基于近红外光谱和SVM算法对琥珀掺伪的定性鉴别与定量分析

Qualitative and Quantitative Analysis of Amber Adulteration Using Near Infrared Spectroscopy Based on Support Vector Machine

目的:本研究利用近红外光谱结合支持向量机算法(SVM)等多种化学计量学方法建立定性模型对正品琥珀与掺松香伪品进行鉴别,并建立定量模型对琥珀中掺入松香的含量进行定量预测。方法:首先采用近红外光谱扫描琥珀、松香以及按照琥珀与松香不同质量比配制的掺伪品。在对正品琥珀与掺有松香的掺伪品定性分析时,对采集的原始图谱分别进行矢量归一化、一阶求导预处理,利用主成分分析法对原始光谱和不同预处理后的光谱降维,采用支持向量机分类法(SVC)对琥珀和掺有松香的琥珀掺伪品建立定性鉴别模型,进行分类预测,对比不同预处理方法下的模型预测性能并确定最佳预处理方法。在定量分析中,分别应用偏最小二乘法(PLS)和支持向量机回归法(SVR),对琥珀掺伪品中松香的含量进行建模和预测。支持向量机(SVM)模型参数采用网格搜索法(grid search)、粒子群算法(PSO)和遗传算法(GA)进行优化,比较不同的优化方法确定的参数所建模型效果。结果:建立的VN-GA-SVC模型校正集和验证集预测准确率分别达到100%和97.37%;应用PLS和SVR建立的定量模型决定系数R^2均达到99.7%以上,MSE最低达到3.03×10-4。结论:该研究建立的定性、定量模型稳定性好,预测结果准确可靠,能够实现琥珀与掺松香伪品的鉴别以及松香含量的定量预测。

Objective：To establish an identification model on amber and its adulterants of colophony by near infrared spectroscopy （NIR） combined with support vector machine（SVM） ,and a quantitative model was built up to predict the content of colophony in adul- terants. Methods ：The near infrared spectra of samples were collected and preprocessed by vector normalization and first derivative re- spectively. Principal component analysis （PCA） method was used to reduct the dimension, qualitative model of adulterated amber was es- tablished by support vector machine classifier（SVC）. The optimal preprocessing method was determined by comparing prediction per- formance. Quantitative model of adulterated amber was established respectively by Partial Least Squares （PLS） and support vector ma- chine regress （SVR）. Model parameters （C ,g）were optimized and determined by grid search, particle swarm optimization and genetic algorithm. Results ：The accuracies of prediction set and calibration set for SVC model were up to 100% and 97. 37% ;the coefficient of determination（R2） of PLS and SVR were all above 99. 7% , the lowest MSE was 3.03 × 10^-4. Conclusion：The established models are relatively stable, accurate and reliable for the rapid identification and the prediction of colophony content in the amber adulteration.