【作者】 舒云星；

【导师】 李刚炎；

【作者基本信息】 武汉理工大学 ， 机械设计及理论， 2008， 博士

【摘要】 水泥工业生产中有效的过程监控和质量控制是保证生产安全、提高产品质量和经济效益的关键,如何充分利用来自生产过程监控系统的数据信息,进行水泥生产过程工艺故障诊断和质量预测是生产实际中亟待解决的问题。支持向量机是一种适用于小样本情况下的故障诊断和质量预测的有效方法。本文以水泥烧成系统工艺故障诊断、熟料游离氧化钙含量预测和生料磨出磨生料细度预测为对象,系统深入地研究支持向量机多类分类算法和回归估计算法,构建水泥烧成系统的工艺故障诊断模型和产品质量的预测模型,为指导生产、提高产品质量和生产效率提供一种有效的方法。本文的主要工作包含以下几个方面:1.针对有限样本情况下故障诊断的特点和传统模式识别方法在故障模式分类中面临的困难,以水泥烧成系统的工艺故障为诊断对象,研究了支持向量机用于故障诊断的关键问题,给出了基于支持向量机故障诊断的基本步骤,提出了基于类间分离性测度的支持向量多类分类算法、基于核聚类的支持向量多类分类算法和半模糊超球面的支持向量多类分类算法,建立了相应的多类分类模型,并对影响多类分类器性能的因素进行了分析。2.为了降低分类机和回归机的计算复杂度、提高分类和回归的精度,针对主成分分析在特征提取上存在的不足,通过计算原始特征空间的内积核函数来实现原始特征空间到高维特征空间的非线性映射,在高维特征空间中对映射数据进行主成分分析来得到原始特征数据的非线性主成分,提出并实现了一种基于核主成分分析的非线性特征提取方法。3.为了降低熟料游离氧化钙含量和出磨生料细度预测的计算复杂度,提高预测精度,提出了基于粗糙集约简和核主成分分析相结合的非线性特征提取方法,利用粗糙集的知识约简算法对样本数据进行预处理,删除冗余的属性,减少样本输入空间的维数,然后利用核主成分分析进行非线性特征提取。在此基础上,提出了集成粗糙集、核主成分分析和支持向量机回归算法;通过粗糙集属性约简和核主成分分析特征提取方法的研究,得出了可以利用粗糙集的约简结果来指导选择核主成分的数目。4.在总结分析烧成系统常见工艺故障的产生原因和主要现象的基础上,利用核主成分分析和支持向量机相结合的故障诊断方法,提出了水泥烧成系统工艺故障诊断的两种二叉树支持向量机多类分类算法和半模糊超球面多类分类算法,建立了新型干法水泥烧成系统工艺故障诊断模型,在小样本情况下对水泥烧成系统的工艺故障诊断方法进行了仿真测试;测试结果表明,所提出的算法可以明显地减少训练和测试时间,且分类精度比较理想。5.针对水泥生产过程中质量不能在线测量或有较长时间延迟的问题,通过分析5000t/d水泥生产过程和工艺流程的特点,选取合适的过程工艺参数,提取在线的过程运行数据,基于所研究的算法,提出了几种水泥生产过程质量预测方法,预测结果验证了所提出多种预测方法的有效性和正确性。更多还原

【Abstract】 In the production process of the cement industry,effective process monitoring and quality control are crucial to guarantee production safety and increase product quality and economic gains.How to make full use of the data and information from the monitoring system in the production process to diagnose technological faults in the production process and predict the quality of the products is an urgent problem to be solved.Support Vector Machines(SVM) is an effective method applicable to the fault diagnosis and quality prediction when the samples are small.This study takes the technological fault diagnosis of the cement burning system,the prediction of the free calcium oxide content in the clinker,and the prediction of the fineness of raw mill finished products as its research objectives.We systematically studied in depth the SVM multi-class classification algorithm and the SVR estimation algorithm and constructed a model for the technological fault diagnosis of the cement burning system and a model for the prediction of product quality.In so doing,we hope to provide an effective method of guiding production and increasing product quality and production efficiency.To be brief,the main work of this study is as follows:1.In view of the characteristics of fault diagnosis under the circumstances of finite samples and the difficulties met by the methods based on traditional pattern recognition in the classification of fault patterns,with the technological faults in the cement burning system as the diagnostic objects,we studied the key issues of using SVM in fault diagnosis,presented the basic procedures for a SVM-based fault diagnosis,proposed a SVM multi-class classification algorithm on the basis of the interclass separability measure,a SVM multi-class classification algorithm on the basis of kernel clustering,a SVM multi-class classification algorithm on the basis of the semi-fuzzy hypersphere,and thereafter established their corresponding multi-class classification models,and then we analyzed the factors influencing the performance of the multi-class classifiers.2.In order to lower the computation complexity of the classifiers and the SVRs and increase the precision of classification and regression,considering the shortcomings of the principal component analysis(PCA) in feature extraction,through the computation of the inner product kernel function in the original feature space,we realized the nonlinear mapping from the original feature space to the high-dimension feature space,and in the high-dimension feature space,the nonlinear principal component of the original feature data are obtained through the PeA of the mapping data.Thus,we proposed and actualized an effective nonlinear feature extraction method on the basis of the kernel principal component analysis (KPCA).3.To lower the free calcium oxide content in the clinker,to reduce the computation complexity of predicting the fineness of raw mill finished products,and to increase the prediction precision,we proposed a nonlinear feature extraction method on the basis of the combination of rough sets(RS) reduction and KPCA,conducted preliminary treatments to the sample data by using the RS knowledge reduction algorithm,deleted the redundant attributes, reduced the number of the dimensions of the sample’s input space,and then conducted the nonlinear feature extraction by using KPCA.On the basis of these procedures,we proposed an integrated RS-KPCA-SVM regression algorithm.By means of RS attribute reduction and the study of the KPCA feature extraction,we obtained the number to be used in guiding the selection of the kernel principal components by using the RS reduction results.4.We analyzed the causes to and the main phenomena of the common technological faults of the cement burning system.By using the fault-diagnosing method which combines the KPCA and the SVM,we proposed two kinds of binary tree SVM multi-class classification algorithm and the semi-fuzzy hypersphere multi-class classification algorithm for the diagnosis of the technological faults of the cement burning system,and we established a new fault diagnosis model for the new-style dry-process cement burning system.After that,we conducted simulation studies of the method of diagnosing the technological faults for the cement burning system when small samples were adopted.The testing results indicate that the proposed algorithms can distinctly reduce the time required for training and testing,and that the classification precision is relatively desirable.5.In view of the problem that it is impossible to monitor the cement quality online in the cement production process or there is a relatively long period of lagging,we analyzed the characteristics of the production process and the technological process of the 5000t/d cement, selected appropriate process technological parameters,extracted online process operational data,and on the basis of the proposed algorithms in this study,we proposed several methods of predicting the product quality in the process of cement production,and finally the prediction results verified the effectiveness and validity of the proposed multiple prediction methods.更多还原