基于关键性能指标的数据驱动故障检测方法研究
发布时间：2019-06-21 00:02
【摘要】：过程监控是保障系统安全性和可靠性的重要研究课题。在过去的几十年中,得益于控制理论体系的不断完善,基于解析模型的过程监控方法获得了长足的发展并产生出了大量的研究成果。然而,诸如化工、冶金、生物制药等工业系统,其内部组成的复杂性和底层反应机理的未知性使得建立其精准的数学模型是非常困难甚至是不可能的。因此,基于解析模型的方法难以在此类系统中有效地推广和应用。值得注意的是,这类工业系统通常能够产生出大量的离线记录数据和在线测量数据。如何利用这些丰富的数据对系统进行有效地监测和控制引起了学术界和工业界的持续关注,这也促进了数据驱动过程监控技术的快速发展。在过去的二十年中,数据驱动过程监控方法的主要任务是快速准确地检测出工业过程中发生的故障,并对故障进行辨识、隔离以及恢复来确保系统的稳定运行。然而,近年来的研究成果和工业实践的反馈信息表明并不是所有的过程故障都会影响到系统最终的产品质量,反而忽略此类故障的报警可以显著减少不必要的停产和检修时间,从而极大地提高系统的生产效率同时降低维护成本。出于这一动机,基于关键性能指标(KPI,Key Performance Indicator)的故障检测方法成为工业界的迫切需求,同时也成为近五年来学术界研究的热门课题。鉴于此,针对线性和非线性静态系统,本文将采用多元统计分析等数据驱动技术,深入开展基于KPI的故障检测方法研究。第一,总结数据驱动故障检测方法的发展历程和研究现状,给出基于KPI的故障检测问题的数学描述,并指出在基于KPI的故障检测领域已有的研究成果的不足之处:1)目前针对线性系统提出的方法大多数是在偏最小二乘(PLS,Partial Least Squares)的基础上通过后处理方式实现的,然而PLS对建模数据中的离群点和丢失点等异常数据非常敏感,导致少量异常数据即可严重影响PLS后处理方法的KPI预测性能,另外,已有的此类方法在故障判定逻辑方面依然比较复杂;2)受限于PLS的分解特性,PLS后处理方法的性能通常不稳定,当故障强度增大时这些方法的故障误报率会显著升高;3)目前针对非线性系统提出的方法在故障检测性能和故障判定逻辑等方面都存在明显缺陷;4)目前已有的线性和非线性方法在算法设计上都是独立进行的,具有一致分解结构的综合设计方法尚未引起注意,而这类方法对于简化基于KPI的故障检测系统的设计步骤是非常重要的。针对这些问题,本文将逐一提出相应的解决方案。第二,针对已有的PLS后处理方法对异常数据敏感以及故障判定逻辑复杂的问题,本文通过引入鲁棒PLS算法和期望最大化算法分别降低建模数据中的离群点和丢失点等异常数据对PLS模型的影响,然后利用奇异值分解(SVD,Singular Value Decomposition)将过程变量空间正交分解为与KPI相关的子空间和与KPI无关的子空间,进而提出一种鲁棒KPI预测和基于KPI的线性故障检测方法。仿真结果表明,所提出的方法在KPI预测方面明显优于已有的PLS后处理方法,并且在故障检测性能和故障判定逻辑等方面都有明显的提升。第三,针对已有的PLS后处理方法当故障强度增大时故障误报率升高的问题,本文通过引入数据预处理的思想,提出一类基于数据预处理和PLS后处理相结合的增强型的基于KPI的线性故障检测方法。该类方法首先通过数据预处理去除过程变量空间中与KPI无关的成分,然后再利用PLS后处理将剩余过程变量空间正交分解为与KPI相关的子空间和与KPI无关的子空间,从而实现过程变量空间的彻底分解。仿真结果表明,所提出的方法明显改善了PLS后处理方法性能不稳定的问题,在故障强度非常大时依然可以保持极低的故障误报率。第四,针对已有的基于KPI的非线性故障检测方法在故障检测性能和故障判定逻辑等方面存在的缺陷,本文提出两种不同实现方式的新方法。首先,本文利用核偏最小二乘(KPLS,Kernel Partial Least Squares)对非线性过程建模,在此基础上构建核空间和KPI之间的线性关系,并通过SVD将核空间正交分解为与KPI相关的子空间和与KPI无关的子空间,进而提出一种基于KPI的非线性故障检测方法。另外,本文充分借鉴非线性模型逼近领域的研究成果,利用统计学习方法将非线性过程等效为若干局部线性模型,并通过在局部模型设计基于KPI的线性故障检测方法进而实现全局的基于KPI的非线性故障检测方法。仿真结果表明,所提出的两种非线性方法在故障检测性能和故障判定逻辑等方面都明显优于已有的方法。第五,针对基于KPI的线性和非线性故障检测方法的综合设计问题,本文提出一种新颖的解决方案。首先,本文在主成分分析(PCA,Principle Component Analysis)的基础上,利用主成分回归(PCR,Principle Component Regression)的建模思想提出一种能够将过程变量空间分解为与KPI相关的子空间和与KPI无关的子空间的算法,进而提出一种新的基于KPI的线性故障检测方法;依据完全相同的分解结构,本文通过建立核主成分分析(KPCA,Kernel Principle Component Analysis)模型,并利用核主成分回归(KPCR,Kernel Principle Component Regression)的建模思想提出一种能够将特征空间分解为与KPI相关的子空间和与KPI无关的子空间的算法,进而提出一种新的基于KPI的非线性故障检测方法。仿真结果表明所提出的线性方法在故障检测性能以及故障判定逻辑等方面都明显优于前面提出的线性方法和已有的PLS后处理方法;而所提出的非线性方法在故障检测性能和故障判定逻辑方面和前面提出的非线性方法基本相同,但明显优于已有的非线性方法。更重要的是,一致的分解结构极大地简化了基于KPI的故障检测系统的设计步骤,同时也便于工程应用人员对算法的理解,因此更加有利于所提出的方法在实际系统中的推广和应用。
[Abstract]:Process monitoring is an important research subject for the safety and reliability of the system. In the past few decades, thanks to the continuous improvement of the control theory system, the process monitoring method based on the analytical model has made great development and has produced a great deal of research results. However, industrial systems such as chemical, metallurgical, biopharmaceutical and the like, the complexity of its internal composition and the unknown nature of the underlying reaction mechanism make it very difficult or even impossible to establish its precise mathematical model. Therefore, the method based on the analytical model is difficult to be effectively promoted and applied in such systems. It is to be noted that such industrial systems are generally capable of producing a large amount of off-line recording data and on-line measurement data. How to use these abundant data to effectively monitor and control the system is the continuous concern of the academic and industry, which also promotes the rapid development of the data-driven process monitoring technology. In the past two decades, the main task of the data-driven process monitoring method is to quickly and accurately detect the faults occurring in the industrial process and to identify, isolate and recover the faults to ensure the stable operation of the system. However, the research results in recent years and the feedback information of industrial practice indicate that not all process failures will affect the final product quality of the system, but the alarm that ignores such failures can significantly reduce the unnecessary shutdown and maintenance time, So that the production efficiency of the system is greatly improved, and the maintenance cost is reduced. For this motivation, the failure detection method based on KPI and Key Performance Indicator becomes the urgent need of industry, and has also become a hot topic for academic research in the past five years. In view of the linear and non-linear static system, a multi-element statistical analysis and other data driving technology will be used in this paper to carry out the research on the method of fault detection based on KPI. First, the development history and the research status of the data-driven fault detection method are summarized, the mathematical description of the fault detection problem based on the KPI is given, and the shortcomings of the research results in the field of fault detection based on the KPI are pointed out: 1) Most of the methods currently proposed for linear systems are implemented by post-processing on the basis of partial least squares (PLS, Partial Least Squares), but the PLS is very sensitive to abnormal data such as outliers and missing points in the modeling data, leading to a small amount of abnormal data can seriously affect the KPI prediction performance of the PLS post-processing method, and in addition, the existing method is still more complex in the fault determination logic;2) is limited by the decomposition characteristic of the PLS, and the performance of the PLS post-processing method is generally not stable, when the failure intensity is increased, the fault error rate of the methods can be obviously increased;3) the method proposed by the non-linear system has obvious defects in the aspects of fault detection performance and fault judgment logic, and the like; 4) The existing linear and non-linear method is independently carried out in the design of the algorithm, and the integrated design method with the consistent decomposition structure has not yet attracted attention, and the method is very important for simplifying the design steps of the KPI-based fault detection system. In view of these problems, the corresponding solutions will be presented one by one. Secondly, according to the problem that the existing PLS post-processing method is sensitive to the abnormal data and the fault decision logic is complex, the influence of the outliers and the missing points in the modeling data on the PLS model is reduced by introducing the robust PLS algorithm and the expectation maximization algorithm, respectively. Then, singular value decomposition (SVD) is used to decompose the process variable space into sub-space related to the KPI and sub-space independent of the KPI, and then a robust KPI prediction and a KPI-based linear fault detection method is proposed. The simulation results show that the proposed method is better than the existing PLS after-processing method in the KPI prediction, and has obvious improvement in fault detection performance and fault decision logic. Thirdly, according to the existing PLS post-processing method, when the failure intensity is increased, the problem of the increase of the fault is raised. In this paper, by introducing the idea of data pre-processing, this paper puts forward a class of enhanced KPI-based linear fault detection method based on data pre-processing and PLS post-processing. The method comprises the following steps of: firstly, the component irrelevant to the KPI in the process variable space is removed through the data pretreatment, and then the residual process variable space is decomposed into the sub-space related to the KPI and the sub-space irrelevant to the KPI by using the PLS after-processing, so that the complete decomposition of the process variable space is realized. The simulation results show that the proposed method can obviously improve the performance of the PLS after-treatment method, and can still maintain a very low fault error when the fault strength is very large. Fourth, in view of the existing defects in the fault detection performance and fault determination logic of the existing KPI-based non-linear fault detection method, two new methods are proposed in this paper. Firstly, the linear relationship between the kernel space and the KPI is built on the basis of the kernel partial least square (KPLS) and the kernel partial least square (KPLS), and the kernel space is decomposed into sub-space related to the KPI and the sub-space irrelevant to the KPI by the SVD. And then a non-linear fault detection method based on the KPI is proposed. In addition, this paper fully uses the research results in the nonlinear model approximation, and uses the statistical learning method to convert the non-linear process into a number of local linear models. And a global KPI-based non-linear fault detection method is realized by designing a KPI-based linear fault detection method in a local model. The simulation results show that the two nonlinear methods are superior to the existing methods in the aspects of fault detection performance and fault decision logic. Fifth, in view of the comprehensive design of the linear and non-linear fault detection method based on the KPI, this paper presents a novel solution. First, on the basis of principal component analysis (PCA) and Principal Component Analysis (PCA), a method of using principal component regression (PCR) to decompose process variable space into sub-space related to KPI and sub-space independent of KPI is proposed. In this paper, a new linear fault detection method based on KPI is proposed. According to the exact same decomposition structure, the kernel principal component analysis (KPCA, Kernel Principal Component Analysis) model is established, and the kernel principal component regression (KPCR) is used. Kernel Principal Component Regression proposed a new method of non-linear fault detection based on KPI, which can decompose the feature space into sub-space related to the KPI and sub-space independent of the KPI. The simulation results show that the proposed linear method is superior to the linear method and the existing PLS after-processing method in the aspects of fault detection performance and fault determination logic. The proposed non-linear method is basically the same as the non-linear method proposed in the fault detection performance and fault decision logic, but it is better than the existing non-linear method. More importantly, the consistent decomposition structure greatly simplifies the design steps of the KPI-based fault detection system, and is also convenient for engineering application personnel to understand the algorithm, so that the proposed method is more beneficial to the popularization and application of the proposed method in the actual system.
【学位授予单位】：哈尔滨工业大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP277
本文编号：2503626
[Abstract]:Process monitoring is an important research subject for the safety and reliability of the system. In the past few decades, thanks to the continuous improvement of the control theory system, the process monitoring method based on the analytical model has made great development and has produced a great deal of research results. However, industrial systems such as chemical, metallurgical, biopharmaceutical and the like, the complexity of its internal composition and the unknown nature of the underlying reaction mechanism make it very difficult or even impossible to establish its precise mathematical model. Therefore, the method based on the analytical model is difficult to be effectively promoted and applied in such systems. It is to be noted that such industrial systems are generally capable of producing a large amount of off-line recording data and on-line measurement data. How to use these abundant data to effectively monitor and control the system is the continuous concern of the academic and industry, which also promotes the rapid development of the data-driven process monitoring technology. In the past two decades, the main task of the data-driven process monitoring method is to quickly and accurately detect the faults occurring in the industrial process and to identify, isolate and recover the faults to ensure the stable operation of the system. However, the research results in recent years and the feedback information of industrial practice indicate that not all process failures will affect the final product quality of the system, but the alarm that ignores such failures can significantly reduce the unnecessary shutdown and maintenance time, So that the production efficiency of the system is greatly improved, and the maintenance cost is reduced. For this motivation, the failure detection method based on KPI and Key Performance Indicator becomes the urgent need of industry, and has also become a hot topic for academic research in the past five years. In view of the linear and non-linear static system, a multi-element statistical analysis and other data driving technology will be used in this paper to carry out the research on the method of fault detection based on KPI. First, the development history and the research status of the data-driven fault detection method are summarized, the mathematical description of the fault detection problem based on the KPI is given, and the shortcomings of the research results in the field of fault detection based on the KPI are pointed out: 1) Most of the methods currently proposed for linear systems are implemented by post-processing on the basis of partial least squares (PLS, Partial Least Squares), but the PLS is very sensitive to abnormal data such as outliers and missing points in the modeling data, leading to a small amount of abnormal data can seriously affect the KPI prediction performance of the PLS post-processing method, and in addition, the existing method is still more complex in the fault determination logic;2) is limited by the decomposition characteristic of the PLS, and the performance of the PLS post-processing method is generally not stable, when the failure intensity is increased, the fault error rate of the methods can be obviously increased;3) the method proposed by the non-linear system has obvious defects in the aspects of fault detection performance and fault judgment logic, and the like; 4) The existing linear and non-linear method is independently carried out in the design of the algorithm, and the integrated design method with the consistent decomposition structure has not yet attracted attention, and the method is very important for simplifying the design steps of the KPI-based fault detection system. In view of these problems, the corresponding solutions will be presented one by one. Secondly, according to the problem that the existing PLS post-processing method is sensitive to the abnormal data and the fault decision logic is complex, the influence of the outliers and the missing points in the modeling data on the PLS model is reduced by introducing the robust PLS algorithm and the expectation maximization algorithm, respectively. Then, singular value decomposition (SVD) is used to decompose the process variable space into sub-space related to the KPI and sub-space independent of the KPI, and then a robust KPI prediction and a KPI-based linear fault detection method is proposed. The simulation results show that the proposed method is better than the existing PLS after-processing method in the KPI prediction, and has obvious improvement in fault detection performance and fault decision logic. Thirdly, according to the existing PLS post-processing method, when the failure intensity is increased, the problem of the increase of the fault is raised. In this paper, by introducing the idea of data pre-processing, this paper puts forward a class of enhanced KPI-based linear fault detection method based on data pre-processing and PLS post-processing. The method comprises the following steps of: firstly, the component irrelevant to the KPI in the process variable space is removed through the data pretreatment, and then the residual process variable space is decomposed into the sub-space related to the KPI and the sub-space irrelevant to the KPI by using the PLS after-processing, so that the complete decomposition of the process variable space is realized. The simulation results show that the proposed method can obviously improve the performance of the PLS after-treatment method, and can still maintain a very low fault error when the fault strength is very large. Fourth, in view of the existing defects in the fault detection performance and fault determination logic of the existing KPI-based non-linear fault detection method, two new methods are proposed in this paper. Firstly, the linear relationship between the kernel space and the KPI is built on the basis of the kernel partial least square (KPLS) and the kernel partial least square (KPLS), and the kernel space is decomposed into sub-space related to the KPI and the sub-space irrelevant to the KPI by the SVD. And then a non-linear fault detection method based on the KPI is proposed. In addition, this paper fully uses the research results in the nonlinear model approximation, and uses the statistical learning method to convert the non-linear process into a number of local linear models. And a global KPI-based non-linear fault detection method is realized by designing a KPI-based linear fault detection method in a local model. The simulation results show that the two nonlinear methods are superior to the existing methods in the aspects of fault detection performance and fault decision logic. Fifth, in view of the comprehensive design of the linear and non-linear fault detection method based on the KPI, this paper presents a novel solution. First, on the basis of principal component analysis (PCA) and Principal Component Analysis (PCA), a method of using principal component regression (PCR) to decompose process variable space into sub-space related to KPI and sub-space independent of KPI is proposed. In this paper, a new linear fault detection method based on KPI is proposed. According to the exact same decomposition structure, the kernel principal component analysis (KPCA, Kernel Principal Component Analysis) model is established, and the kernel principal component regression (KPCR) is used. Kernel Principal Component Regression proposed a new method of non-linear fault detection based on KPI, which can decompose the feature space into sub-space related to the KPI and sub-space independent of the KPI. The simulation results show that the proposed linear method is superior to the linear method and the existing PLS after-processing method in the aspects of fault detection performance and fault determination logic. The proposed non-linear method is basically the same as the non-linear method proposed in the fault detection performance and fault decision logic, but it is better than the existing non-linear method. More importantly, the consistent decomposition structure greatly simplifies the design steps of the KPI-based fault detection system, and is also convenient for engineering application personnel to understand the algorithm, so that the proposed method is more beneficial to the popularization and application of the proposed method in the actual system.
【学位授予单位】：哈尔滨工业大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP277
【参考文献】
相关期刊论文 前1条
1 王丽;侍洪波;;采用改进核偏最小二乘法的非线性化工过程故障检测(英文)[J];Chinese Journal of Chemical Engineering;2014年06期
，本文编号：2503626
本文链接：https://www.wllwen.com/shoufeilunwen/xxkjbs/2503626.html
