基于选择稳定性优化的卷烟吸阻快速预测  

Rapid prediction of cigarette pressure drop based on selective stability optimization

作　　者：胡素霞 陈琳 王娜 章桥新[2] 龚静雯[2] 刘思琪 程占刚 赵诗棋 HU Suxia;CHEN Lin;WANG Na;ZHANG Qiaoxin;GONG Jingwen;LIU Siqi;CHEN Zhangang;ZHAO Shiqi(Technology Center,China Tobacco Hubei Industrial Co.Ltd.,Wuhan 430000;School of Mechanical and Electronic Engineering,Wuhan University of Technology,Wuhan 430070)

机构地区：[1]湖北中烟工业有限公司技术中心,武汉430000 [2]武汉理工大学机电工程学院,武汉430070

出　　处：《安徽农业大学学报》2023年第2期372-378,共7页Journal of Anhui Agricultural University

基　　金：湖北中烟工业有限责任公司科技项目(2020JSCL3JS2B029)资助。

摘　　要：卷烟吸阻是卷烟设计制造中的核心指标。因涉及影响因素多且具有复杂的非线性特性,无论是基于多孔介质流体力学模型还是基于大量工程实践的经验模型,均无法定量指导设计与生产,至今卷烟吸阻仍以实验测试数据为评价依据。针对卷烟生产过程中产生的大量检测数据及数据的复杂多源和不断更迭的特性,提出了一种利用生产历史积累数据,通过K均值聚类算法清洗数据消除样本差异,结合自适应套索方法对输入变量进行降维处理和辅助变量选择,并利用选择稳定性评估对过程进行一致性约束,在多源数据和滚动过程一致选择出与吸阻原理模型匹配的关键影响指标,并将其作为径向基函数神经网络(RBFNN,radical basis function netural network)的输入,建立吸阻的推理预测模型。经验证,预测模型的均方误差为0.004,相对误差率控制在3%以内,实现了生产场景下的吸阻快速预测。Pressure drop is the key indicator for cigarettes design and manufacture.As pressure drop is affected by many factors and exists complex nonlinear characteristics,neither the fluid mechanics model of porous media nor the empirical model can guide the design and production quantitatively based on a large number of engineering practices.Hence,pressure is still evaluated on the basis of experimental test data.Aiming at the characteristics of a large number of test data generated in the process of cigarette manufacture and the complex multi-source and updating,a method was proposed to eliminate the sample differences and using K-means to clean the data.The approach applied adaptive lasso to reduce the dimension and select key variables from input and used stability indicator to evaluate the process of variables screening.Afterwards,the key variables were used as inputs to RBFNN(radical basis function netural network).Through the training of RBFNN,the study created a pressure drop prediction model.After example validating,the mean square error of the prediction model was 0.004,and the relative error rate was controlled within 3%,realizing the rapid prediction of the pressure drop in production scenarios.

关 键 词：吸阻 预测 径向基神经网络 选择稳定性 自适应套索 

分 类 号：TS452[农业科学—烟草工业]