铀矿是核领域最重要的矿产资源之一,核领域迫切需求一种能快速、有效勘探铀矿资源的分析技术,快速勘探出优质矿产资源有助于核工业平稳、健康发展.激光诱导击穿光谱(LIBS)是一种发射光谱元素分析技术,具备目标元素现场快速检测的优点,能够实现铀矿资源快速、准确现场勘探和分析的目的.基于LIBS技术结合机器学习对铀矿中U进行定量分析.共制备 12 组实验样本,9 组样本设置为训练集,用于模型建立和超参数优化;3 组样本设置为验证集,仅用于模型验证.使用偏最小二乘(PLS)和随机森林(RF)两种算法建立定量模型,采用十折交叉验证方法对两个模型的超参数进行优化,最终采用三个验证集验证和对比两种模型的定量效果.经过超参数优化后,两个定量模型均具备良好的线性相关性和模型稳定性,其中RF定量模型的线性相关系数为 0.996,而PLS定量模型的线性相关系数为 0.997.在模型验证方面,RF 模型对三个验证集的相对误差分别为 22.33%、12.79%和12.04%;PLS模型对三个验证集的相对误差分别是 4.33%、6.63%和 6.85%.对比两种定量模型的验证结果,PLS模型在三个验证集上的相对误差均低于RF模型,表明PLS模型比RF模型在验证集上具有更高的定量准确度.验证结果表明与RF算法相比,PLS算法更适用于铀矿中U的LIBS定量分析.
Uranium ore is one of the most important mineral resources in nuclear industry,and the nuclear industry is in urgent need of an analytical technology that can quickly and effectively explore uranium ore resources.Rapid and effective exploration of uranium ore resources could promote the stable and healthy development of the nuclear industry.Laser induced breakdown spectroscopy(LIBS)is a kind of emission spectral element analysis technology,which has the advantages of rapid detection of multi-target elements,and can realize the purpose of rapid and accurate field exploration and analysis of uranium resources.The quantitative analysis of U in uranium ore was carried out based on LIBS technology and machine learning.A total of 12 groups of experimental samples were prepared in this work,and 9 groups of samples were set as training sets for model establishment and hyper-parameter optimization.Three sets of samples were set for model validation.Partial least squares(PLS)and Random forest(RF)algorithms were used to establish a quantitative model,and the hyper-parameters of the two models were optimized by the ten-fold cross-validation method.Finally,the quantitative effects of the two models were verified and compared by three verification sets.After hyper-parameter optimization,both quantitative models had good linear correlation and model stability.The linear correlation coefficient of RF quantitative model was 0.996,while that of PLS quantitative model was 0.997.In terms of model validation,the relative errors of RF model for the three validation sets w ere 22.33%,12.79%and 12.04%,respectively.T he relative errors of PLS model for the three verification sets were 4.33%,6.63%and 6.85%,respectively.Compared with the verification results of the two quantitative models,the relative error of the PLS model on the three verification sets was lower than that of the RF model,which indicated that the PLS model had higher quantitative accuracy than the RF model.Therefore,compared with RF algorithm,PLS algorithm is more suitable for quantitative analysis of U LIBS in uranium ore.