Xiao Qinlin, Gu Ruifang, Li Li, Wen Jing, Zhang Xixiang, Shen Yi, Liu Yang, Xiao Lan, Tang Qinqin, Yang Jun, He Yong, Yang Juan
Technology Center, China Tobacco Sichuan Industrial Co., Ltd., Chengdu, China.
College of Biosystems Engineering and Food Science, Zhejiang University, Hangzhou, China.
Front Plant Sci. 2025 Aug 7;16:1617958. doi: 10.3389/fpls.2025.1617958. eCollection 2025.
Accurate detection of blending proportions in tobacco formulations is crucial for ensuring the quality consistency and flavor stability of cigarette products. In recent years, modeling approaches based on near-infrared spectroscopy (NIRS) have attracted significant attention for the quantitative analysis of tobacco blending. However, due to variations in tobacco composition and spectral characteristics across different cigarette brands, the generalization ability of NIRS-based models often declines when applied to cross-brand prediction tasks. To address this issue, this study takes the detection of blending proportions of tobacco silk in tobacco formulations as the research focus, and investigates transfer learning strategies aimed at enhancing the cross-brand adaptability of NIRS-based models. A partial least squares regression (PLSR) model was first developed using NIRS data from four different tobacco brands, achieving high prediction accuracy on the combined dataset (RMSEP = 1.20%). However, when the model trained on a single brand was applied to predict other brands, the prediction performance decreased notably. To improve model adaptability, three approaches were explored: Transfer Component Analysis (TCA), Correlation Alignment (Coral), and model updating. The results show that TCA-PLSR achieved substantial reductions in prediction error in most transfer tasks involving large discrepancies in feature distributions. Coral-PLSR demonstrated superior performance in transfer tasks involving similar spectral feature distributions. Additionally, in transfer tasks characterized by substantial distribution differences, the Updated-TCA-PLSR model, which incorporates a small proportion of target domain samples into the source domain before domain adaptation, yielded accurate predictions of tobacco silk blending proportions. These findings demonstrate that transfer learning and model updating offer practical, flexible, and robust approaches for enhancing the performance of NIRS-based models, supporting more accurate and consistent quality control in industrial-scale formulated tobacco production.
准确检测烟草配方中的混合比例对于确保卷烟产品的质量一致性和风味稳定性至关重要。近年来，基于近红外光谱（NIRS）的建模方法在烟草混合定量分析方面引起了广泛关注。然而，由于不同卷烟品牌的烟草成分和光谱特征存在差异，基于NIRS的模型在应用于跨品牌预测任务时，其泛化能力往往会下降。为了解决这一问题，本研究以检测烟草配方中烟丝的混合比例为研究重点，探讨旨在提高基于NIRS的模型跨品牌适应性的迁移学习策略。首先使用来自四个不同烟草品牌的NIRS数据开发了偏最小二乘回归（PLSR）模型，在组合数据集上实现了较高的预测准确率（RMSEP = 1.20%）。然而，当在单个品牌上训练的模型用于预测其他品牌时，预测性能显著下降。为了提高模型适应性，探索了三种方法：迁移成分分析（TCA）、相关对齐（Coral）和模型更新。结果表明，在大多数涉及特征分布差异较大的迁移任务中，TCA - PLSR实现了预测误差的大幅降低。Coral - PLSR在涉及相似光谱特征分布的迁移任务中表现出卓越性能。此外，在以分布差异较大为特征的迁移任务中，在域适应之前将一小部分目标域样本纳入源域的更新TCA - PLSR模型，对烟丝混合比例做出了准确预测。这些发现表明，迁移学习和模型更新为提高基于NIRS的模型性能提供了实用、灵活且稳健的方法，支持在工业规模的配方烟草生产中进行更准确和一致的质量控制。
