基于DE_BPANN的单料烟烟气指标预测方法

DE_BPANN-Based Method for Predicting Smoke Components of Cigarette Made of Single Grade Tobacco

为了分析和挖掘单料烟的化学成分和烟气指标之间的关系,采用基于差分进化的人工神经网络用于预测单料烟的烟气指标。建立了单隐层的人工神经网络,并针对基于误差反向传播的人工神经网络的缺陷,将差分进化算法应用于神经网络的训练过程。该预测方法的主要思想是结合了人工神经网络的局部搜索能力和差分进化的全局搜索能力。通过采用某烟草公司的单料烟烟叶、烟气数据,将单料烟的7种常规化学成分作为预测模型的输入变量,将主流烟气中的焦油、烟气烟碱和CO作为预测模型的输出变量,建立了人工神经网络预测模型。实验结果表明:焦油、烟气烟碱和CO的预测均方差达到了较好水平,与传统神经网络相比,分别提高了27%,10%和26%,表明该方法的预测准确度更高。

In order to analyze the relation between the chemical components in tobacco and the smoke components of cigarette, an artificial neural network based on differential evolution was used to predict the routine smoke components of cigarette made of single grade tobacco. An artificial neural network with single hidden layer was established. A differential evolution algorithm was used in the training process to avoid the shortcomings of artificial neural network based on error back propagation. This prediction method combined the local searching ability of artificial neural network and the global searching ability of differential evolution. On the basis of historical data of single grade tobacco and smoke of cigarettes made of single grade tobacco saved by a tobacco company, an artificial neural network prediction model was established, seven routine chemical components in the single grade tobacco were taken as input variables and the deliveries of tar, nicotine, carbon monoxide in mainstream cigarette smoke as output variables. The results showed that the root mean square errors of prediction of tar, nicotine, carbon monoxide reached a higher level and increased by 27%,10% and 26%, respectively, comparing with the traditional neural network.