基于近红外光谱与Transformer的烟叶感官指标预测方法

Prediction Method of Tobacco Sensory Indicators Based on Near Infrared Spectroscopy and Transformer

摘要:

为克服传统卷烟配方设计与维护过程中存在的主观性强、过度依赖人工经验及感官评吸等技术瓶颈，利用“近红外光谱-化学成分-感官指标”的间接关联，提出了一种基于近红外光谱与Transformer架构的端到端烟叶感官质量指标预测方法。首先采用Savitzky-Golay卷积平滑法（SG）、一阶导数法（D1）、多元散射校正（MSC）3种光谱预处理技术有效消除基线漂移和散射干扰；进而设计了一种面向光谱数据特征的Transformer预测模型，实现了烟叶感官质量三维评价体系（风格特征：清香、甜香、焦香；烟气特征：浓度、劲头；质量特征：香气质、香气量、杂气、刺激、余味）的精准预测，并采用了SHAP方法对模型进行分析，增强了模型的可解释性。结果表明，模型对各感官指标测试集预测的平均绝对误差均不高于0.56，具有较好可用性；针对不同感官指标，模型表现出对不同光谱特征波段的捕捉，有效挖掘了光谱特征的协同作用机制，具有较好可解释性。在此基础上，进一步结合多维相似度分析设计了一种辅助烟叶替代方法，可为烟叶替代与配方优化提供量化决策支持。

Abstract:

In order to overcome the technical bottlenecks of strong subjectivity, over-reliance on manual experience and sensory evaluation in the process of traditional cigarette formula design and maintenance, an indirect correlation model of “near infrared spectroscopy-chemical composition-sensory indicators” was constructed, and an end-to-end tobacco sensory quality indicators prediction method was proposed based on near infrared spectroscopy and Transformer architecture. Firstly, three spectral preprocessing techniques, Savitzky-Golay convolution smoothing method (SG), first derivative method (D1), and multivariate scattering correction (MSC), were used to effectively eliminate baseline drift and scattering interference;then a Transformer prediction model oriented to spectral data features was designed to achieve accurate prediction of the three-dimensional evaluation system of tobacco sensory quality (style characteristics: freshness, sweet, and burnt;smoke characteristics: concentration and strength;quality characteristics: quality of aroma, volume of aroma, offensive taste, irritating, and pleasant aftertaste). The model was analyzed by using the SHAP method to enhance its interpretability. Results showed that the model’s mean absolute error for each sensory indicators test set was no more than 0.56, demonstrating good usability. For different sensory indicators, the model demonstrated strong capture of distinct spectral feature bands, effectively exploring the synergistic mechanism of spectral features and demonstrating good interpretability. Furthermore, a method for assisting tobacco leaf substitution was designed by combining multidimensional similarity analysis, providing quantitative decision support for tobacco leaf substitution and blend optimization.