Rapid quantitative analysis of raw rocks by LIBS coupled with feature-based transfer learning

Abstract

Quantitative analysis of rock samples using laser-induced breakdown spectroscopy (LIBS) is a challenging task due to the significant differences in spectra between pressed pellet samples and rock samples, with pressed pellet samples usually exhibiting stronger spectral lines. The predictive ability of LIBS quantitative analysis models for rock samples is poorer compared to that for pressed pellet samples, and quantitative analysis models constructed using pressed pellet samples cannot be directly applied to predict the composition of rock samples. To address the issue, a quantitative analysis method for raw rocks based on feature transfer learning using transfer component analysis (TCA) is proposed in this paper. By establishing a feature mapping relationship between pressed pellet samples and rock samples, the differences between data features are reduced, thus enabling the training of a more quantitative analysis model. The transfer learning was introduced into a multivariate regression machine learning model for training using pressed pellet samples, which successfully addresses the problem of prediction accuracy of rock sample element contents. In this model, all data were mapped to a high-dimensional reproducing kernel Hilbert space, and a subset of the most similar features was selected to train the quantitative analysis model. Upon testing the model with independent rock samples, the difference between the predicted and true values was significantly reduced. The model yielded a root mean square error of prediction (RMSEP) of 3.7131, 1.0185, 0.2985, 13.0439, and 1.5450 for Si, Al, Fe, Ca, and Mg in rock samples, respectively. These results indicate that LIBS coupled with the transfer learning algorithm can effectively eliminate the differences between the spectra of the pressed pellet and the raw rock and provide another idea for in situ LIBS detection.