Identification Of Typical Chinese Medicinal Materials Based On Terahertz Spectra

Species identification is an important link in the quality control of Chinese herbal medicines,and the authenticity of the species is directly related to the clinical efficacy of patients and even the safety of life and property.The traditional methods of herbal medicine identification have disadvantages such as strong professionalism and lack of objectivity,while the modern identification techniques of infrared spectroscopy have problems such as serious overlapping of spectral peaks and Raman spectroscopy has the defect of fluorescence interference.In recent years,people have started to combine machine learning with terahertz time-domain spectroscopy to achieve the identification of Chinese herbal medicines.However,there is no standard on how to preprocess the terahertz spectra of herbal medicines;in addition,there is no definite conclusion on what feature extraction method should be used to extract spectral features and reduce redundant information so as to improve the model accuracy;in addition,most of the current research on recognition based on terahertz spectra of herbal medicines adopts closed-set recognition model(i.e.,the training set and test set have the same category),but such recognition model will produce a large number of false identifications.Based on the above problems,this paper takes six typical herbal medicines as the research object and combines machine learning with their terahertz spectra to achieve the accurate identification of different varieties of herbal medicines,and the main work is as follows:(1)Acquisition of terahertz absorption spectra of six typical Chinese herbs.The timedomain spectra of different varieties of herbs(White monkshood tablet,Rheum officinale,Dried tangerine or orange peel,Codonopsis pilosula,Ophiopogon japonicus and Rhizoma gastrodiae)were obtained by THz-TDS system,and their absorption spectra were calculated.The results showed that the spectra of the six herbs were different from each other,but machine learning is still needed to achieve accurate identification of the herbs.(2)In order to make the classification model have better prediction accuracy,the raw terahertz spectra of the herbs need to be preprocessed.In this paper,Min-max normalization,S-G smoothing,and MSC are used to preprocess them respectively,and the results of different treatments are compared.The results show that Min-max normalization + MSC is the best pretreatment method.(3)Establishing a typical Chinese herbal medicine species identification model.Support vector machine(SVM),random forest(RF)and partial least squares discriminant analysis(PLS-DA)algorithms were used to establish three herbal species classification models,and the results showed that: SVM had the best classification effect.Subsequently,principal component analysis(PCA),linear discriminant analysis(LDA),and LASSO feature selection were used to extract features from the spectra,and the extraction results were input into the support vector machine classification model.The results showed that the LDA-SVM variety recognition model had the highest accuracy(99.12%).Finally,in order to make the model have rejection recognition ability for out-of-set herbs,fuzzy discrimination(FDA)was introduced and LDA-SVM-FDA open-set recognition model was constructed,which used three herb spectra as training set and six herb spectra as open test set,and the results showed that: the recognition rate of the recognition model was 95.45% for herbal samples in the open test set,and the method can effectively reduce the unknown class samples of false recognition.