【作者】 耿志强；

【导师】 朱群雄；

【作者基本信息】 北京化工大学 ， 控制理论与控制工程， 2005， 博士

【摘要】 随着石油、化工等流程工业的日益大型化、复杂化和现代化,产生了大量的有关物质、能源、工艺设备和操作信息等方面的数据。从大量的生产、管理数据中挖掘深层次的生产知识、最优操作条件和管理模式,即流程工业数据挖掘,是实现流程工业过程的在线监测、故障诊断、安全评估、生产管理、市场分析与预测等的关键技术之一,为流程工业的稳定操作、高效生产提供更有效的决策支持。 流程工业数据挖掘应用的首要任务就是选择和建立有效的适合流程工业数据性质的挖掘算法,粒度数据挖掘技术可以从不同系统级别的角度研究系统,根据实际的应用划分不同的粒度空间,在不同的粒度空间寻找系统的操作模型和相关的约束变量,挖掘流程变量之间的关系与规律,寻求不同粒度空间的局部的或全局的优化方法,有效的解决流程工业诊断和优化的实际问题。 裂解炉系统是乙烯生产过程中的龙头装置,具有一般连续流程工业的各种典型特性。本文以此为主要背景来研究流程工业粒度数据挖掘技术。 针对流程工业数据的高维数、不确定的特点,研究适合处理流程数据的模糊集、粗糙集的粒度数据挖掘理论和方法。针对粗糙度不能完全区分知识粒度的缺点,研究了信息粒度原理与知识粗糙性本质关系,提出了基于粒度熵的信息量化方法和最优属性约简算法;为寻求快速有效的流程工业粒度数据挖掘方法,根据模糊信息粒度矩阵的原理,提出了基于信息粒度矩阵算法的粗糙数据挖掘模型,以及信息粒度压缩矩阵算法和增量式规则获取;提出的粗糙数据挖掘模型和挖掘算法操作方便、易于理解,降低了存储空间,提高了流程工业数据挖掘的运行效率。 针对流程工业数据包含大量噪声、多时标和动态性的特点,进行了以下研究:在数据移动窗内,采用小波变换进行特征提取和滤波,研究了输入训练神经网络的非线性主元分析粒度数据挖掘方法,改进了输入训练神经网络的学习算法;利用小波的多粒度空间分析,提出了自适应多尺度非线性主元分析(MS-NLPCA)的流程工业时序数据的特征提取与非正常工况监测方法;研究了基于流程参数正态更多还原

【Abstract】 With the large-scale, complication and modernization of process industry such as petroleum and chemical engineering, a large number of data about material, product, equipment, process, operation and so on, are generated from manufacture and research of them. Extracting deeply knowledge, optimal operation condition and manageable pattern from large data of production and management, namely, process industrial data mining, is one of the most important technologies to realize online monitoring, fault diagnosis, safety estimation, product management, marketing analysis and prediction and so on of process industry. In addition, it can provide more effective decision support for industrial safety operation and efficient manufacture.The first important task of process industrial data mining is to select and build effective and suitable data mining algorithms to process industrial data. The granularity data mining can research system from different level versions, mine process operating model and relative variables in different granularity space according to practical applications. Moreover, it can discover the relationships and rules among process variables and find the local or global optimization among different granularity space to solve the process diagnosis and operating optimization effectively.The cracking furnace system is the key equipment in ethylene manufacturing process, which has a lot of typical characteristic of general continues petrochemical process. In this paper the process industrial granularity data mining is mainly based on the ethylene cracking furnace system.According to the high dimensions and uncertainty of process industrial data, the fuzzy set and rough set of granularity data mining are studied for process data. To overcome the roughness that it can not completely discern knowledge granularity, the nature relationship between information granularity principle and roughness ofknowledge is studied, and the algorithms of granularity computing and optimal attribute reduct based on granularity entropy are proposed. To pursue fast and efficient granularity data mining algorithm for process industry, rough data mining model based on information granularity matrix algorithm is proposed according to fuzzy information granularity matrix principle. And on the basis of information granularity matrix algorithm, the information compression granularity matrix algorithm and incremental rule acquisition are proposed. The proposed rough data mining model and data mining algorithm are understood easily and operating conveniently, can decrease the store space and improve the efficiency of process industrial data mining.According to process data with noise, multi-frequency and dynamics characteristic, the paper makes several researches as follows: Adopting wavelet transformation based on data moving window to extract feature and filter noise, and then studies granularity data mining on nonlinear principal component analysis (NLPCA) based on inp(?) training neural network (ITNN), moreover improves the learning algorithm of ITN(?) Using the multi-granularity space analysis of wavelet transformation, adapt multi-scale nonlinear PCA (MSNLPCA) granularity data mining method is proposed extract feature and abnormal states monitoring for process industrial time-serial da(?) Using the proposed information granularity matrix algorithm to acquire fuzzy diagnos(?) rules, fuzzy discretization method of continued data is studied on normal distribution o(?) process data and fuzzy-rough information state space model about real-time proces data is built. Based on above researches multi-granularity process monitoring an diagnosis model integrating MSNLPCA-Rough set is proposed.According to the strong coupling and high interrelation of process data, tw granularity data mining methods are studied, namely, dynamical fuzzy clustering-ranking and kernel clustering algorithm. Using PCA to decide the number of fuzzy clustering and ranking the variables in each cluster based on interrelation index, fuzzy clustering-ranking algorithm is used in process alarms optimal management to improve operating efficiency and avoid blindness of dealing with alarms. Dynamic kernel clustering algorithm is used to recognize optimal operating pattern and select the better cracking crude oil to improve the operating ability of ethylene cracking furnace更多还原