摘要
烟叶化学成分是影响烟叶质量的物质基础,烟叶化学成分与评吸质量的关系探究一直是非常重要的研究领域。传统的数理统计方法无法直接给出烟叶感官质量的预测。BP神经网络具有很强的非线性映射能力,但是容易出现"过拟合"。由于烟叶样本数据的噪声较大,为了提高预测模型的泛化性能,本文在应用BP神经网络建模时,采用"权值衰减"策略和样本验证策略进行训练。结果表明,设计合理的BP网络可以较好的对烟叶的常规化学成分进行感官质量预测。
Chemical constituents in tobacco leaf are substance bases of tobacco's quality.It has been always an important research sublect on relationship between chemical constituents and sensory quality.Conventional statistic methods can not get estimation of sensory quality directly.BP neural networks have strong capability of mapping nonlinear relations,though "over-fitting" emerges frequently.Because sample data can be affacted by noise,therefore strategy of weights-fading and validation samples were adopted in network training to enhance model generalization.This study showed that it is acceptable to use BP networks with good-design to predict sensory quality in tobacco leaf by chemical constituents.
出处
《中国烟草学报》
EI
CAS
CSCD
2011年第1期19-25,共7页
Acta Tabacaria Sinica
基金
川渝中烟工业公司科技攻关项目(CYZY200701)资助
关键词
烟叶
化学成分
评吸质量
BP网络
过拟合
tobacco leaf
chemical constituents
degustation qualities
BP network
over-fitting