芘作为多环芳烃(PAHs)类物质广泛存在于自然环境中, 亲脂性强, 对人体有致癌影响。 因此, 食用油中芘的含量的判定对品质的把控具有深远的意义。 采用拉曼光谱与人工智能算法相结合进行多环芳烃的定量分析是当前的一个研究热点。 将一毫升食用油与不同固定浓度的芘液体混合制作样本, 然后制作薄层色谱板与金粒子, 采用薄层色谱和表面增强拉曼散射(SERS)光谱相结合的方法进行实验获得光谱数据, 选取自适应迭代加权惩罚最小二乘算法进行预处理, 再采用Multi parameter-Principal Component Analysis-Back Propagation Neural Network模型方法进行定量分析。 该模型首先在预处理后的光谱中选取两个特征峰进行分峰拟合获取特征峰的高度、 半高宽、 面积等参数。 将两个特征峰的拉曼数据与通过拟合获取的参数进行归一化再采用主成分分析获取关键参数, 将获取的关键参数作为输入层输入基于L2正则化的BP神经网络中, 输出预测浓度。 实验分别采用不同的算法进行浓度预测, 实验结果表明, 通过偏最小二乘算法预测的芘浓度, 其测试集决定系数R2为0.58, 均方根误差(RMSEC)为1.85; 采用线性回归拟合特征峰面积与浓度的规律最终预测的芘浓度, 其测试集决定系数R2为0.26, 均方根误差(RMSEC)为2.28; 采用Multi parameter-Principal Component Analysis-Back Propagation Neural Network模型预测芘浓度, 其测试集决定系数R2为0.99, 均方根误差(RMSEC)为0.31, Multi parameter-Principal Component Analysis-Back Propagation Neural Network模型预测精准度更高, 误差更小。 模型是针对光谱数据信息与样本浓度之间非线性、 高维度的关系, 而建立的预测精度及建模效率均高于同类对比的算法模型。 模型拟合特征峰获取关键变量, 将关键变量与特征峰的拉曼位移都作为特征向量, 因此特征向量较为充分, 模型利用PCA提取拉曼光谱非线性特征并且采用基于L2正则化BP神经网络泛化力强的优点, 防止过拟合, 因此可以更加精准快捷地预测出芘的浓度。

Pyrene, a kind of polycyclic aromatic hydrocarbons (PAHs), widely exists in the natural environment. It has strong lipophilicity and carcinogenic effect on the human body. Therefore, the rapid analysis of pyrene content in edible oil has far-reaching significance for quality control. The quantitative analysis of polycyclic aromatic hydrocarbons using Raman spectroscopy and artificial intelligence algorithm is a current research hotspot. One milliliter of edible oil is mixed with pyrene liquid with different fixed concentrations to make samples, and then a thin-layer chromatography plate and gold particles are made. The experiment is carried out by combining thin-layer chromatography, and surface-enhanced Raman scattering (SERS) spectrum to obtain the spectral data. The adaptive iterative weighted penalty least square algorithm is selected for preprocessing, Then the Multi parameter-Principal Component Analysis- Back Propagation Neural Network model was used for quantitative analysis. Firstly, two characteristic peaks are selected in the preprocessed spectrum for peak fitting, and the parameters such as height, half-width, height and area of characteristic peaks are obtained. Normalized the Raman data of the two characteristic peaks and the parameters obtained by fitting, and then use the principal component analysis to obtain the key parameters. The obtained key parameters are input into the BP neural network based on L2 regularization as the input layer to output the predicted concentration. The experimental results show that the R2 determination coefficient of the test set is 0.58 and the root mean square error (RMSEC) is 1.85; The linear regression is used to fit the law between the characteristic peak area and pyrene concentration. The final predicted pyrene concentration has an R2 determination coefficient of 0.26, and a root mean square error (RMSEC) of 2.28; For the pyrene concentration predicted by the Multi parameter-Principal Component Analysis-Back Propagation Neural Network model, the R2 determination coefficient of the test set is 0.99, and the root mean square error (RMSEC) is 0.31. The multi-parameter principal component analysis-back propagation neural network model has higher measurement accuracy and less error. The model is aimed at the nonlinear and high-dimensional relationship between spectral data information and sample concentration. The prediction accuracy and modeling efficiency are higher than similar comparison algorithms. The model fits the characteristic peak to obtain the key variables and takes the Raman displacement of the variable and the characteristic peak as the characteristic vector, so the characteristic vector is sufficient. The model uses PCA to extract the nonlinear characteristics of the Raman spectrum and adopts the advantages of strong generalization based on L2 regularization BP neural network to prevent overfitting, so that it can predict the concentration of naphthalene more accurately and quickly.