基于广义线性模型的激光诱导击穿光谱铜含量检测  

Determination of copper content in raw ores via laser-induced breakdown spectroscopy with generalized linear model

作　　者：何静 刘泉澄 熊中刚 陈林宇 HE Jing;LIU Quancheng;XIONG Zhonggang;CHEN Linyu(School of Information Engineering,Mianyang Teachers'College,Mianyang 621000,China;Key Laboratory of IOT Security at Mianyang Teachers'College of Sichuan Province,Mianyang 621000,China;School of Information Engineering,Southwest University of Science and Technology,Mianyang 621010,China;School of Mechanical Engineering,Guilin University of Aerospace Technology,Guilin 541004,China)

机构地区：[1]绵阳师范学院信息工程学院,四川绵阳621000 [2]物联网安全四川省高校重点实验室,四川绵阳621000 [3]西南科技大学信息工程学院,四川绵阳621000 [4]桂林航天工业学院机电工程学院,广西桂林514004

出　　处：《红外与激光工程》2025年第2期48-57,共10页Infrared and Laser Engineering

基　　金：国家自然科学基金项目(62241304,22305198);四川省基础教育研究专项(MSJJ24-A03);绵阳师范学院博士科研启动基金项目(QD2023A20);绵阳师范学院创新团队项目(CXTD2023PY03)。

摘　　要：在众多激光诱导击穿光谱(Laser-induced Breakdown Spectroscopy,LIBS)分析方法中,主成分分析(Principal Component Analysis,PCA)和偏最小二乘回归(Partial Least Squares Regression,PLSR)是对原始光谱进行线性特征变换来降低光谱冗余信息,但是上述两种方法无法确定哪些谱线属于冗余谱线,导致模型的物理解释性也较差。为深入了解原始谱线在降维-定量模型中的物理意义,采用Lasso、Ridge和Elastic Net等3种广义线性模型(Generalized Linear Models,GLM)对天然铜矿/精矿中的铜含量进行检测。首先对9种铜矿/精矿样本的光谱特性进行了简要分析,然后选定了11条原子谱线和18条离子谱线用于预测建模,最后详细分析了Elastic Net模型中参数α对模型性能和有效分析谱线数量的影响。定量结果表明,Lasso、Ridge和Elastic Net的测试集均方误差(Mean Square Error,MSE)分别为1.706、1.180和1.231,相对于PLSR而言,上述3种方法的MSE分别降低了7.4%、33.2%和36.0%。在分析谱线数量方面,Ridge和Elastic Net模型中29条分析谱线均为有效分析谱线,而Lasso中仅有21条有效分析谱线。显著性分析结果表明,Ridge和Elastic Net的整体性能优于传统的PLSR,而Lasso的模型性能与PLSR相当。Objective Chemometrics is one of the most effective analytical methods among various laser-induced breakdown spectroscopy(LIBS)techniques,with principal component analysis(PCA)and partial least squares regression(PLSR)being the most typical.The above two methods perform linear feature transformation on the original spectra to reduce the issue of redundant spectral lines.It failed to determine which spectral lines are redundant although the achieved features can represent the effective information of the original spectra,finally resulting in poorer physical interpretability of the above two models.Elastic Net is a feature selection and regression modeling method that not only allows nonlinear feature dimensionality reduction of the original spectrum,but also eliminates the multicollinearity problem between spectral lines.Compared to PCA and PLSR,the Elastic Net model has a stronger capacity for model interpretation,and its regression coefficients directly reflect the influence of all original features on the response variable.To insight into the impact of effective spectral lines on quantitative results and to further comprehend the physical implications of LIBS quantitative analysis,this paper proposes three generalized linear models(GLM),Lasso,Ridge,and Elastic Net,to predict the copper content in raw copper ores/concentrates.Methods To understand the physical significance of the original spectral lines in the dimensional reductionquantitative models,Lasso,Ridge,and Elastic Net were utilized to determine the copper content in raw copper ores/concentrates.A spectrum acquisition device was designed using a nanosecond laser and an Echelle spectrometer(Fig.1).Five types of copper ores and four types of copper concentrates were selected as sample materials,finally obtained nine pieces of copper ore tablets(Fig.2).The gate delay of the spectrometer was set to 1μs with a gate width of 2μs.A total of 90 spectra were obtained from the nine types of copper ores/concentrates,of which 70 spectrums were used for training a

关 键 词：激光诱导击穿光谱 定量分析 广义线性模型 弹性网络 套索回归 岭回归 

分 类 号：O43[机械工程—光学工程]