Research On Material Identification Technology Based On Terahertz

Terahertz ray is electromagnetic whose frequency ranges from 0.1 to 10 THz.Due to the lack of high power terahertz source and high sensitive detector,before 1980 s,people know little about terahertz and terahertz hasn't been effectively developed.So terahertz is called "terahertz gap".In recent twenty years,terahertz generation and detection technology has been improved.More and more scientific research institutions have begun to invest in the research of terahertz technology.The following paper mainly studies the terahertz time-domain spectroscopy technology,including the experimental principle and data processing methods.The main work and achievements of this paper are as follows:(1)According to that THz-TDS is easy interfered by noise,terahertz spectral de-noising method is studied.The traditional spectrum de-noising methods have a main shortcoming that it will lead to the decline of SNR of signal,at the same time these methods can not eliminate pulse noise.A new de-noising method based on fractional order integration is proposed.This new method uses median filter firstly to eliminate pulse noise,then uses traditional fractional order integration method,which can improve the SNR of spectrums and robustness.(2)According to baseline drift,a baseline correction method for single side insensitive SVM is proposed.This algorithm using different loss functions to plus error and minus error,which leads to the fitting curve is not the outline but the baseline of the spectrum.So this algorithm can correct the baseline of spectral at the same time keep the original sharp of the peaks.(3)According to the traditional recognition methods treat the THz spectral as a normal vector,and ignore the particularity of THz spectral,which leads to the recognition rate is not high,a feature selection method based on multi-scale analysis is proposed.This algorithm can select the real peaks of spectrums.According to the numbers of feature are not the same,traditional classifiers could not deal to this problem.A kernel function based on matching degree is proposed.When the database is too big,SVM will expend very big storage and CPU.Traditional support vector selection methods are analyzied.According to leak selection of SV,these methods have low recognition rate.A support vector selection method based on boundary points is proposed.A multi-classification method based on adjacency graph is proposed.This algorithm can improve the classification performance of SVM.