Research On Fast Detection Algorithm And Application Of Infrared Spectroscopy Of Panax Notoginseng Powder Quality

To achieve the rapid detection of the quality of traditional Chinese medicine and improve the detection ability of inferior and counterfeit products in the mixture of traditional Chinese medicine powder have great research value in the aspects of drug safety spot checks and monitoring,as well as safeguarding the lives and health and legitimate rights of consumers.Panax notoginseng is a precious resource of traditional Chinese medicine,so it is of great significance to realize the rapid and non-destructive testing of Panax notoginseng powde.Due to its high medicinal value,difficulty in planting,and high prices,a large number of fake and shoddy products of Panax notoginseng appeared on the market,which will harm consumers' interests and endanger consumers' health.In this paper,the infrared spectroscopy is used to measure the molecular spectrum information of Panax notoginseng powder and its common inferior and counterfeit products.Machine learning methods are used to establish rapid detection models for the quality of Panax notoginseng powder.A co-occurrence feature learning method is proposed to enhance the generalization ability of the model.An automatic analysis platform for spectral data is developed to speed up the efficiency of data analysis.The main research work and results of the paper are as follows:(1)Achieve rapid identification of Panax notoginseng powder mixture of different grades,and combine data fusion and optimization algorithms to improve the identification accuracy of the mixture.20 Tou and 140 Tou Panax notoginseng are used as the research objects.Panax notoginseng powder of different grades are mixed according to 12 kinds of blend ratios(the minimum blend ratio is 5%).The Fourier transform mid-infrared spectrum(FT-MIR)of the sample is measured.Machine learning model is used to achieve rapid identification of Panax notoginseng powder adulteration of different grades.First,the feature variables are selected using the Interval Partial Least Squares(i PLS)algorithm and Principal Component Analysis(PCA)in order.Then,according to the change of the calibration accuracy of the Linear Discriminant Analysis(LDA)and Support vector machine(SVM)model with the number of principal components,it is determined that 11 principal components are used for modeling.Finally,the calibration accuracy and test accuracy of PCA-LDA are 99.72% and 100%,and the calibration accuracy and test accuracy of PCA-SVM are 98.61% and 100%.In order to further improve the detection accuracy,FT-MIR and Near infrared(NIR)spectral data are combined,Support Vector Machine(SVM)model is optimized based on Particle Swarm Optimization(PSO)algorithm.The experiment is designed with 14 kinds(L14,minimum blend ratio of 1%)and 15 kinds(L15,minimum blend ratio of 0.5%)for comparison experiments.When using 9 principal components,the prediction accuracy of SVM on L14 and L15 is 92.46% and 91.97%,the prediction accuracy of PSO-SVM on L14 and L15 is 96.65% and 96.97%.The experimental results show that FT-MIR spectroscopy combined with machine learning can effectively identify Panax notoginseng powder adulteration of two different grades,moreover,data fusion and optimization algorithms in combination can improve the model's discrimination ability.There are many grades of Panax notoginseng according to the number of Tou.This study provides a research basis for the identification of other different grades of Panax notoginseng powder adulteration.(2)Realize Multi-label identification for the powder and origin of Panax notoginseng infected with root-knot nematode based on Attenuated Total ReflectionFourier Transform Infrared(ATR-FTIR)spectroscopy,and accomplish the fast and nondestructive identification of the adulteration of healthy and unhealthy powder.Healthy and unhealthy Panax notoginseng from three different geographical origins are collected.The ATR-FTIR spectra of Panax notoginseng powder are measured.Multiplicative Scatter Correctioin(MSC)is as preprocessing method.Competitive Adaptive Reweighted Sampling(CARS)and Successive Projection Algorithm(SPA)select 17 feature variables.Density-based Spatial clustering of Application with Noise(DBSCAN)is used to observe subsets of the samples.Each subset is a cluster,and a total of 6 clusters are obtained.Binary Relevance Method(BR),Classifier Chain(CC),Ensembles of Classifier Chains(ECC),and Multilayer Perceptron Classifier(MLPC)are applied to create multi-label classification models.The precision,recall,F-value,and accuracy of the test set are used as evaluation indicators.The experimental results show that using the ECC to consider the effect of label order can significantly improve the performance of multi-label classification.The Panax notoginseng powder infected with root-knot nematode is mixed into healthy Panax notoginseng powder at various blend ratio.And then,The ATR-FTIR spectra of the mixure are measured.First deirative with 7-point Savitzky-Golay smoothing is as preprocessing method.i PLS and CARS are applied to select feature variables.Finally,Back-propagation Neural Network(BPNN)and nu-SVM are used to establish fast identification models of the mixure of healthy and unhealthy Panax notoginseng powder.The performance of the nu-SVM model is better,the calibration accuracy and prediction accuracy are 91.97% and 98.67%,respectively.(3)Co-occurrence feature learning method is proposed to improve the generalization ability of the model.Qualitative and quantitative identification studies of saponins are performed on different parts of Panax notoginseng root from different geographical origins.The ATR-FTIR spectra of the main root,branch root,and rhizome of Panax notoginseng are measured.SVM,BPNN,Long Short-term Memory(LSTM),and Convolutional Neural Networks(CNNs)combined with co-occurrence feature learning methods are applied to identify the powder of the different parts of Panax notoginseng.The samples of Panax notoginseng are collected from 21 different geographical origins.3065 infrared spectra are measured.7 different preprocessing methods are used to preprocess the original spectral data.1662.366-941.5894 cm-1 is manually selected as the characteristic band.Extremely randomized trees(Extra-trees)algorithm combined with the Gini importance is applied to calculate the variables' importance of the preprocessed spectral data and the original spectral data,and variables are ranked when getting the results.Then,19 co-occurrence feature variables are selected from ordered variables.The modeling results of the 19 co-occurrence feature variables selected by co-occurrence feature learning method are compared with the modeling results of the first 19 variables selected by traditional methods.The experimental results show that SVM and BPNN have better generalization ability on independent test sets based on co-occurrence feature learning method.The classification results obtained by LSTM and CNNs based on cooccurrence feature variables is the best.The accuracy of the test set is 96.00% and 95.20%.The results show that the co-occurrence feature learning method proposed in this paper can effectively improve model classification accuracy and generalization ability.In addition,High Performance Liquid Chromatography(HPLC)is adopted to measure the content of ginsenoside Rg1,Rb1,and notoginsenoside R1 in the main root,branch root,rhizome of Panax notoginseng from 20 different geographical origins.Combined to ATR-FTIR spectra,Partial Least Squares(PLS)Regression model is used to achieve rapid prediction of saponin content of the different parts of Panax notoginseng.(4)Research and develop an automatic analysis platform for spectral data to realize automatic screening of the best preprocessing methods and a variety of commonly used machine learning models to maximize the efficiency of spectral data analysis.In addition,based on the automatic analysis platform and the co-occurrence feature learning method,we build a rapid identification model for the mixture of Panax notoginseng powder and 7 counterfeit products.Based on the spectral data derivation,Savitzky-Golay smoothing,MSC,and SNV algorithm,the automatic analysis platform has designed 20 commonly used preprocessing methods with different combinations.PLS is used to establish the evaluation model of each preprocessing method.And then,we design the evaluation rules of the model results to automatically select the best preprocessing method.In order to enhance the usability of the computing platform,the client of the automatic analysis platform based on J2 EE is developed to upload the data file and display the modeling results.The Python language implements the computing model of server-side.Thrift as a remote procedure call framework is used to implement cross-language parameter transfer between the client and server.Based on this platform,the rapid identification of the mixture of Panax notoginseng powder and 7 counterfeit products is achieved.First,Panax notoginseng powder blend with 7 counterfeit products in different proportions,then 3550 ATR-FTIR spectra are obtained.Next,automatic analysis platform is applied to select the best preprocessing method,co-occurrence feature learning method is used to select the feature variables.Finally,based on the computing platform,LDA,BPNN and SVM models are quickly established.The experimental results show that the modeling effect of SVM is the best,and the accuracy of cross-validation and prediction is 97.33% and 97.41%.The automatic analysis platform can automatically screen the best preprocessing method and analyze the data based on the machine learning model in the platform,which can greatly improve the analysis and modeling efficiency of spectral data.