深層学習による複雑なラマンスペクトルの分析を用いたゲラニウムおよびローズエッセンシャルオイル混合物の比率定量

Due to their high economic value, essential oils (EOs) are increasingly subject to adulteration, posing significant challenges for quality assessment and control. Owing to its excellent molecular specificity, Raman spectroscopy has been widely employed for the analysis and evaluation of EO products. In this study, we collected a total of 2700 Raman spectra comprising rose essential oil (REO, n = 900), geranium essential oil (GEO, n = 900), and their mixtures (n = 900). We first constructed a conventional convolutional neural network to distinguish spectra corresponding to varying EO mixing ratios. Interpretability analysis revealed that spectral peaks at 800, 1000, 1002, 1028, 1200, 1378, and 1668 cm^-1, which show notable intensity variations in the averaged spectra of REO and GEO, also played a critical role in model decision-making, indicating that these spectral features can serve as discriminative markers for different EO ratios. Subsequently, we developed a channel attention residual feature extraction network (CARFENet), which employs spectral capturing and spectral separation modules to deconstruct mixed EO spectra into their constituent pure components. CARFENet demonstrated robust performance on the validation set and yielded predicted spectra that closely resembled the true spectra in an external test set, with similarity indices exceeding 0.99. These findings indicate that CARFENet enables the effective quantitative analysis of pure EO components within mixed Raman spectra.