Qualitative And Quantitative Analysis Of Chinese Medicinal Materials Based On Machine Learning And LIBS Technology

There are many Chinese medicinal materials,some of which the morphology is so similar that to distinguish them is difficult.In recent years,the problem of heavy metal pollution in soil has led to the increasingly serious phenomenon of excessive heavy metal content in Chinese medicinal material,which is caused by industrial development and extensive use of pesticides.To quickly and reliably identify the species of Chinese medicinal material through a non-visual means,and realize in-situ online detection of heavy metal elements in Chinese medicinal material,it is significant and practical to carry out qualitative identification and quantitative analysis of Chinese medicinal material based on rapid element analysis method.Laser-induced breakdown spectroscopy(LIBS)is a new element analysis technology based on atomic emission spectrometry,which possesses some characteristics including simple sample pretreatment,rapid real-time detection of multiple elements and so on.Machine learning algorithm can finish predicting the information composition of unknown samples through the characteristics of self-iterative learning samples.Besides,it can promote the application process of LIBS technology if used in the research of LIBS spectral data analysis,recognition and prediction.The combination of machine learning algorithm and LIBS technology for qualitative and quantitative analysis of Chinese medicinal material can reduce the difficulty of analysis and improve the accuracy and precision of analysis,which can provide more advanced and effective technical means for improving the quality of Chinese medicinal material.Based on the above research background and objectives,the main work is as follows.(1)A method of Chinese medicinal material classification and recognition based on LIBS technology and PCA-PSO-SVM was proposed and verified by experiments.Firstly,the spectra of licorice,astragalus,angelica sinensis and codonopsis pilosula were collected and pretreated.During the process,the wavelet threshold denoising method was used to smooth the spectrum,and the denoising results of different wavelet basis functions and decomposition levels were compared and analyzed.The results showed that the denoising effect of using the db5 wavelet base to split the spectrum with three layers is the best.The segmented eigenvalue extraction method was used to correct the baseline of the denoised spectrum,and then the spectrum was normalized.Secondly,the Chinese medicinal material classification model based on PCA-PSO-SVM was constructed.The classification results of different principal components as the input of PSO-SVM model were compared,and then the first nine principal components after dimension reduction were selected as the input of PSO-SVM model.The result obtained was that the best classification accuracy of four medicinal materials was 100%.PCA-LDA and PCA-PSO-BP classification models were constructed respectively,and comparative experiments were carried out.The results showed that PCA-PSO-SVM model had the best classification performance for four kinds of medicinal materials,and its average classification accuracy reached 100%.This work provides an effective technical strategy for the rapid and high accuracy identification of Chinese medicinal materials by laser-induced breakdown spectroscopy.(2)A quantitative analysis method of Cr in Chinese medicinal material based on LIBS technology and machine learning was proposed and validated.Firstly,the standard samples of licorice,astragalus,angelica sinensis and codonopsis pilosula with different concentrations of Cr were prepared,and then their spectra were collected and pretreated.The single variable standard curve models of Cr concentration in four kinds of Chinese medicinal materials were established.The standard curve model was used to predict the Cr element concentration in the four medicinal materials.The obtained results were that the average relative error(REC)of the calibration set was between 21.23% and 30.52%,the average relative error(REP)of the validation set was between 25.68% and 40.67%,the relative standard deviation(RSD)was between 38.54%and 42.86%,and the limit of detection(LOD)was between 73.53 ppm and 124.61 ppm.Secondly,the multivariable quantitative analysis models of Cr concentration in four kinds of Chinese medicinal material were constructed based on machine learning algorithm.To realize the automatic selection of spectral features,the Select KBest algorithm was used to replace the traditional method of artificially selecting characteristic spectral lines.And the selected spectral features were used to train and test the corresponding BP neural network model and support vector machine model of four kinds of medicinal materials.The obtained result were that REC,REP,RSD and LOD of the two models were between 1.93% and 7.45%,between 2.57% and 7.43%,between 3.06%and 10.51%,between14.42 ppm and 67.49 ppm,respectively.The results showed that compared with the traditional standard curve method,the machine learning quantitative analysis method combining Select KBest algorithm for spectral feature selection and BP neural network model or support vector machine model could effectively improve the accuracy and stability of quantitative analysis of Cr element concentration in Chinese medicinal material,and reduce the detection limit of Cr element.