Research On Calibration Transfer And Quantitative Analyse Algorithm Of Martian-like Mineral Based On LIBS

Laser-induced Breakdown Spectroscopy(LIBS)has been widely used due to its rapid detection,no need for sample preparation,and long-distance detection,especially in the field of planetary exploration.fruitful results have been achieved.However,there is such a problem in the practical application of LIBS technology:the quantitative analysis model established on the original spectroscopic instrument cannot be directly applied to another new instrument.The inability of spectral models to be universal among different instruments and the inability to adapt to changes in instrument and experimental conditions has become a difficult problem that hinders the large-scale application of this technology.The solution to this problem is called model transfer(Calibration Transfer)in the field of spectroscopy.To this end,this paper does the following research on this issue:The first is for the "ground fire correction" problem.The master and slave instruments of this problem are from the LIBS instrument(master instrument)on Curiosity and the backup prototype(slave instrument)in the Mars Science Laboratory(MSL).The samples are " The CCCT(ChemCam Calibration Target)sample carried by Curiosity.In this paper,the spectral transfer model based on the Extreme Learning Machine(ELM)method is used to convert the spectra of the master and slave instruments to achieve the purpose of "ground fire correction".At the same time,in order to enhance the universality of the method,this paper deepens the network structure,introduces the auto-encoder(AE)structure,and builds a spectrum based on the deep extreme learning machine(ML-ELM)method.Transfer the model,and transfer the model to the rock standard sample spectral set with more samples and greater spectral difference.The results show that the method used in this paper has better transfer effect than the linear correction and PDS methods.Most of the existing research methods solve the problem of model transfer between spectra of the same resolution,but there is no in-depth research on the model transfer of spectra with different resolutions.Especially when low-resolution spectra are reconstructed and transferred to high-resolution spectra,the prior knowledge of high-dimensional LIBS spectra is often extremely complex,and its degradation is often unknown.In this paper,the super-resolution method is introduced to solve this problem.The convolutional neural network algorithm based on the dynamic attention mechanism is used for model transfer,and three spectral data sets with different resolutions are built for verification,namely the Super-Chem data.dataset,ChemSDU dataset,simulation dataset.The results show that the model transfer algorithm has a good transfer effect and can reconstruct the detailed information of low-resolution spectral loss to the greatest extent.The model transfer methods are roughly divided into two categories,namely the standard model transfer method and the standard model transfer method.In practical situations,it is difficult to obtain spectra of the same set of standard samples on the master and slave instruments respectively,which makes it difficult to implement model transfer with standard samples,but model transfer without standard samples can solve this problem.In this paper,based on the parameter transfer method of transfer learning,the ChemCam data set is used as the source domain data set,and the SDU-LIBS data set constructed by making standard samples and collecting spectra by itself is used as the target domain data set.A quantitative analysis model of convolutional neural network based on Inception structure is built on the source domain dataset,and the Fine-Tune method is used to transfer the model to the target domain dataset.Compared with quantitative analysis of the target domain dataset directly,the transfer learning-based model transfer method has better quantitative accuracy.The quantitative analysis of the spectrum can obtain the information of the component content of the sample,and the model transfer work ultimately also serves the quantitative analysis.The experimental results show that:(1)The random forest(VI-RF)algorithm based on variable importance can effectively extract spectral features and simplify the input variables.Compared with the random forest algorithm and the least squares algorithm,the prediction results are more accurate,and the root mean square Error(RMSEP)is low.(2)Using the VIRF algorithm to quantitatively analyze the sample spectra before and after the transfer,the spectrum of the slave instrument after the model transfer is closer to the model of the master instrument in RMSEP,which verifies the validity of the model transfer.