提出了一种基于激光拉曼光谱和人工蜂群智能优化支持向量回归机(ABC-SVR)算法的快速定量检测三组分混和油中3种脂肪酸含量的方法。 该方法针对光谱数据信息与样本之间非线性、 高维度的关系, 建立了预测精度及建模效率均高于同类对比算法的数学模型, 同时避免了气相色谱法、 液相色谱法等对混合油脂肪酸含量的检测方式, 根据纯种油中3种脂肪酸含量的国际标准, 由油品配置体积得到脂肪酸质量, 有效降低了检测成本与实验复杂程度, 提高了检测工作的实用价值。 首先根据一定梯度配置66组混合油检测样品, 使用便携式拉曼光谱仪采集样本的拉曼光谱信息, 扣除背景噪声; 观察多组样本的拉曼光谱图可知, 由于官能团浓度的差异, 食用油的拉曼特征峰位移基本相同, 特征峰的峰值明显不同, 因此基于特征峰信息可以区分食用调和油的不同混合物; 其次对拉曼光谱做背景扣除、 光谱平滑、 最大值谱线归一化三步预处理, 以降低实验中不可控的外界因素及背景荧光的影响, 准确提取光谱特征峰强度信息; 然后根据纯种油中3种脂肪酸的国际标准含量, 结合国家食品法典委员会标准CODEX STAN210—1999《指定的植物油法典标准》中规定的纯种油密度中值, 由油品体积得到脂肪酸质量数; 随机选取56组样本数据作为训练集, 剩余10组样本数据作为预测集; 以训练集光谱特征峰强度和脂肪酸质量分别作为回归模型的输入及输出值, 建立SVR和PSO-SVR, ABC-SVR三种混合优化算法对比的定量分析模型, 对测试集的3种脂肪酸含量分别进行预测; 最后通过均方误差(MSE)、 相关系数(r)及建模时间(Elapsed time)分别进行对比, 建立数据表对模型精准度进行检验。 实验结果表明, 通过ABC-SVR定量分析模型效果最佳, 3种脂肪酸含量预测值与真实值的均方差分别为0.88×10-4, 16×10-4和8×10-4, 均低于0.002; 相关系数分别为93.43%, 99.65%和99.43%, 均高于93%; 预测时间(Elapsed time)分别为1.26, 2.42和2.14 s。 因此, 所提出的检测方法, 具备较高的精确度、 较快的建模时间, 且在理论上的类似条件下可适用于其他样品检测工作, 可为振动光谱学对食用油掺伪分析的进一步工作提供可行的理论依据。

In this paper, a rapid and quantitative detection method combing the laser Raman spectroscopy with Artificial Bee Colony-Support Vector Machine for Regression (ABC-SVR) is proposed for the determination of fatty acids content in three-component blend oil. This method establishes a mathematical model with higher prediction accuracy and higher modeling efficiency than similar comparison algorithms in solving the nonlinear and high-dimensional complex relationship between spectral data information and samples. And it can avoid complicated detection ways like gas chromatography and liquid chromatography, etc. The quality of fatty acids is obtained from the oil configuration volume according to the international standards for the content of three fatty acids in pure oils, which effectively reduces the cost and complexity of the experiment, and increases the practical value of the inspection work. Firstly, 66 groups of mixed oil test samples were arranged according to a certain gradient. The Raman spectroscopic information of the samples was collected from a portable Raman spectrometer and the background noise was subtracted at the time. Through the spectrum of the samples, we could see that the Raman spectra had the same characteristic peak shifts basically, but the intensities of the characteristic peaks were obviously different because of the difference in functional group concentration. Therefore, different components could be distinguished according to the characteristic peak information. Secondly, the spectra were pretreated by background subtraction, spectral smoothing and normalization to reduce the effect of uncontrollable external factors in the experiment. Then the mass of fatty acid was obtained from the oil volume by the international standard content of three kinds of fatty acids in pure oil in National Codex Alimentarius Commission Standard CODEX STAN210—1999.2/3 of sample data were randomly selected as the training set, and the remaining 1/3 of sample data were used as the prediction set. The characteristic peak intensity and the quality of fatty acid of train set were used as the input and output values of the regression model, and the quantitative analysis model of hybrid optimization algorithms of SVR, PSO-SVR and ABC-SVR were established to predict the content of fatty acids of test set. The accuracy of the model was tested by using mean squared error (MSE), the correlation coefficient (R2) and elapsed time. The experimental results showed that the ABC-SVR quantitative analysis model was effective: the MSE of the predicted and true values of three fatty acid contents were 0.88×10-4, 16×10-4 and 8×10-4, respectively. The R2 were 93.43%, 99.65% and 99.43%, respectively. The elapsed time were 1.26, 2.42 and 2.14 s, respectively. Therefore, the proposed method has higher accuracy, faster modeling time than other ways, and it can be applied to other sample detection work under theoretically similar conditions. This method can provide a viable theoretical basis for further study on the analysis of adulterated edible oils by vibration spectroscopy.