Method for qualitative and quantitative determination of key substances in mixture based on terahertz spectrum

Patent metadata

Field	Value
Publication number	US-11353396-B2
Application number	US-201916243315-A
Country	US
Kind code	B2
Filing date	Jan 9, 2019
Priority date	Jan 9, 2019
Publication date	Jun 7, 2022
Grant date	Jun 7, 2022

Abstract

Disclosed is a method for qualitative and quantitative determination of key substances in mixture based on terahertz spectrum. Terahertz spectrum of a reference mixture is trained through an SVR algorithm, and predicting parameters of key substances in the mixture to be determined after a model is generated. According to the method, an initial pure spectrum corresponding to each composition in the mixture does not need to be separately determined, no limitation on the number of samples contained in the mixture, and no limitation on frequency range to be determined, and the proportion requirement of the mixture in the early test stage is not limited, and the SVR model does not need to be re-trained after a database is formed in the later stage, and the result can be obtained immediately after the spectrum of item to be determined is introduced into the algorithm model.

First claim

What is claimed is:

1. A method for qualitative and quantitative determination of key substances in mixture based on terahertz spectrum, comprising:
1) selecting n reference mixture samples containing key substances in a plurality of mixing proportions, and recording an actual concentration of the key substances; scanning the n reference mixture samples for multiple times by using a terahertz time domain spectroscopy system and obtaining graphic spectra of time domain signals of the reference mixture samples;
2) intercepting a reflection peak for each of the graphic spectra of the time domain signals obtained in step 1), and performing a Fourier transform to convert the intercepted reflection peak into absorption coefficient frequency spectrum; intercepting a graphic spectrum between effective regions after a wavelet transform, obtaining corresponding frequency spectra of each time domain signal;
3) grouping data of the frequency spectra obtained in step 2), one group is training data, and an other group is a test data; using frequency spectra of the training data as feature vectors, and establishing a comparison database wherein the feature vectors are in a one-to-one correspondence with known parameters of corresponding reference mixture sample; performing a support vector regression (SVR) model training on each composition of the key substances to be determined, finding out a characteristic relation between various parameters of the key substances of said kind of reference mixture sample and the terahertz spectrum, and obtaining a trained SVR model;
4) predicting frequency domain spectra of the test data by using the trained SVR model obtained in step 3), obtaining various parameters of the key substances in the reference mixture sample corresponding to the calculated test data;
5) comparing the calculated various parameters of the key substances in the reference mixture sample corresponding to the test data obtained in step 4) with the known parameters corresponding to the test data, verifying an accuracy rate of the trained SVR model obtained in step 3), if the accuracy rate meets requirements, using the SVR model to perform a parameters prediction of the key substances selected in step 1);
if the accuracy rate does not meet requirements, increasing the number of reference mixture samples, after performing step 1) and step 2) process, re-entering step 3) to perform training, and obtaining an adjusted SVR model, then performing verification: wherein comparing the calculated various parameters of the key substances in the reference mixture sample corresponding to the test data obtained in step 4) with known parameters corresponding to the test data, verifying accuracy rate of the adjusted SVR model obtained in step 3) until the accuracy rate meets requirements, performing the parameters prediction of the key substances selected in step 1) by using the adjusted SVR model;
wherein the parameters prediction is to input the frequency domain spectra of the training data and the actual concentration of each key substance into the SVR model, and establish a correlation between spectral characteristic of frequency domain and the actual concentration of the key substances, and then input the frequency domain spectra of the test data into the SVR model, and the SVR model provides qualitative and quantitative information of the key substance.

Assignees

• Univ Of Shanghai For Science And Technology

Inventors

• Peng Yan
• Zhu Yiming
• Shi Chenjun
• Yang Yujian
• Zhao Jiayu
• Liu Zhijia
• Tang Xinyu
• Liu Yang
• Guo Lin
• Liu Keying
• Zhuang Songlin

Classifications

• G06N5/04

  Inference or reasoning models · CPC title

• G01N21/3586 Primary

  by Terahertz time domain spectroscopy [THz-TDS] · CPC title

• G06N20/10

  using kernel methods, e.g. support vector machines [SVM] · CPC title