李力群　纪旭东　乔月梅　牛文广　叶亚军　郭春生

摘要[目的]科学地评价卷烟配方中劲头的大小，通过建立BP神经网络模型预测卷烟劲头。[方法]以烟叶游离烟碱百分含量、总烟碱百分含量、结合态烟碱百分含量、游离烟碱占总烟碱比率和水浸液pH作为BP神经网络的输入，感官劲头作为输出，网络训练前对输入指标作归一化处理，然后通过训练样本数据对网络进行充分的训练，获得适宜的参数矩阵，得到卷烟劲头的网络预测模型，最后用训练好的网络模型对检验样本数据进行预测。[结果]卷烟配方中劲头大小的预测值与实际值相对标准偏差小于5%，达到了较好的预测结果。[结论]建立了卷烟劲头的BP神经网络预测模型，该模型对于预测卷烟劲头具有指导意义。

关键词 BP神经网络；烟碱；卷烟劲头

中图分类号 S126 文献标识码 A 文章编号 0517-6611（2016）05-021-03

Abstract[Objective]To scientifically evaluate the cigarette impact， and to predict the cigarette impact through the BP neural network.[Method]The percentage of free nicotine in leaves， the percentage of total nicotine， the percentage of combined state nicotine， the percentage of free nicotine in total nicotine， and pH value of aqueous extracts were used as the input of BP neural network. And sensory momentum was used as the output. Normalization processing of input index was carried out before network training. Network was fully trained before network training. Then， network was fully trained by training sample data， so as to obtain the proper parameter matrix， and to obtain the network forecast model of cigarette impact. Finally， test sample data were forecasted by the trained network model.[Result]Relative standard deviation between predicted value and actual value was smaller than 5%， which reached relatively good predicted value.[Conclusion]Prediction model of cigarette impact through the BP neural network is established， which has guiding significance for the prediction of cigarette impact.

Key words BP neural network； Nicotine； Cigarette impact

烟草中烟碱含量是烟草和卷烟质量控制的一项重要指标，烟碱可以以游离态、单质子态和双质子态3种形态存在[1-2]。质子态烟碱被口腔吸收的速度较慢，而非质子化游离态烟碱挥发性极强，可以穿过口腔黏膜迅速被人体吸收，对中枢神经的药理作用非常强烈，抽吸时表现为劲头较强[3-4]。卷烟劲头不仅与烟气中总烟碱量有关，而且与烟碱的形态密切相关，而卷烟的水浸液pH能够影响烟碱的存在状态[5]。

BP（Back Propagation）网络是一种按误差逆传播算法训练的多层前馈网络，BP网络能学习和存贮大量的输入-输出模式映射关系，而无需事前揭示描述这种映射关系的数学方程。……