Enhancing LIBS analysis accuracy of C element in low-carbon alloy steel by automatic segmented modelling with nonlinear-regression-based spectral line selection and Mahalanobis distance kernel space classification

Abstract

The instability of the Laser Induced Breakdown Spectroscopy (LIBS) spectral intensity amplitude of C element in low-carbon alloy steel greatly decreases the analysis accuracy. Traditional one model strategies usually suffer from the over-fitting problem, while multi-model strategies heavily depend on the prior knowledge of segmentation. In this paper, an automatic segmented model with nonlinear-regression-based spectral line selection and Mahalanobis distance kernel space classification is proposed to enhance the LIBS quantitative analysis accuracy of C element in low-carbon alloy steel. The nonlinear logarithmic PLSR method is used for analysis line selection. For predicting the concentration of an unknown test sample, the Mahalanobis Distance Kernel Space Classification (MD-KSC) method is utilized to determine which segment model should be applied. Experiments on the alloy steel dataset which consists of 7 standard samples and 30 industrial production samples, with C concentrations less than 0.102%, were carried out. Results show that with the proposed automatic segmented modelling the total Mean Relative Errors (MREs) for predicting C element can reach 2.75%, which is better than those of the traditional methods. The comparative experiments also verified that the nonlinear logarithmic PLSR based analysis line selection method is superior to the Ridge-RFE and the second derivative method.