Abstract: Raman spectroscopy is used for the rapid prediction of the content of trans fatty acids (TFAs) after oil heating. Three edible oils were heated at 190℃ (common frying temperature) for different time, and 36 Raman spectra were collected for each sample. Firstly, the original spectral data is preprocessed by polynomial smoothing and standard normal variable transformation to remove the background and noise interference; then the spectral data is filtered by the characteristic variables using uninformative variable elimination (UVE); qualitative and quantitative analysis models were established based on the full spectrum data and the selected characteristic spectral variables. The experimental results were compared and analyzed. The results show that. Fisher discriminant analysis (FDA) is used to establish a qualitative discriminant model, and the discriminant accuracy rate is improved from 40%-50% to over 90% based on the selected variables, indicating that the selected variables can be better characterized. Based on the screening variables and the full-spectrum data, the mathematical prediction model of TFAs content in different samples was established using PLSR, BP neural network and SVR method. Through the comparative analysis of the prediction results, it is shown that the non-information variable elimination combined with the support vector regression machine method (UVE-SVR) had a good detection effect. The R² of the TFAs content rapeseed oil, soybean oil, and corn oil were upgraded from 0.850 4, 0.943 5, and 0.753 4 to 0.952 6, 0.954 8, and 0.958 5, respectively. Therefore, the UVE-SVR method not only simplifies the prediction model, but also improves the stability and accuracy of the model. It also provides a feasible method for the rapid detection of TFA in the edible oil.

Key words: edible oil, trans fatty acids, uninformative variable elimination, Fisher discriminant analysis, support vector regression

摘要： 为实现拉曼光谱技术对食用油加热后反式脂肪酸(TFAs)含量的快速预测,将3种食用油在190℃下(常用煎炸温度)进行不同时间加热,每个样品采集36条拉曼光谱。首先,采用多项式平滑与标准正态变量变换(SNN)对原始光谱数据进行预处理,以去除背景和噪音的干扰,然后采用无信息变量消除法(UVE)对光谱数据进行特征变量筛选,最后分别基于全谱数据和筛选的特征光谱变量建立定性和定量分析模型,并对试验结果进行对比分析。结果表明,基于筛选后的变量,运用Fisher判别分析(FDA)建立定性判别模型,其判别正确率由40%~50%提升至90%以上,表明筛选后的变量能较好表征样品的特征信息;分别基于筛选变量和全谱数据,运用偏最小二乘回归(PLSR)、BP神经网络(BPNN)和支持向量回归机(SVR)方法,建立不同样品中TFAs含量的数学预测模型。通过对预测结果对比分析,表明UVE结合SVR方法具有良好的检测效果,菜籽油、大豆油、玉米油的测试集R²分别从0.850 4、0.943 5和0.753 4升至0.952 6、0.954 8和0.958 5。因此,利用UVE-SVR方法不仅简化了预测模型,提高了模型的稳定性和精度,也为食用油中TFA的快速检测提供了一种可行的方法。

关键词: 食用油, 反式脂肪酸(TFAs), 无信息变量消除法(UVE), Fisher判别分析(FDA), 支持向量回归机(SVR)