反射光谱结合光谱基二维卷积回归网络快速检测食用油中饱和脂肪酸

doi:10.3964/j.issn.1000-0593(2022)05-1490-07

摘要
参考文献
相关文章 (15)

全文: PDF (2705 KB)
输出: BibTeX | EndNote (RIS)

摘要：人们日常膳食中常见的食用油含有丰富的饱和脂肪酸，饱和脂肪酸能为人体提供能量和必须营养物质，但过量摄入会导致多种心血管疾病。结合反射率光谱和深度学习方法发展一种食用油中饱和脂肪酸含量的分析方法。首先，测量了菜籽油、大豆油、葵花籽油、玉米油、橄榄油、芝麻油及花生油等7种食用植物油350～2 500 nm范围的反射光谱，并通过气相色谱-质谱分析法获得其软脂酸、花生酸及山嵛酸等饱和脂肪酸的含量。使用中心化、多元散射校正、标准正态变量变换及标准化等算法做光谱预处理消除光谱噪声。然后，构建了一种新型的二维光谱卷积回归网络（S2DCRN）用于脂肪酸分析，而全卷积网络（FCN）、偏最小二乘回归（PLSR）、支持向量回归（SVR）及随机森林（RF）用于与S2DCRN模型相对比。最后，采用序列前向选择（SFS）、随机蛙跳（RFrog）及遗传算法选取光谱特征的重要波长，进而构建更为简单稳健的分析模型。实验结果表明，对食用油的全光谱预处理后，S2DCRN模型性能最优，其模型对预测集的决定系数（R²_P）达到0.987 9，均方根误差（RMSE_P）为0.510 0。基于重要波长的S2DCRN模型，RFrog-SFS为S2DCRN提供了最佳的预测结果R²_P=0.967 9，RMSE_P=0.462 7。虽然变量选择后所取得的分析效果略差，但光谱波长数目不足全光谱的1%，节省了光谱数据采集工作并大幅降低了模型复杂度，有助于后续便携式简化检测装置的研发。为进一步探究S2DCRN模型的通用性能，S2DCRN被用于食用油中花生酸和山嵛酸含量分析。其中，S2DCRN对花生酸的预测结果较好R²_P=0.950 1，RMSE_P=0.152 9。所提出的S2DCRN可实现反射率光谱对食用油中多种脂肪酸的准确快速分析。

关键词：食用油；饱和脂肪酸；反射光谱；卷积神经网络

Abstract：The edible oil in the humandaily diet is rich in saturated fatty acids, which can provide energy and other healthy nutrients for the human body, but excessive intake of saturated fatty acids can lead to a variety of cardiovascular diseases. In this study, a method for analyzing the content of saturated fatty acidsin edible oils was developed by combining reflectance spectroscopy and machine learning. Firstly, the reflectance spectra of 7 edible vegetable oils, such as rapeseed oil, soybean oil, sunflower seed oil, corn oil, olive oil, sesame oil and peanut oil, were measured in the range of 350～2 500 nm, as well as the contents of palmitic acid, arachidonic acid and behenic acid were obtained by GC-MS. Spectral preprocessing algorithms were employed to eliminate the noise in spectra, including centralization, multiple scattering correction, standard normal variable transformation and standardization. Then, a novel two-dimensional spectral convolution regression network (S2DCRN) was constructed for fatty acids analysis, and a full convolutional network (FCN), partial least squares regression (PLSR), support vector regression (SVR) and random forest (RF) were compared with S2DCRN. Finally, sequential forward selection (SFS), random frog (RFrog) and genetic algorithm were used to select important wavelength spectra to re-build more simple and robust analysis models. The results showed that the S2DCRN obtained optimal performance after pretreatment of edible oil spectra with the determination coefficient of prediction set (R²_P) of 0.987 9 and the root mean square error of prediction set (RMSE_P) of 0.510 0. Based on important wavelengths selected by combination RFrog and SFS, the S2DCRN exhibited excellent performance with R²_P=0.967 9 and RMSE_P=0.462 7. Although the results based on important wavelengths obtained is slightly worse, the number of wavelengths is less than 1% of the full spectra. It is convenient for the measurement of spectra and remarkably reduces the complexity of the model, which is helpful for the further development of portable and simplified detection devices. In addition, to further explore the generalization and applicability of S2DCRN, S2DCRN was used to analyze the content of arachidonic acid and behenic acid and gain a prediction result for arachidonic acid R²_P=0.950 1, RMSE_P=0.152 9. Therefore, the proposed method accurately and rapidly analysed various fatty acids in edible vegetable oils by reflectance spectroscopy.

Key words：Edible oil; Saturated fatty acids; Reflectance spectroscopy; Convolutional neural network

收稿日期: 2021-01-06 修订日期: 2021-06-29

中图分类号:

O657.39

基金资助: 国家自然科学基金项目（32001421），安徽省重点研究与开发计划项目（202004a06020032）资助

通讯作者: 翁士状

作者简介: 翁士状，1989年生，安徽大学电子与信息工程学院副教授 e-mail: weng_1989@126.com *通讯作者

引用本文:

翁士状，储昭结，王满琴，王年. 反射光谱结合光谱基二维卷积回归网络快速检测食用油中饱和脂肪酸[J]. 光谱学与光谱分析, 2022, 42(05): 1490-1496.
WENG Shi-zhuang, CHU Zhao-jie, WANG Man-qin, WANG Nian. Reflectance Spectroscopy for Accurate and Fast Analysis of Saturated Fatty Acid of Edible Oil Using Spectroscopy-Based 2D Convolution Regression Network. SPECTROSCOPY AND SPECTRAL ANALYSIS, 2022, 42(05): 1490-1496.

链接本文:

https://www.gpxygpfx.com/CN/10.3964/j.issn.1000-0593(2022)05-1490-07 或 https://www.gpxygpfx.com/CN/Y2022/V42/I05/1490