数据驱动的过渡过程建模与监测

发布时间：2018-01-13 05:01

本文关键词：数据驱动的过渡过程建模与监测　出处：《浙江大学》2017年博士论文　论文类型：学位论文

【摘要】：由于工况输入点改变、原料变化、季节因素以及设备老化等原因,工业过程运行状态会发生不同程度的改变。在多工况过程建模和故障检测中,需要对这些变化加以识别以便实现不同阶段的故障检测。传统的多工况过程监测通常着眼于稳态工况而忽略了过渡过程的存在。在过渡过程中,数据特性的剧烈波动会对过程的平稳运行产生一定程度的影响。因此,对过渡过程的建模和在线监测是工业过程监测领域中一个重要的研究课题。本文通过对过渡过程中的不同数据特性,包括非线性特性、非高斯特性及动态特性进行建模分析,着重研究了过渡过程的识别和在线故障检测方法。全文的主要研究内容如下:首先,考虑过渡过程数据中的非线性特性的影响,将一个过渡过程视为多个子阶段,提出了一种多投影模型迭代更新的数据分类方法,并综合多次数据分类情况得到最终多工况数据的分类结果。同时设置过渡过程判别规则来判断分类结果中过渡过程各子阶段和各稳态工况的分布,从而达到离线识别过渡过程的目的。之后,提出了一种基于KPLS模型序列的过渡过程在线识别和故障检测方法,通过模型的在线更新完成过渡过程的在线识别和故障检测的任务。其次,考虑过渡过程数据中的非高斯特性的影响,提出了一种基于独立元交互信息差异度的过程故障检测方法。针对过渡过程的时变特性,引入即时学习的思想,通过求取在线数据和历史数据之间的交互信息,选取相似历史数据作为在线训练集。同时运用核密度估计方法,并结合独立元交互信息差异度分析方法对统计限进行在线更新。之后,通过比较在线数据和在线训练集之间的独立元交互信息差异度,实现过渡过程的故障检测。再次,考虑过渡过程数据中的自相关特性的影响,提出了一种基于动态交互信息的方法用以提取过程的动态信息,并通过度量多工况过程中不同阶段之间动态特性的相似度,完成过渡过程的离线识别工作。同时,为兼顾算法快速性和高效性,提出了基于两步移动窗策略的多工况过程识别方法。随后,提出了基于DPLS模型序列的过渡过程在线识别和故障检测方法,通过模型的在线更新完成过渡过程在线识别和故障检测的工作。最后,对本文所做工作进行了总结,并对相关领域的未来研究工作进行了简要分析和展望。
[Abstract]:Due to the change of input point, material change, seasonal factors and aging of equipment, the operating state of industrial process will change to some extent. These changes need to be identified in order to achieve different stages of fault detection. Traditional multi-condition process monitoring usually focuses on steady state conditions and neglects the existence of transition process. The violent fluctuation of data characteristics will have a certain degree of impact on the smooth operation of the process. Modeling and on-line monitoring of transition process is an important research topic in the field of industrial process monitoring. Non-#china_person0# characteristics and dynamic characteristics of modeling analysis, focusing on the transition process identification and on-line fault detection methods. The main contents of this paper are as follows: first. Considering the influence of nonlinear characteristics in transition process data, a new data classification method based on iterative updating of multi-projection models is proposed, in which a transition process is regarded as multiple sub-stages. Finally, the classification results of the final multi-condition data are obtained by synthesizing the classification of multiple data. At the same time, the transitional process discriminant rules are set to judge the distribution of each sub-stage and each steady state of the transition process in the classification results. In order to achieve the purpose of off-line identification of transition process, a method of on-line identification and fault detection of transition process based on KPLS model sequence is proposed. The task of on-line identification and fault detection of transition process is completed by online updating of the model. Secondly, the influence of non-Gaussian characteristics in transition process data is considered. In this paper, a process fault detection method based on the difference degree of independent element interactive information is proposed. According to the time-varying characteristics of transition process, the idea of real-time learning is introduced, and the interactive information between online data and historical data is obtained. The similar historical data is selected as the online training set and the statistical limit is updated online by using the kernel density estimation method and the independent element interactive information difference analysis method. By comparing the information difference between online data and online training set, the fault detection of transition process is realized. Thirdly, the influence of autocorrelation in transition process data is considered. A method based on dynamic interactive information is proposed to extract the dynamic information of the process and measure the similarity of the dynamic characteristics between different stages in the multi-condition process. At the same time, in order to give consideration to the fast and high efficiency of the algorithm, a multi-condition process recognition method based on two-step moving window strategy is proposed. A method of on-line identification and fault detection of transition process based on DPLS model sequence is proposed. The on-line identification and fault detection of transition process are completed by on-line updating of the model. Finally. The work done in this paper is summarized, and the future research work in related fields is briefly analyzed and prospected.
【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP274

【参考文献】