【作者】 邓鹏程；

【导师】 桂卫华；

【作者基本信息】 中南大学 ， 控制科学与工程， 2011， 博士

【摘要】 密闭鼓风炉铅锌熔炼过程是一个复杂的物理化学变化过程,具有多变量、非线性、强耦合、大滞后和不确定性等特点,由于机理反应复杂,原料成分多变及工作环境恶劣,导致铅锌熔炼过程工况频繁变化、多工作点运行、数学描述困难和故障时常发生,从而影响铅锌的产量和质量。目前,铅锌熔炼过程在线监控与故障诊断主要凭人工经验进行判断,无法及时发现和诊断故障,难以保证生产过程持续稳定运行。因此,研究铅锌熔炼过程在线监控与故障诊断新技术,对于保证生产安全稳定运行,提高企业经济效益具有重要的现实意义。本文针对铅锌熔炼过程工况频繁变化和多工作状态切换的特点,利用生产现场采集的大量工况数据,研究了基于过程工况数据的自适应在线监控与故障诊断新方法。针对铅锌熔炼过程在线监控与故障诊断存在的问题,提出了基于一阶扰动的递归动态主元分析方法、基于潜空间变换的自适应建模方法、基于多模型切换的自适应在线监控方法、基于奇异值分解的故障二次分离方法和基于极大似然估计的故障辨识方法。并将其应用于铅锌熔炼过程实际生产中,取得了较好效果。主要工作和创新点包括以下几个方面：(1)针对铅锌熔炼过程机理反应复杂、工况频繁变化和多工作状态频繁切换的问题,将铅锌熔炼过程每个加料周期划分为三个子过程,包括初始阶段,过渡阶段和稳定阶段。然后根据三个子过程工况变化的不同特点,采用静态主元分析方法建立正常工况下铅锌熔炼过程初始阶段静态监控模型,采用动态主元分析方法建立过渡阶段动态监控模型,提出一种基于潜空间变换的自适应建模方法提取稳定阶段正常工况变化的模型特征,建立稳定阶段自适应监控模型。(2)针对传统监控统计变量(如SPE)阈值计算不精确和在线计算复杂的问题,提出两个改进的监控统计量。首先深入分析传统监控统计量存在的问题,构造两个改进的监控统计量。然后比较各个监控统计量的故障敏感度,从理论上证明了改进的统计量具有更强的故障检测能力,同时在线计算方法简单、阈值确定更精确。(3)针对铅锌熔炼过程多工作状态运行和工况频繁变化引起的系统故障误报警的问题,提出了一种基于动态递归主元分析和多模型切换方法的铅锌熔炼过程自适应在线监控策略。采用基于标准主元分析的在线监控方法对铅锌熔炼过程初始阶段工况实行在线监控。然后提出一种基于一阶扰动的动态递归主元分析方法实时跟踪过渡阶段工况的缓慢变化,实现过渡阶段工况自适应在线监控。最后提出一种基于正常工况变化模型(工况漂移,工况扩大和工况突变模型)的多模型切换方法,实现稳定阶段工况自适应在线监控。实验结果表明,本文提出的白适应在线监控策略能有效跟踪铅锌熔炼过程工况的正常变化,消除系统故障误报警并有效检测故障。(4)针对铅锌熔炼过程不同类型故障引起相同状态变量异常而导致系统故障误诊断的问题,提出了一种基于奇异值分解的铅锌熔炼过程故障二次分离方法。首先根据异常状态变量和故障类型的对应关系,提出一种基于奇异值分解的故障分离方法初步确定故障类型。然后根据异常状态变量与故障类型的特殊对应关系实现故障的二次分离。实验结果表明,该方法能够有效分离故障,减少系统故障误诊断的发生。(5)针对铅锌熔炼过程故障辨识困难的问题,提出了一种基于极大似然估计和奇异值分解的故障辨识策略,实现故障发生程度的评价。首先从理论上将故障分为均值故障和方差故障两大类,然后采用多元正态分布均值和协方差假设检验方法判断故障类别(均值故障或方差故障)。最后采用极大似然估计方法辨识均值故障的大小,采用奇异值分解的方法辨识方差故障的大小。实验结果表明,本文提出的故障辨识策略能有效辨识均值故障和方差故障的大小。更多还原

【Abstract】 Zinc-Lead smelting process, a complicated process with some physicochemical reactions, is always characterized as multivariate, nonlinearity, strong coupling, large delay and uncertainty. The complicated reactions, unstable component of raw materials and poor working environment make the smelting process changeable and cause many faults, which consequently lead to poor output and quality of Lead-Zinc. Therefore, some researches on new methods for Zinc-lead process monitoring and fault diagnosis have an important and practical significance for guaranteeing the production safety and stability and improving industrial economic profits.According to the characteristics of Zinc-lead smelting process, this paper focuses on the new methods of adaptive process monitoring and fault diagnosis with the aid of abundant process data. To solve the problems in process monitoring and fault diagnosis system, some new and efficient methods are proposed as follows:latent space transformation (LST) and multimode switch based process monitoring approaches, first order perturbation based recursive principal component analysis (PCA), Singular value decomposition (SVD) based 2-steps fault isolation approaches, maximum-likelihood estimation based fault identification approach. All this methods are applied to the process monitoring and fault diagnosis system in the smelting process, and some good performances are achieved, the main content and some innovative achievements are depicted as follows:(1) To solve the problem of hardly constructing a mathematical model due to the complicated reactions and changeable process in actual Zinc-Lead smelting process, a PCA based smelting process off-line modeling strategy is presented. To improve the quality of model data, some methods called the exponentially weighted moving average (EWMA) is firstly adopted to preprocess the process data. Secondly the process in one circle is classified as 3 sub-processes that are initial stage, transition stage and stable stage. The static PCA modeling approach is selected to deal with initial stage and get a static process monitoring model. The dynamic PCA modeling approach is chose to cope with transition stage and get a dynamic process monitoring model. Finally, a LST approach is proposed to extract the characteristics of normal process changes including drifting process, enlarging process and bias and achieve an adaptive process monitoring model.(2) To solve the problem of inefficient threshold calculation of traditional monitoring statistical variables such as T2 and SPE, two new statistical variables are proposed. First some detailed analysis inefficient of threshold calculation is carried out, and then the fault sensitivity of different statistical variables is compared and testified to illustrate that the proposed new statistical variables have a better performance to detect the fault. Experimental results show that the proposed methods can detect faults efficiently and the corresponding thresholds are able to be computed much simpler and more accurate.(3) To solve the problem of false alarm due to multiple operating points switch and normal process changes, first order perturbation dynamic recursive PCA and multi-model switch approach based smelting process adaptive on-line monitoring strategy is proposed to solve the problem. The traditional PCA based monitoring approach is selected to deal with the initial stage, and then the FOP based recursive PCA approach is chose to cope with transition stage and for adaptive process monitoring, finally a normal process changing model based multi-model swith approach is proposed to handle the stable stage for adaptive process monitoring. Experimental results show that the proposed methods can adapt to the normal process changes and eliminate the false alarm and detect process faults efficiently.(4) To solve the problem of false diagnosis due to the fact that different faults in Zinc-Lead smelting process may make the same variables to be abnormal, a SVD based 2-steps fault isolation approach is proposed. Firstly, the fault types can be preliminarily determined with the aid of one-to-one relationships between abnormal variables and fault types, and then some specific relationships between abnormal variables and fault types are analyzed to realize final fault isolation. Experimental results show that the proposed methods can reduce the false diagnosis rate and isolate the fault effectively.(5) Maximum-likelihood estimation (MLE) and SVD based Zinc-Lead smelting process fault diagnosis strategy is proposed to achieve fault evaluation. Firstly, faults are theoretically divided into two types that are off-set fault and scaling fault, and then Multivariate normal distribution mean and variance hypothesis testing approach is selected to distinguish the two faults. Finally, a MLE based fault identification approach is presented to identify the magnitude of off-set fault, and a SVD based fault identification approach is proposed to identify the magnitude of scaling fault.更多还原