source-detector separation.Taking advantage of the spatial differences in brain activity activation of different cognitive tasks,digital memory task,word-sentence task,and mental arithmetic task were designed to improve the distinguishability of mental tasks.Based on the processing of the NIRS signal,various sources of noise in the signal are analyzed.Aiming at the motion artifacts generated by various abnormal movements,a two-stage motion artifact recognition and correction algorithm is proposed,which uses different moving standard deviation thresholds combined with cubic spline interpolation and S-G filters to deal with different types of artifact signals.Our algorithm improved the signal quality after correction,and the classification accuracy increased by 10.56%.With the self-built NIRS data set of 4 types of brain activities,the feature extraction method based on random forests is used to build a deep forest classification model.Compared with support vector machine,decision tree,k-nearest neighbor and other classic machine learning algorithms,deep forest algorithm is more suitable to extract structured spatial and temporal characteristics of the NIRS signal.By using a combination of multi-grained scanning and cascaded forest structure,this paper finally achieved a classification accuracy of 86.95%,which is the best result of 4-class NIRS brain activity signal classifications.