【作者】 张鑫；

【导师】 罗汉；

【作者基本信息】 湖南大学 ， 概率论与数理统计， 2018， 硕士

【摘要】 卷烟材辅料的上机适应性,是影响卷烟生产设备的有效作业率、产品质量、生产物耗成本的关键所在,已经成为卷烟生产的重点关注问题.实际中影响卷烟产品质量的上机适应性问题通常以机台故障现象的形式反映出来.将不同类型的机台故障现象进行分类,并对由卷烟材辅料上机适应性而造成卷烟产品的残次品剔除情况进行预测,进一步对卷烟材辅料的关键指标进行参数设计,从而有效提高生产效率和控制卷烟产品质量,具有重要的理论意义和实际应用价值.支持向量机(Support Vector Machines,SVM),作为一种机器学习方法,对解决实际工业问题中高维数、非线性数据的分类预测问题,在训练学习能力和泛化能力方面相较于其他大数据分析及挖掘技术方法有着较强的优势.本文在参与H中烟公司国家局项目《基于大数据分析的卷接包工艺适应性研究与应用》研究的同时,结合卷烟“包装工序辅料上机适应性诊断系统”的生产数据和仿真数据,基于SVM理论开展以下三个方面的主要研究工作:一,针对卷烟商标纸的上机适应性问题对应的机台故障现象的分类问题,基于半监督SVM的两种分类学习算法(自我训练算法,协同训练算法),训练得到机台故障现象关于卷烟商标纸关键指标的半监督SVM分类器群.结合数据验证两种算法均具有较好的学习能力,并与传统二阶聚类方法进行比较.讨论训练数据的类标记个数对分类精度的影响.二,针对由卷烟商标纸上机适应性问题而造成的烟包剔除率的预测问题,基于?-不敏感损失函数SVR理论,训练得到机台烟包剔除率关于商标纸关键指标体系的SVR预测模型.通过数值结果比较基于?-不敏感损失函数SVR与传统回归对测试数据的预测准确度.讨论对预测模型检验和训练数据处理的统计方法.三,针对上机适应性的过程质量控制和卷烟商标纸关键指标的参数设计问题,利用核主成分分析方法(KPCA)进行二次关键特征提取,基于多元统计的预测控制理论和核技巧思想,在上述研究结果的基础上实现了卷烟商标纸摩擦系数和压痕挺力的参数设计,进而模拟烟包剔除数和剔除率的波动趋势.目前参与的H中烟公司研究项目已成功结题并被推荐申报中国烟草总公司科学技术奖,主要研究结果从今年起在H中烟公司推广应用,卷烟产品剔除率较往年相比由0.051降低为0.032(包/万支),降低了37.3%,大幅提高了卷烟设备的有效作业率和卷烟材辅料的利用率.更多还原

【Abstract】 The machine applicability of cigarette auxiliary materials is the key to the effective operation rate of cigarette production equipment,cigar ette product quality and materials cost of production,which has become the focus of cigarette production.The practical machine applicability problems affecting the quality of cigarette products are usually reflected in the form of machine failure phenomen on.Categorizing different types of machine failure phenomenon,forecasting and controlling the elimination of cigarette products caused by the machine adaptability problems of cigarette auxiliary materials,parametric designing key indexes of cigarette auxiliary materials,are the prerequisite for effective measures to improve production efficiency,and have important theoretical research value and practical significance.SVM(Support Vector those,SVM),as a new machine learning method,to solve the practical classification and prediction of industrial high dimension and nonlinear data,has a strong advantage over other big data analysis and mining techniques in training learning ability and generalization ability.Participated in the national bureau project《The research and application of the technology adaptability of cigarette packaging based on big data analysis 》of H company,combined the production data and simulation experimental data of the big data platform of the cigarette dynamic diagnosis system(MES),this paper carries out the following three main research work based on the SVM theory method:Firstly,focusing on the classification of cigarette auxiliary materials of machine adaptability,two classification algorithms(Self-training algorithm and Co-training algorithm)based on semi-supervised SVM are proposed in this paper.The semi-supervised SVM classifier group of machine adaptability under the key index system of cigarette trademark paper is obtained.Combined with data validation and compared with traditional second-order clustering method,the two classification algorithms have good learning ability.The influence of the number of training data on the classification accuracy is discussed.Secondly,a support vector regression machine based on ?-insensitive loss function is proposed for the prediction of cigarette packet rejection rate caused by the machine adaptability problems of cigarette label paper.The SVR prediction model of the key index system of label paper is obtained by training.The predictive accuracy of test data of SVR based on ?-insensitivity loss function and of traditional multiple regression is compared with numerical results.In addition,statistical methods for predicting model test and training data processing are also discussed.Thirdly,based on SVR,a predictive control theory is proposed to solve the problem of parameter design of key indexes of cigarette label paper and quality control of machine adaptability.The multivariate nonlinear model is transformed into a linear regression model by combining kernel trick.Using kernel principal component analysis(KPCA)to secondary extraction of key features,the parameter design of friction coefficient and indentation force of cigarette label paper is realized,and the fluctuation of the cigarette packet rejection number caused by the machine adaptability problems of cigarette label paper is predicted.Currently,the participating research project has been successfully finished and recommended to declare the science and technology award of China national tobacco corporation.The main research results have been applied in the production of H company since this year.Compared with previous years,the rejection rate of cigarette products decreased from 0.051 to 0.032,which decreased by 37.3%.The effective operation rate of cigarette equipment and the utilization ratio of cigarette materials are greatly improved.更多还原