流程工业过程故障检测的特征提取方法研究
发布时间：2018-08-08 17:44
【摘要】：机器的出现使生产效率和生产安全得到了提升,但机械设备的故障同样也会导致财产损失和人员伤亡。随着工业过程不断向着大型化、复杂化的趋势发展,传统的依靠对每个变量实施独立监控的方法不再可行,并且建立精确的解析模型也越来越困难。与此同时,随着集散控制系统的广泛应用,从多个角度反映系统运行状态的大量数据被记录了下来。于是,为了进一步提高过程监控的效率,数据驱动的故障检测方法开始兴起。 实际工业过程通常有多种复杂的特性,如多工况、时变、复杂数据分布等,而在最初进行研究时,人们会通过设定一些假设条件如过程运行在单一稳定工况下,变量服从高斯分布等,以此来简化研究问题。但是随着研究工作的深入开展,就需要克服这些假设,研究适用性更强的故障检测方法。本文旨在针对工业过程中的复杂数据分布、多工况和时变问题,通过利用数据局部和全局结构信息,提出新的特征提取方法,进一步提高故障检测模型的精度。具体内容如下： (1)针对单工况下训练数据具有复杂数学分布的问题,提出了一种基于局部线性嵌入(Locally linear embedding, LLE)和支持向量数据描述(Support vector data description, SVDD)的故障检测算法。该算法充分利用了LLE在特征提取和SVDD在建立统计量时均不受数据数学分布约束的优势,所以可以建立精确的监控模型。此外,利用最小二乘回归策略,解决了LLE不能得到原始空间到特征空间投影矩阵的问题,保证了在线监控的效率。 (2)针对传统监控方法在特征提取时会破坏原始数据局部或全局结构的问题,提出了一种局部-非局部嵌入(Local and nonlocal embedding, LNLE)算法。LNLE利用LLE的目标函数约束样本与其近邻点间的相对位置,同时设计了一个新的目标函数约束样本与其非近邻点间的相对位置,最终通过求解双重优化问题,将原始数据的局部和全局结构信息完整的保存在特征空间中,进一步降低了特征提取过程中的信息损失。 (3)针对训练数据来自多个生产工况即训练数据的数学分布具有多峰性的问题,提出了一种协调混合因子分析(Aligned mixture factor analysis, AMFA)算法。AMFA在传统多模型法的基础上,通过把样本在特征空间中的表达具有唯一性作为优化过程的约束条件,将各个局部模型协调整合得到一个全局模型,从而使监控模型不仅包含每个工况的独有特征而且包含工况间数据的相关性信息。同时,在线监控时,由于不需要判断新样本属于何种工况或应该使用哪个局部模型进行故障检测,所以监控效率得到了提升。 (4)针对现有聚类算法在处理多工况训练数据时无法自动求得工况数以及部分聚类算法可能陷入局部最优的问题,提出了一种工业过程数据聚类算法。该算法利用了样本的时序相关性和工业数据中同类数据总是相邻的特点,通过寻找扩展矩阵中的“断裂点”,提高了聚类的效率和精度。此外,在将局部模型整合为全局模型时,通过同时约束样本与其近邻点及非近邻点间的相对位置,并引入权重系数平衡二者的比例,在特征提取过程中更完整的保存了原始数据的局部和全局结构。 (5)针对同时具有时变、多工况和复杂数据分布的工业过程,提出了移动窗局部离群因子和移动窗局部离群概率两种故障检测方法。利用两种方法基于近邻点计算局部密度从而不受复杂数据分布影响的优势,保证了监控模型的精度。同时,分别针对两种算法提出了相应的移动窗快速更新策略,提高了模型更新速度,保证了监控效率。而对于工况变化过程,又提出了半监督的模型切换机制,通过短时间内强制接受每个新样本使监控模型能迅速跟踪工况的变化。为了减小将故障或扰动更新到窗口中的概率,分别针对两种算法提出了“盲更新”的终止条件,保证了算法的持续有效性。 本文在对上述方法进行理论分析的同时,在数值例子、非等温连续式搅拌釜(Non-isothermal continuous stirred tank reactor, CSTR)和田纳西伊斯曼(Tennessee eastman,TE)过程的仿真框架下,通过设计不同的测试情景并与文献中类似的方法进行对比,验证了本文提出算法的有效性和实用性。
[Abstract]:The emergence of machines makes production efficiency and production safety improved, but mechanical equipment failures also lead to property losses and casualties. As the industrial process continues to become larger and more complex, the traditional method of independent monitoring of each variable is no longer feasible and an accurate analytical model is established. At the same time, with the wide application of distributed control system, a large number of data which reflect the running state of the system from many angles have been recorded. So, in order to further improve the efficiency of process monitoring, the data driven fault detection method began to rise.
The actual industrial process usually has a variety of complex characteristics, such as multiple working conditions, time-varying, complex data distribution, and so on. In the initial research, people will simplify the research problem by setting some assumptions, such as the process running in a single stable condition, the variable obeys the Gauss distribution, so as to simplify the research problem. We need to overcome these hypotheses and study more applicable fault detection methods. This paper aims at the complex data distribution in the industrial process, multi condition and time-varying problem. By using the local and global information of the data, a new feature extraction method is proposed to further improve the accuracy of the obstacle detection model.
(1) a fault detection algorithm based on Locally linear embedding (LLE) and support vector data description (Support vector data description, SVDD) is proposed to solve the problem of complex mathematical distribution of training data under single operating conditions. This algorithm takes advantage of LLE in feature extraction and SVDD in the establishment of statistics. In addition, the least squares regression strategy is used to solve the problem that LLE can't get the original space to the feature space projection matrix, and the efficiency of on-line monitoring is guaranteed.
(2) aiming at the problem that the traditional monitoring method can destroy the local or global structure of the original data when feature extraction, a local non local embedding (Local and nonlocal embedding, LNLE) algorithm.LNLE is used to use the phase pair position between the target function constraint sample and its near neighbor of the target function of LLE, and a new objective function constraint sample is designed. By solving the dual optimization problem, the local and global structure information of the original data is preserved in the feature space, and the information loss in the process of feature extraction is further reduced.
(3) aiming at the problem that the mathematical distribution of training data from multiple production conditions is multi peak, a coordinated mixed factor analysis (Aligned mixture factor analysis, AMFA) algorithm.AMFA is proposed on the basis of the traditional multi model method, by which the expression of the sample in the feature space is unique as the optimization process. The constraints of each local model are coordinated and integrated to get a global model, so that the monitoring model not only contains the unique features of each working condition but also contains the correlation information between the working conditions and the data. At the same time, in the online monitoring, it does not need to judge what working conditions of the new sample or which local model should be used for fault detection. So the monitoring efficiency has been improved.
(4) in view of the problem that the existing clustering algorithm can not automatically get the number of working conditions and the partial clustering algorithm may fall into local optimum when processing the multi condition training data, an industrial process data clustering algorithm is proposed. The algorithm uses the temporal correlation of the sample and the characteristics of the similar data of the same industry in the industrial data. The "breaking point" in the extended matrix improves the efficiency and accuracy of clustering. In addition, when the local model is integrated into a global model, the proportion of the two persons is balanced by the relative position between the constraint samples and their nearest neighbors and the non nearest neighbors, and the original data are preserved more completely during the process of feature extraction. The Department and the global structure.
(5) aiming at the industrial process with time-varying, multi working conditions and complex data distribution, two fault detection methods are proposed for local outliers of mobile windows and local outlier probability of moving windows. The accuracy of the monitoring model is ensured by using two methods to calculate the local density based on the nearest neighbor and thus not affected by the complex data distribution. In order to improve the speed of updating the model and ensure the efficiency of monitoring, a semi supervised model switching mechanism is put forward for two kinds of algorithms, which can improve the speed of model updating and ensure the efficiency of monitoring. The probability of failure or disturbance updating to the window is analyzed, and the termination condition of blind updating is proposed for the two algorithms, which guarantees the continual validity of the algorithm.
In this paper, in the theoretical analysis of the above methods, in the simulation framework of numerical examples, non isothermal continuous stirred kettle (Non-isothermal continuous stirred tank reactor, CSTR) and Tennessee Eastman (Tennessee Eastman, TE) process, different test scenarios are designed and compared with similar methods in the literature. The validity and practicability of the algorithm are proposed in this paper.
【学位授予单位】：华东理工大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TH165.3
本文编号：2172540
[Abstract]:The emergence of machines makes production efficiency and production safety improved, but mechanical equipment failures also lead to property losses and casualties. As the industrial process continues to become larger and more complex, the traditional method of independent monitoring of each variable is no longer feasible and an accurate analytical model is established. At the same time, with the wide application of distributed control system, a large number of data which reflect the running state of the system from many angles have been recorded. So, in order to further improve the efficiency of process monitoring, the data driven fault detection method began to rise.
The actual industrial process usually has a variety of complex characteristics, such as multiple working conditions, time-varying, complex data distribution, and so on. In the initial research, people will simplify the research problem by setting some assumptions, such as the process running in a single stable condition, the variable obeys the Gauss distribution, so as to simplify the research problem. We need to overcome these hypotheses and study more applicable fault detection methods. This paper aims at the complex data distribution in the industrial process, multi condition and time-varying problem. By using the local and global information of the data, a new feature extraction method is proposed to further improve the accuracy of the obstacle detection model.
(1) a fault detection algorithm based on Locally linear embedding (LLE) and support vector data description (Support vector data description, SVDD) is proposed to solve the problem of complex mathematical distribution of training data under single operating conditions. This algorithm takes advantage of LLE in feature extraction and SVDD in the establishment of statistics. In addition, the least squares regression strategy is used to solve the problem that LLE can't get the original space to the feature space projection matrix, and the efficiency of on-line monitoring is guaranteed.
(2) aiming at the problem that the traditional monitoring method can destroy the local or global structure of the original data when feature extraction, a local non local embedding (Local and nonlocal embedding, LNLE) algorithm.LNLE is used to use the phase pair position between the target function constraint sample and its near neighbor of the target function of LLE, and a new objective function constraint sample is designed. By solving the dual optimization problem, the local and global structure information of the original data is preserved in the feature space, and the information loss in the process of feature extraction is further reduced.
(3) aiming at the problem that the mathematical distribution of training data from multiple production conditions is multi peak, a coordinated mixed factor analysis (Aligned mixture factor analysis, AMFA) algorithm.AMFA is proposed on the basis of the traditional multi model method, by which the expression of the sample in the feature space is unique as the optimization process. The constraints of each local model are coordinated and integrated to get a global model, so that the monitoring model not only contains the unique features of each working condition but also contains the correlation information between the working conditions and the data. At the same time, in the online monitoring, it does not need to judge what working conditions of the new sample or which local model should be used for fault detection. So the monitoring efficiency has been improved.
(4) in view of the problem that the existing clustering algorithm can not automatically get the number of working conditions and the partial clustering algorithm may fall into local optimum when processing the multi condition training data, an industrial process data clustering algorithm is proposed. The algorithm uses the temporal correlation of the sample and the characteristics of the similar data of the same industry in the industrial data. The "breaking point" in the extended matrix improves the efficiency and accuracy of clustering. In addition, when the local model is integrated into a global model, the proportion of the two persons is balanced by the relative position between the constraint samples and their nearest neighbors and the non nearest neighbors, and the original data are preserved more completely during the process of feature extraction. The Department and the global structure.
(5) aiming at the industrial process with time-varying, multi working conditions and complex data distribution, two fault detection methods are proposed for local outliers of mobile windows and local outlier probability of moving windows. The accuracy of the monitoring model is ensured by using two methods to calculate the local density based on the nearest neighbor and thus not affected by the complex data distribution. In order to improve the speed of updating the model and ensure the efficiency of monitoring, a semi supervised model switching mechanism is put forward for two kinds of algorithms, which can improve the speed of model updating and ensure the efficiency of monitoring. The probability of failure or disturbance updating to the window is analyzed, and the termination condition of blind updating is proposed for the two algorithms, which guarantees the continual validity of the algorithm.
In this paper, in the theoretical analysis of the above methods, in the simulation framework of numerical examples, non isothermal continuous stirred kettle (Non-isothermal continuous stirred tank reactor, CSTR) and Tennessee Eastman (Tennessee Eastman, TE) process, different test scenarios are designed and compared with similar methods in the literature. The validity and practicability of the algorithm are proposed in this paper.
【学位授予单位】：华东理工大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TH165.3
【参考文献】
中国期刊全文数据库 前10条
1 胡益;马贺贺;侍洪波;;On-Line Batch Process Monitoring Using Multiway Kernel Partial Least Squares[J];Journal of Donghua University(English Edition);2011年06期
2 王洪江;孙保民;田进步;;定性仿真在锅炉状态监控和故障诊断中的应用[J];工程热物理学报;2007年01期
3 陈国金,梁军,钱积新;独立元分析方法(ICA)及其在化工过程监控和故障诊断中的应用[J];化工学报;2003年10期
4 葛志强;宋执环;;一种新的多工况过程在线监测方法[J];化工学报;2008年01期
5 张少捷;王振雷;钱锋;;基于LTSA的FS-SVDD方法及其在化工过程监控中的应用[J];化工学报;2010年08期
6 王凌,王雄;流程工业CIMS体系结构和生产执行系统[J];计算机工程与应用;2003年10期
7 柴天佑,金以慧,任德祥,邵惠鹤,钱积新,李平,桂卫华,郑秉霖;基于三层结构的流程工业现代集成制造系统[J];控制工程;2002年03期
8 许建新;侯忠生;;数据驱动系统方法概述(英文)[J];自动化学报;2009年06期
9 周东华;胡艳艳;;动态系统的故障诊断技术[J];自动化学报;2009年06期
10 常玉清;王姝;谭帅;王福利;杨洁;;基于多时段MPCA模型的间歇过程监测方法研究[J];自动化学报;2010年09期
，本文编号：2172540
本文链接：https://www.wllwen.com/jixiegongchenglunwen/2172540.html
