CALIBRATION TRANSFER FOR ChemCam SPECTRAL DATA FROM DIFFERENT LASER-INDUCED BREAKDOWN SPECTROMETERS VIA A DEEP EXTREME LEARNING MACHINE

Abstract

Laser-induced breakdown spectroscopy (LIBS) technology has been applied to many fields, so it is crucial for quantitative analyses of LIBS spectra. However, there is a problem in the field of LIBS spectra. Even in the same experimental setting, the same sample exhibits different spectra with different instruments, which is mainly reflected in the intensity, wavelength shift, and peak width differences. These differences cause standardization problems in LIBS spectroscopy and serious interference in quantitative analyses. The aim of this study is to correct the difference by applying the extreme learning machine method and the deep extreme learning machine method to two different spectral datasets. The first dataset is the ChemCam calibration target sample set, which contains two spectral datasets produced by using ChemCam on instruments on the Curiosity rover and at the Mars Science Laboratory (MSL). The other dataset comprises spectra obtained from calibration target samples produced by the ChemCam team at MSL. The performance of the two algorithms is tested, and the results show that our calibration transfer methods are stable predictive methods that provide significantly lower prediction error compared with linear transfer and the piecewise direct standardization method. The model established by the partial least square method is used for quantitative analyses of the transferred spectra, and the transmitted spectra showed improved quantitative accuracy.