雷竞技ray官网王加华副教授团队：使用拉曼光谱法检测山茶油中多物种植物油的掺假：化学计量学和深度学习方法的比较Raybet在线官网食品品质检测与安全评价新技术课题组在国际著名期刊《Food Chemistry》（Q1区，IF=8.5）上发表了题为“Adulteration detection of multi-species vegetable oils in camellia oil using Raman spectroscopy: Comparison of chemometrics and deep learning methods”的文章。王加华副教授为论文第一作者，2022级硕士研究生钱奖金为主要完成人，刘小丹博士、皮付伟教授为论文通讯作者。山茶油是从山茶树的种子中提取的，中国的产量约占全球总量的 90%。山茶油富含不饱和脂肪酸、脂溶性维生素、番红皂苷、多酚等功能性物质，由于其高品质，现在通常用作食用油。由于山茶油与其他植物油之间的巨大价格差异，它可能成为掺假的目标，例如，将更便宜的植物油混合到高价值的山茶油中，以欺骗消费者以获得不应有的利益。因此，构建快速、准确、易作的掺假检测方法是保障山茶油行业有序发展和市场健康的必要手段之一。
在本研究中，设计了三种深度学习方法，即 LSTM 、 ConvLSTM 和 Transfomer ，来识别和量化山茶油的掺假。将深度学习模型的定性和定量结果与传统化学计量学方法进行统计比较。模型效用的比较研究结果可为复杂食品的掺假检测模型选择提供有益的参考。Fig. 1.Network structure of LSTM (a), ConvLSTM (b) and transformer (c). The LSTM, ConvLSTM and Transformer programs were run in the Python (Versions 3.9.18).Fig. 2. Raman spectra (the spectral range of 1800–2700 cm1 was removed due to the absence of significant peaks) of six edible oil samples (a), plots of principal component scores based on 11 characteristic peak intensities (b) and 7 PIRs (c).
Fig. 3. ROC curves and AUC scores for machine learning classifiers based on pure CAOs and adulterated CAOs with adulteration levels greater than 10 % (a), 5 % (b) and 1 % (c); results of the deep learning classifiers with adulteration levels greater than 10 % (d), 5 % (e), and 1 % (f); the line indicates the average ROC curve for each method, calculated from the individual ROC curves for each data split.
Fig. 4. Comparison of FRP, TNR, FNR and TPR (a) and significance analysis of AUCs (b) for machine learning and deep learning classifiers with adulteration levels greater than 10%, 5% and 1%; One-way ANOVA of RMSEP and RPD for chemometrics and deep learning models of adulteration levels (c); Scatterplot of the optimal ConvLSTM model for the adulteration level (d).
