基于LSSVM的土壤重金属定量分析

Quantitative Analysis of Soil Heavy Metals Based on LSSVM

为了提高土壤定量分析的精度,分别把偏最小二乘法(PLS)和最小二乘支持向量机(LSSVM)与激光诱导等离子体技术相结合对土壤中的Cu元素进行分析。对比分析了CuⅠ324.75 nm和CuⅠ327.40 nm两条特征谱线,最终选择CuⅠ324.75 nm作为分析谱线。首先对实验参数进行优化。通过对比激光能量、采集延时与信噪比之间的关系,确定最佳能量为90 mJ,最佳采集延时为1000 ns。然后在最佳实验条件下采集五个不同浓度样品的特征光谱,并用内标法、PLS和LSSVM建立定标模型。对比三种模型的拟合系数、均方根误差和平均相对误差,发现由于土壤基体效应和自吸收效应的影响,内标法的定标模型性能较差,拟合程度未达到实验要求,而均方根误差和平均相对误差的数值过大,无法满足实验对于精确度和稳定性的要求。用PLS对定标模型进行校准,相对于内标法而言,定标模型的精确度和稳定性均有明显的提高,R^2由0.8701提高到0.9851,训练集和预测集的均方根误差均下降到了0.1 Wt%量级,平均相对误差虽有所下降,但仍然无法达到实验要求,说明PLS虽然可以在一定程度上提高定标模型的精确度,但在提高稳定性方面仍有欠缺,并不能很好的降低土壤的基体效应与自吸收效应。与内标法和PLS的定标模型相比,LSSVM定标模型的精确度和稳定性最好,R^2提高到了0.9976,模型中的数据点基本分布在拟合曲线上,具有良好的线性相关性。相比于内标法,LSSVM定标模型训练集的均方根误差由3.4488 Wt%下降到0.0187 Wt%,预测集的均方根误差由1.2807 Wt%下降到0.1491 Wt%,体现稳定性的平均相对误差降低了6.24倍。与PLS定标模型相比,LSSVM定标模型的各个参数均有大幅降低,特别是平均相对误差由7.4556%下降到2.1370%,可以满足稳定性要求。说明在提高定标模型精确度与稳定性方面,LSSVM算法更具有优势,能够更好地降低土壤基体效应和自吸收效应带来的影响。

In order to improve the accuracy of soil quantitative analysis,the partial least squares(PLS)and least squares support vector machine(LSSVM)were combined with laser-induced plasma technology to analyze the Cu elements in the soil.Two characteristic lines of CuⅠ324.75 nm and CuⅠ327.40 nm were compared and analyzed,and CuⅠ324.75 nm was selected as the analytical line.Firstly,the experimental parameters were optimized.The relationship between laser energy,acquisition delay time and signal-to-noise ratio were compared.The optimal energy and optimal acquisition delay time were 90 mJ and 1000 ns,respectively.Then,the characteristic spectra of five different concentrations of samples were collected under the optimal experimental conditions.The calibration model was established by standard internal method,PLS and LSSVM.By comparing the fitting coefficient,root mean square error and average relative error of the three models,the calibration model of the standard internal method had poor performance due to the influence of soil matrix effect and self-absorption effect.And the fitting degree did not meet the experimental requirements.The values of the root mean square error and the average relative error were too large to meet the accuracy and stability requirements of the experiment.The calibration model was calibrated with PLS.Compared with the standard internal method,the accuracy and stability of the calibration model were significantly improved.R^2 was increased from 0.8701 to 0.9851.The root mean square error of the training set and the root mean square error of the prediction set were reduced to the order of 0.1 Wt%.But the decrease of the average relative error with PLS model can’t meet the experimental requirement.It indicated that PLS could improve the accuracy of the calibration model rather than the stability of the calibration model.The matrix effect and self-absorption effect of soil cannot be reduced.Compared with the former calibration models,the LSSVM calibration model has better accuracy and stability.The R^2 increased to 0.9976.The data points in the model were basically distributed on the fitted curve with good linear correlation.Compared with the standard internal method,the root mean square error of the LSSVM training set decreased from 3.4488 Wt%to 0.0187 Wt%.The root mean square error of the prediction set decreased from 1.2807 Wt%to 0.1491 Wt%.The average relative error reduced by 6.24 times.Compared with the PLS calibration model,the parameters of the LSSVM calibration model were greatly reduced.The average relative error reduced from 7.4556%to 2.137%,which can meet the stability requirements.It shows that the LSSVM algorithm has advantages for improving the accuracy and stability of the calibration model.It can reduce the matrix effects and the self-absorption effects of soil.