Rapid Quantitative And Discriminative Analysis Of Ceramics By Laser-induced Breakdown Spectroscopy Combined With Machine Learning Strategies

As a treasure of Chinese culture,ceramics contain a wide range of information about ancient social activities and play a crucial role in the study of the evolution of ancient civilisations.In view of the sensory limitations of traditional identification methods and the precious and non-reproducible nature of ceramics,laser induced breakdown spectroscopy(LIBS)techniques need to be introduced to obtain information on the provenance of ceramics,the production processes and elemental composition of the period they represent,providing a scientific basis for conservation and restoration work.However,due to the complexity of the composition and structure of ceramics,the use of LIBS produces a large number of complex spectra,and how to obtain valid information from these complex spectral for quantitative or qualitative analysis is still one of the challenges faced.This thesis addresses the practical needs of non-destructive,fast analysis of ceramics and conducts research on the composition analysis of ceramics based on LIBS technology,focusing on the quantitative multi-element analysis of ceramics based on chemometric strategies and the discriminative analysis model of ceramic dynasties,in order to solve the problem of complex spectral analysis by LIBS.This study will provide a theoretical basis and technical support for the realisation of micro-damage inspection of ancient ceramics.The full text is divided into three chapters,and the main research contents are as follow:1.A multi-elemental analysis method for the quantitative analysis of Mg,Al,Ca and Fe in ceramics based on the combination of LIBS technique and random forest regression(RFR)algorithm was developed with ceramics as the research object.(1)A multivariate regression analysis method based on RFR is proposed to investigate the effects of different parameters such as delay time,laser energy and number of spectral accumulations on the prediction results,in view of the influence of interference factors such as laser energy fluctuations,sample inhomogeneity and matrix effects.A full-spectrum,four different spectral bands based RFR calibration model was developed for the rapid determination of the elemental content of Mg,Al,Ca and Fe in ceramics.The results show that the RFR model exhibits better prediction performance than partial least squares regression(PLSR)and support vector machine(SVM)models.The obtained R~2for Mg,Al,Ca and Fe element of ceramics were0.9726,0.9619,0.9805 and 0.9695,and the RMSE were 0.8324,0.8654,0.6892 and0.7395,respectively.(2)To address the problem of unsatisfactory prediction results of the RFR algorithm,a feature selection method based on sequence backward selection(SBS)is proposed.The SBS-RFR calibration model was also constructed to achieve rapid determination of the multi-element content of ceramics.The results show that the SBS-RFR model exhibits superior prediction performance compared with PLSR,SVM and RFR calibration models.The R~2for all four elements were above 0.9800,and the RMSE obtained were 0.3954,0.4627,0.2532 and 0.3791,respectively.Thus,this chapter establishes a multi-element quantitative analysis method for ceramics based on the LIBS technology combined with the RFR algorithm,while the use of suitable feature selection methods can further improve the predictive performance of the calibration model.2.A ceramic discriminant analysis method based on LIBS technology combined with RF algorithm is established with ceramics as the research object.(1)A supervised pattern recognition method based on random forest(RF)combined with variable importance(?)is proposed to address the differences in elemental information between different categories of ceramics,and a ?-RF discriminant model is constructed to achieve the classification discrimination of ceramics in different periods.The prediction performance was compared with that of an unsupervised pattern recognition method based on principal component analysis(PCA).The results show that the ?-RF discriminant model can better achieve the discrimination of different periods of ceramics,and the discrimination sensitivity,specificity and accuracy obtained are 0.8528,0.9710 and 0.9433 respectively.(2)To address the problems of high dimensionality and complex information of LIBS spectral data in the discriminant analysis of ceramic,a feature selection method based on generalized sequence backward selection(GSBS)was proposed,and a GSBS-RF discriminant model was constructed to achieve the classification discrimination of ceramics in different periods.The results show that the discrimination sensitivity,specificity and accuracy obtained by the GSBS-RF model are 0.9526,0.9910 and 0.9782 respectively,which are all higher than those obtained by the RF,?-RF and SBS-RF discriminant models.(3)To address the limitations of the Wrapper-based feature selection method which is computationally intensive,a Filter/Wrapper hybrid feature selection method(MI-DBS)based on mutual information(MI)and bi-directional selection selection(DBS)was proposed,and the MI-DBS-RF discriminant model was constructed to achieve the classification discrimination of ceramics in different periods.The predicted results obtained are compared with RF,?-RF,SBS-RF,GSBS-RF and MI-RF discriminant models.The results show that the MI-DBS-RF model has better prediction performance,with the discrimination sensitivity,specificity and accuracy obtained being 0.9722,0.9956 and 0.9850,respectively.The results show that the MI-DBS-RF model is able to reduce the computational time required for discriminant analysis compared to the ?-RF,SBS-RF and GSBS-RF discriminant models.Thus,this chapter establishes a discriminative analysis method for ceramics of different periods based on LIBS technology combined with RF algorithms,while the use of superior hybrid feature selection methods can further improve the predictive performance and computational efficiency of the classification model.