Research On Rapid Identification Of Ginseng Origin Based On Laser-induced Breakdown Spectroscopy Technology

As a representative of the characteristic cultural products of the three northeastern provinces,ginseng contains ginsenosides and trace elements that help prevent fatigue,resist oxidative damage and treat cancer,and play an active role in the treatment of cardiovascular disease,immunity and the central nervous system.The types of ginsenosides in ginseng from different producing areas are basically the same,but the contents of trace elements are different,and the medicinal value of ginseng from different origins is also different.In recent years,laser-induced breakdown spectroscopy(LIBS)has been developing rapidly beyond imagination.It has many advantages that other spectral analysis methods can not compare with,such as no complex sample preparation,almost no additional reagents,and on-line multi-element detection.Due to its advantages,LIBS is widely used in various fields,but LIBS also has its own limitations,such as large amount of full-spectrum data,high detection limit,and low quantitative and qualitative accuracy.In view of the limitations of the LIBS method,a machine learning algorithm is introduced to combine with it to solve the problem of large and complex LIBS data and to improve the qualitative accuracy.Therefore,this paper aims to quickly and accurately identify the origin of ginseng through LIBS combined with machine learning algorithms.The work completed in this paper is as follows:1.To prepare experimental samples and build the LIBS experimental system,firstly,the ginseng samples were processed into ginseng samples with smooth surface and uniform specifications,and then the optical path required for the experiment was built.By observing the collected spectrum,the experimental instrument was adjusted for optimization,and the experimental system under the optimal conditions and parameters was obtained.2.Perform noise reduction processing on the spectrum collected in the experiment.Since the original noise of the light source and the stray light in the optical path in the experiment will generate a certain degree of noise,the preprocessing process of the spectrum should be performed before the analysis.Four methods(moving window average smoothing method,S-G smoothing filtering method,Fourier fast filtering method and wavelet transform method)were selected to de-noise the spectrum respectively,and the signal-to-noise ratio(SNR)was used as an indicator to compare the de-noising effects of the four methods,and the optimal spectral pre-processing method was obtained.3.Extract the representative spectral features and compare with NIST to obtain the characteristic spectral lines information of the ginseng spectrum.Two feature vector selection methods(principal component analysis and random forest)were used to reduce the dimension of the experimental data,and the selected feature information was obtained.The random forest feature vector selection method was used as the feature extraction method after comparison,and the selected feature vector was used as the input variable of the subsequent model.4.Classification and identification of ginseng from five different origins,the selected feature vector was used as a representative of ginseng classification data,and was substituted into random forest-support vector machine(RF-SVM)and random forest-back-propagation neural network(RF-BPNN).The classification accuracy of the two models was 99.75% and 100%,respectively.Rf-SVM model and RF-BPNN model were compared by using the performance evaluation parameters of the model.The results showed that RF-BPNN model could achieve 100% classification accuracy for ginseng samples from five habitats,and the classification speed was faster and the classification effect was the best.The optimal classification model-RF-BPNN was obtained.5.To verify the robustness of the model,the ginseng classification model was validated according to the data collected with the introduction of two variables(time and impurities)respectively.Under the condition of error tolerance,the classification accuracy of the optimal classification model(RF-BPNN)was still 100%,and the optimal model was analyzed to be robust.