基于多核支持向量回归的浓香型白酒风味成分逐步预测模型研究

Study on stepwise prediction model for flavor components of strong-flavor Baijiu based on multicore support vector regression

中文关键词:  白酒感官  风味成分  数据增强  遗传算法  多核支持向量回归

英文关键词:Baijiu sensory  flavor components  data enhancement  genetic algorithm  multi-kernel support vector regression

中文摘要:

      目的 探究白酒感官品评与白酒风味成分之间的关系, 实现通过感官品评对风味成分进行预测。方法 采用变分自编码器(variational auto encoder, VAE)对原始数据进行增强, 以多核支持向量回归(multi-kernel support vector regression, MKSVR)结合遗传算法(genetic algorithm, GA)建立单预测模型, 再采取逐步预测的方式按照酸、酯、醇、醛类物质的顺序进行预测, 从而构建最终模型。结果 在经过VAE对数据进行增强的条件下, 多元线性回归(mixed logistic regression, MLR)对酸、酯、醇、醛类物质预测的拟合优度分别为0.9660、0.9106、0.8767、0.8686, 随机森林(random forests, RF)对酸、酯、醇、醛类物质预测的拟合优度分别为0.9663、0.9186、0.8805、0.8708, GA-MKSVR对酸、酯、醇、醛类物质预测的拟合优度分别为0.9715、0.9423、0.9072、0.8809, GA-MKSVR逐步预测对酸、酯、醇、醛类物质预测的拟合优度分别为0.9715、0.9447、0.9102、0.8851, GA-MKSVR逐步预测的效果均为最优。结论 GA-MKSVR逐步预测方法相较于传统的机器学习方法, 具有更好的性能, 对数据具有更高的适应性, 能更好地构建白酒感官与风味成分之间的关系模型。

英文摘要:

      Objective To explore the relationship between the sensory evaluation of Baijiu and the flavor composition of Baijiu, the flavor composition can be predicted through sensory evaluation. Methods The original data was enhanced by variational auto encoder (VAE), and a single-prediction model was established by multi-kernel support vector regression (MKSVR) combined with genetic algorithm (GA), and then the stepwise prediction method was adopted to predict in the order of acids, esters, alcohols and aldehydes, this builds the final model. Results Under the condition that the data were enhanced by VAE, the goodness-of-fit of mixed logistic regression (MLR) for the prediction of acids, esters, alcohols and aldehydes was 0.9660, 0.9106, 0.8767 and 0.8686, respectively, and the goodness of fit of random forests (RF) for acids, esters, alcohols and aldehydes was 0.9663, 0.9186, 0.8805, 0.8708, respectively. The goodness of fit of GA-MKSVR for the prediction of acids, esters, alcohols and aldehydes was 0.9715, 0.9423, 0.9072 and 0.8809, respectively, and the goodness of GA-MKSVR stepwise prediction for acids, esters, alcohols and aldehydes was 0.9715, 0.9447, 0.9102 and 0.8851, respectively, and the effect of GA-MKSVR stepwise prediction was the best. Conclusion The GA-MKSVR stepwise prediction method has better performance and higher adaptability to data than traditional machine learning methods, and can better construct the relationship model between Baijiu sensory and flavor components.