【作者】 赵春晖；

【导师】 王福利；

【作者基本信息】 东北大学 ， 控制理论与控制工程， 2009， 博士

【摘要】 作为工业生产中一种重要的生产方式,间歇过程与人们的生活息息相关,已被广泛应用于精细化工、生物制药、食品、聚合物反应、金属加工等领域。近年来,随着现代社会对多品种、多规格和高质量产品更迫切的市场需求,工业生产更加倚重于生产小批量、高附加值产品的间歇过程。间歇生产的安全可靠运行以及产品的高质量追求已成为人们关注的焦点。基于数据的多元统计分析技术因其只需要正常工况下的过程数据来建立模型,同时它们在处理高维、高度耦合数据时具有独特的优势,越来越受到研究人员和现场工程师的青睐。间歇过程的统计建模、在线监测、故障诊断及质量预测已成为广泛的研究课题。与连续工业过程相比,间歇生产的过程特性更加复杂,数据的统计特征亦更为丰富。同一操作周期内又分成多个子时段,每个时段都有其特定的控制目标,有不同的过程主导变量,呈现不同的过程相关特性。因此,面向多时段间歇生产过程的统计分析及在线应用更具挑战性,不仅仅要关注整个过程的运行状况及其与最终产品质量的因果关系,更应该深入分析过程的每一个子时段,揭示其不同的潜在过程特性并发掘它们对质量指标不同的影响效果和作用能力。本论文在深入研究间歇过程多时段特性的基础上,从解决实际问题的角度出发,提出了一系列基于时段的间歇过程统计建模、在线监测和质量预测算法：(1).针对多时段间歇过程中的时段过渡现象,引入模糊时段的概念,提出了一种基于过渡的软时段划分方法,将间歇过程按其潜在特性的不同详细划分成若干子时段及主要时段之间的过渡区域,并针对它们不同的数据特征建立了基于软划分的统计建模及在线监测算法。(2).针对建模数据不充分的问题,仅仅基于少量几个正常批次进行统计分析,从中提取子时段信息,提出了基于时段的统计建模及在线监测算法；同时建立了模型在线更新策略,利用逐渐增多的正常间歇操作批次不断自适应更新和完善监测模型以实现准确的过程监测。(3).受各种因素影响,间歇过程批次间通常呈现缓慢波动特性。对于慢时变间歇过程,本文提出了基于批次间“相对变化”的子时段统计建模和在线监测算法。该算法从表征“相对变化”的批次间差分轨迹中提取、训练慢时变运行模式本身的统计规律及演化特性,将慢时变模式主动容纳到初始监测系统,从而赋予初始模型对于批次间慢时变行为的自动适应能力而无需在线更新,增强了监测模型的鲁棒性。(4).针对“段型质量指标”,鉴于它只决定于某一个或几个特定的子时段,其他时段对其没有显著影响,本文建立了一种基于子时段PLS回归模型的质量分析及预测算法。一方面,从时段整体的角度识别了影响质量的各个关键时段及其关键变量,并建立了基于子时段的PLS在线质量预测模型；另一方面从时段平均运行水平的角度分析了各个关键时段对于质量指标的平均作用效果,建立了更为稳定的质量预测关系。(5).针对“过程型质量指标”,鉴于其受间歇操作周期内所有子时段的共同作用,本文进一步发展了基于时段的质量预测算法,深入理解并分析了各个子时段对质量指标的局部影响力及其共同作用效果。一方面,在各个局部时段内,该算法定量化提取了与质量指标相关的关键过程信息以及各个子时段分别独立表征的局部质量性能,增强了二者之间的因果关系；另一方面,从全过程的角度,它将各子时段的局部作用累加起来从而获得整个间歇过程对于质量指标的综合解释与预测能力。基于时段的统计分析策略可以更细致地反映过程运行状态及其潜在相关特性在时段间的发展变化,改善在线监测性能；同时它能够深入理解过程与质量二者之间的因果关联,揭示各个时段对质量指标的不同影响作用,有利于质量解释与预测。上述算法在间歇过程实验系统及仿真研究中验证了其有效性与可靠性,从而丰富了间歇过程统计建模、在线监测及质量预测的研究成果,并揭示了进一步研究的必要和可能。更多还原

【Abstract】 As an important industry production way, batch processes, which have a close relationship with people’s everyday life, have been widely applied to fine chemical, biopharmaceutical, food, polymer, and metal industries etc. Recently, with the urging market requirement for various product types and high product quality, the manufacturing of higher-value-added products that are mainly produced through batch processes have become increasingly important in many industries. The batch process safety and product quality have been the focus of people’s attention. Data-based multivariate statistical analysis techniques only require the normal process data for modeling and show particular advantages to deal with the high-dimension and coupling data, which makes them specially and increasingly attractive. Multivariate statistical modeling, online monitoring, fault diagnosis and quality prediction have been under wide investigation for batch processes.Batch processes are fairly more complex with more rich data statistical characteristics compared with continuous processes. The operation process covers multiple phases, which have specific control objects, different dominant process variables and distinct process correlation characteristics. Therefore, it is more challenging to conduct statistical analysis and online application for multiphase batch processes. It should not only pay attention to the whole operation status and find the causal relationship between process variables and quality variables, but also focus on different phases to analyze their local nature and reveal their different effects on quality.Based on the further research on multiplicity of operation phase, this dissertation developes a series of phase-based statistical modeling, process monitoring and quality prediction methods for batch processes to solve the practical problems:(1). For the phase transition behaviors in multiphase batch processes, the concept of fuzzy phase is introduced and a transition-based soft phase partition algorithm is developed, which divides the process into different subphases and transition regions between neighboring phases according to the changes of underlying process correlations along time. Consequently, focusing on their different data nature, different statistical models are respectively developed as well as the corresponding online monitoring strategy. (2). For limited reference batches, a new statistical analysis strategy is proposed to explore the phase information and thus develop the phase-based statistical models for online monitoring. Meanwhile, an online adaptive updating strategy is adopted to adjust the monitoring models with the accumulation of new successful batches, which can enhance the reliability of monitoring results.(3). Under the influence of various factors, batch processes commonly involve normal slow variations over batches. For slow-varying batches, a phase-based statistical modeling and nonline monitoring algorithm is developed using between-batch "relative changes". It extracts and models the statistical rules and evolving characteristics of slow-varying behaviors from the between-batch difference trajectory that represents the batch-wise relative changes, which thus accommodates the slow-varying mode into the initial monitoring models. In this way, it endows the initial monitoring system with adaptive competency to batch-to-batch slow-varying behaviors, avoids the online updating requirement and enhances the robustness of monitoring models.(4). For phase-type quality, which is only determined by one or several specific phases and has no close relationship with others, a phase-based PLS regression modeling algorithm is developed for quality analysis and prediction. On the one hand, it identifies the critical-to-quality phases and key predictors from an overall phase viewpoint, and then develops phase-based PLS models for online quality prediction. On the other hand, it analyzes the average effects of process behaviors in each phase on quality based on phase-specific average trajectory and thus develops a stabler quality prediction relationship.(5). For process-type quality, which depends upon all phases within the batch operation, it further develops phase-based quality analysis strategy to comprehend and analyze the local and cumulative effects of various phases on quality more detailedly. On the one hand, it quantitatively extracts the critical-to-quality feature information within each phase and the part of quality attribute interpreted by each local phase, which, thus, enhances the causal relationship between them. On the other hand, it captures the overall effects of the whole process on quality interpretation and prediction by stacking the different local effects of various phases.Phase-based multivariate statistical analysis can more detailedly reveal the process operation status and the changes of underlying characteristics over different phases, which will help to improve the online monitoring performance. Moreover, it can further comprehend the causal relationship between process variables and quality attribute, and explore the different effects of various phases on quality, which will be of great benefit to quality interpretation and prediction. The successful applications to batch process systems and simulation experiments demonstrate the effectiveness of the present methods, which, thus, enrich the achievement of statistical modeling, online monitoring and quality prediction for batch processes and also suggest the necessarity and potential of further research.更多还原