Spectral knowledge-based regression for laser-induced breakdown spectroscopy quantitative analysis. (1st November 2022)

• Record Type: Journal Article
• Authors: Song, Weiran; Afgan, Muhammad Sher; Yun, Yong-Huan; Wang, Hui; Cui, Jiacheng; Gu, Weilun; Hou, Zongyu; Wang, Zhe
• Is Part Of: Expert systems with applications. Volume 205(2022)
• Publication Date: 2022-11-01
• Subjects: Domain knowledge -- Laser-induced breakdown spectroscopy -- Machine learning -- Multivariate analysis -- Quantitative analysis -- Regression; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.117756
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: Spectral knowledge-based regression for LIBS data analysis. Nonlinear modelling with linear and physical interpretability. State of the art in LIBS quantitative analysis. Abstract: Laser-induced breakdown spectroscopy (LIBS) is a promising atomic emission spectroscopic technique for multi-elemental analysis and has the advantages of real-time multi-element measurement, minimal sample preparation and remote detection. However, the quality of LIBS data can be low due to matrix effects and signal uncertainty which hinders the wide application of LIBS. Recent studies attempt to improve the performance of LIBS quantitative analysis using linear and nonlinear multivariate analysis models. Linear models can easily present how key variables contribute to the prediction but suffer from performance degradation if data has a high degree of nonlinearity. Nonlinear models tend to have good performance, but they lack simple and intuitive explanations for the contribution of variables and are prone to overfitting. Moreover, nonlinear models used in LIBS quantitative analysis, such as support vector regression (SVR) and kernel extreme learning machine (K-ELM), are not designed to incorporate domain knowledge of spectral data. In this work, a new machine learning algorithm is proposed, namely spectral knowledge-based regression ( SKR ), which integrates linear and nonlinear models to improve the performance of LIBS quantitative analysis. The linear model is knowledge-driven and built on key variables correlated with analyte composition. The nonlinear model is data-driven and transforms the input data into a kernel matrix. The proposed SKR is tested on 4 LIBS datasets and outperforms 5 baseline methods on 12 of 18 quantification tasks. Moreover, it intuitively explains the contribution of key variables towards prediction and has the same low computational complexity as ridge regression These results demonstrate that SKR inherits the high accuracy of nonlinear modelling and the simple variable interpretability of linear models. Therefore, it can serve as a promising method for improving the accuracy and reliability of LIBS quantitative analysis.