基于高维多元数据的酒体感官评价可视分析

张贵宇; 庹先国; 曾祥林; 彭英杰; 王昆; 陈霏; 付朝帅

doi:10.13386/j.issn1002-0306.2020090006

基于高维多元数据的酒体感官评价可视分析

Visual Analysis of Liquor Sensory Evaluation Based on High-Dimensional Multivariate Data

摘要

摘要: 酒体质量评价以感官鉴定为主，但感官评价易受人的身体条件和经验等因素的影响。为提高评价的稳定性和有效性，建立以理化指标数据为依据的评价方法。本文提出了一种基于大数据可视分析的研究方法，挖掘高维多元指标数据对酒体感官特性的影响。首先，对高维多元指标数据的相关性，以及与感官评价的相关性进行分析，通过图模型阐释关系特性，初步建立面向领域的可视分析方法。然后，采用基于机器学习的数据分析技术，结合感官评价构建酒体质量评价模型。最后，结合评价模型对酒体指标参数的重要性进行分析。在此基础上，对模型的评价效果进行了验证，分别选取两类样本进行对比，一类样本包含重要性得分较高的6 项理化指标，包括酒体密度、残糖、挥发性酸、酒精度、硫酸酯和非挥发性酸，另一类样本包含全部11 项理化指标，验证结果显示两类样本的预测结果相近，分类预测误差仅相差0.4%，表明以上6项理化指标是影响感官评价的主要成分。该可视分析方法可降低理化指标的维数，并保留酒体质量的特征信息，用于酒体质量的评价，对酒体质量科学化评价起到重要作用。

Abstract: The quality evaluation of liquor is mainly based on sensory identification, but sensory evaluation is easily affected by various factors such as the body conditions and experience. In order to improve the stability and effectiveness of evaluation, an evaluation method based on physical and chemical indicator data was established. This paper presented a research method based on visual analysis of big data to explore the effects of high-dimensional multivariate indicator data on the sensory characteristics of liquor. First, the correlation between high-dimensional multivariate indicator data and the correlation between data and sensory evaluation was analyzed in the paper. The characteristics of the relationship had been explained through graph models, and built a domain-oriented visual analysis method initially. Then, the data analysis technology based on machine learning was combined with sensory evaluation to build a liquor quality evaluation model. Finally, the importance of indicator parameters was analyzed by combining with the liquor quality evaluation model. On this basis, the evaluation effect of the model was verified. Two types of samples for comparison were selected. One type of sample contained 6 physical and chemical indicators with high importance scores, including density, residual sugar, volatile acidity, alcohol, sulphates and fixed acidity.The other type of sample contained all 11 physical and chemical indicators. The verification results showed that the prediction results of the two types of samples were similar, and the difference in classification prediction error was only 0.4%, indicating that the above 6 physical and chemical indicators were the main components that affect sensory evaluation. It indicated that the visible analysis method could reduce the dimensionality of the physical and chemical indicators, and retain the characteristic information of the liquor quality. This method could be used for the evaluation of liquor quality and played an important role in the scientific evaluation of liquor quality.