【作者】 陈国金；

【导师】 钱积新； 梁军；

【作者基本信息】 浙江大学 ， 控制科学与工程， 2004， 博士

【摘要】 人类对技术的追求永无止境，保障生产安全和减小产品质量波动一直是工业生产过程的两大追求目标。在工业生产过程中，及时有效地发现、检测和修复过程故障是提供性能优良、品质一致产品的先决条件，这也是进行工业过程监控的目的和动机。工业过程监控技术从单变量过程监控算起，已有近七十多年的历史，但基于多变量信息的过程监控技术至今也不过十几年的时间。在此领域虽已获得了大量成果，但研究基本上是在过程检测数据服从多元正态分布和独立同分布的两个假设下进行的。然而，在实际工业过程中，过程信息是非常复杂的，所提取的过程特征信息服从何种分布很难确定。本文的研究正是着眼于克服这两大假设条件，使过程监控技术能更好地适用于实际工业生产过程而进行的。 为此，本文采用了主元分析和盲源信号分析这两类过程数据驱动的方法，作为研究的主要数学工具。主元分析方法不仅作为一种过程特征的提取方法(在过程信息服从多元正态分布的情况下)，而且也作为一种过程数据降维的主要工具(在过程盲源信号提取的情况下)；盲源信号分析是从信息论的角度，从过程信息中提取出尽可能独立的过程特征信号或过程原始信源信号，它具有比主元分析更好的刻画过程运行特征的性能。本文的主要内容概括如下： 1) 介绍了过程监控的基本概念和内容，并指出了流程工业中基于子空间特征信息抽取进行过程监控的优越性。此外，还简要地描述了主元分析方法和盲源信号分析方法及它们在过程监控中的应用。 2) 由于过程信息并非均服从正态分布，提出了一种基于支持向量分类器主元分析方法的过程监控方法，仿真表明提高了过程监控的性能。 3) 根据过程信息能够用若干“尽可能独立”的过程特征信号进行描述的原理，提出了一种基于独立成分分析的过程监控方法。仿真研究表明，这种方法是有效的。 4) 通常过程信息或多或少地受到噪声的污染，提出了一种先提取过程盲源信号，随后用小波变换进行去噪的传统过程监控方法。仿真研究结果表明，这些处理方式能够提高过程的监控性能。 5) 噪声往往会导致过程特征提取的失效。为了提高盲源信号描述过程的能力，提出了首先利用小波变换去噪，然后提取盲源信号进行过程监控的方法，对过程监控仿真的结果表明，这种方法比基于传统盲源信号分离具有更好的监控性能。加声，认掌摘要和目录6)针对工业过程中过程信息的复杂性，采用了多元统计投影方法(独立成分分 析方法和主元分析方法)，先后从过程信息中提取非正态分布特征信号和正 态分布特征信号，然后用这些过程特征去刻画过程、监控过程性能和进行故 障诊断。该方法避免了传统多元统计过程的正态分布假设，提高了多元投影 方法进行过程监控和故障诊断的适用性和可靠性。7)过程信息并非均是独立同分布，对于很多过程，过程信息往往存在着一定的 时间结构，有鉴于此，提出了利用过程信息时间结构的过程监控方法，仿真 研究表明，这种方法具有比传统ICA的方法更好的性能。8)鉴于在过程中，过程信息的平稳性并不确定，提出了一种不考虑过程平稳性 能的过程监控方法，仿真表明该方法比基于传统ICA的过程监控方法具有更 少的误报率和漏报率，而比基于MSPC的过程监控方法具有更少的误报率， 从而说明该方法的有效性。9)随着许多工业过程转向半间歇和间歇操作，对这些过程的监控技术变得越来 越重要。为此，提出了基于多元统计信号处理的过程监控技术。这种方法将 过程信息空间划分为由盲源信号描述的信号子空间、主元描述的信号子空间 和残差信号子空间，对过程监控仿真的结果表明，这种方法比传统MSPC更 好的故障分离性和定位性，从而也更为有效。 论文的内容安排顺序是与研究过程的逐步深入和完善相适应的。最后，对以多元过程监控技术的研究方向进行了一些有益的探索。更多还原

【Abstract】 Human won’t be satisfied with obtaining knowledge. Similarly, the safety of production procedure and consistency of product quality are always two goals of the process industry. It is only timely and effectively finding, detecting and restoring fault in process that can create conditions for providing products with good performance and consistent quality, which is also the object and motivation of process monitoring. Industrial process monitoring has developed for seventy years from first appearance of quality control diagram by Shewhart, however, the research for multivariate process monitoring is only longer than ten years. Lots of research results are obtained in this field, though which are always based on two assumptions: One is that process variables are subjected to multivariate normal distribution; the other is that samples are subjected to independent and identical distribution (iid). In fact, the process information in real process is complex and the probability distribution of extracted features is indeterminate. Of course, it is often effective to apply conventional multivariate statistical process control (MSPC) to the process whose process variables are subjected (or approximatively subjected) to multivariate normal distribution. For the process with information subjected to nonnormal distribution, a more effective signal processing method (blind source analysis, BSA) is applied to extract features of process. The research results of this dissertation indicate that process monitoring methods based on BSA will improve the monitoring performance of process and enlarge the range of the application.Two primary mathematical tools used in this dissertation are principal component analysis (PCA) and blind signal analysis (BSA), which are both data-driven methods. PCA is not only used as feature extracting method (where process variables are subjected to multivariate normal distribution), but also as a tool for dimension reduction; BSA is used to extract independent features or process blind source signals from process information in information theory sense, which is more effective than PCA in describing the process.The main contributions of this dissertation are as follows:1) The elementary concepts and scope of process monitoring are introduced. Moreover, PCA and BSA with their application in process monitoring are simpledescribed2)Due to the fact that process information isn’t always subjected to multivariate normal distribution, a process monitoring method based on PCA with support vector classifier is provided, which improves the monitoring performance.3)Based on the idea that the process information is driven by a few of components as independent as possible, a novel process monitoring method is provided whose effectiveness is verified by the research results.4)In order to reduce the influence of noises an improved conventional process monitoring method is present, which includes as following steps: firstly extract blind source signals from process information, then denoise each blind source signal with wavelet transform, finally build process statistics to monitor process. The research results verify that it can improve the monitoring performance of process.5)Due to the failure of extracting process features by noise, a process monitoring method based on blind source signal separation with denoising information by wavelet transform is provided. The results of process monitoring indicate that this method is more effective than the process monitoring method based on conventional blind source signal separation.6)Due to the complexity of process information, a process monitoring method which applies independent component analysis and principal component analysis to extract nonnormal distributed process features and normal distributed process features is presented, which avoids the assumption that process information is subjected to multivariate normal distribution. The results of process simulation verify the effectiveness of the presented method.7)Lots of更多还原