【作者】 谢磊；

【导师】 王树青；

【作者基本信息】 浙江大学 ， 控制科学与工程， 2005， 博士

【摘要】 间歇过程是现代流程工业中常见的一种生产方式，由于其本身具有的灵活性，间歇过程被广泛应用于医药、染料、香料、生化制品等小批量、高附加值产品的生产和制备当中。间歇过程统计性能监控从历史生产数据出发，通过统计信息处理的方法建立相应的性能监控与故障诊断模型，并用于监视生产过程的进行，及时发现并消除过程的异常状况，实现过程的高效、安全、稳定的运行，最终达到提高产品质量一致性和企业经济效益的目的。由于间歇过程统计性能监控主要依赖于过程数据，通用性较强，因此近年来成为过程控制领域的研究热点之一。 本文系统地研究了间歇过程统计性能监控实施过程中的若干问题，提出了多种针对间歇过程具体数据特点的监控算法。具体包括： (1) 采用蒙特卡罗(Monte Carlo)实验系统地分析了PCA算法给出的统计量及其控制限的性质。在主元数目选择不合适或建模样本偏少的情况下，PCA算法的实际误警率将和理论值产生很大的差异。在理论上分析产生该差异的原因后，提出了一种二阶段PCA分析算法，并在理论上给出了其相关统计量分布情况的证明。实验结果表明，二阶段PCA算法有效克服了样本数目和主元数目对PCA监控结果造成的影响，其实际误警率更加吻合理论值。 (2) 提出了一种鲁棒多向主元分析(MPCA)算法，用于解决历史数据中存在离群点时的间歇过程统计性能建模问题。该方法通过引入抗差估计理论及投影寻踪算法有效地克服了离群点对监控模型的影响。传统主元分析(PCA)与多向主元分析算法基于过程数据的奇异值分解，在建模数据中存在离群点时，其主元方向会偏离实际方向，从而无法得到正确的监控模型并导致整个监控系统的失效。对Dupont间歇聚合反应过程的监控与故障诊断应用研究表明了鲁棒MPCA算法的有效性。 (3) 提出了步进MPCA算法用于多阶段间歇过程的性能监控。该方法通过建立一系列的PCA分析模型避免了传统MPCA算法在线监控时需要对未来数据进行预测的不足，基于数据相似性度量的遗忘因子的引入使得步进MPCA算法可以自然地推广到多阶段间歇过程。通过对多阶段链霉素发酵过程的监控应用表明，相对于传统MPCA，步进MPCA算法能够更精确地描述过程的行为。 (4) 提出一种微粒群一非高斯投影算法，该算法通过引入统计信号非高斯性的度量准则和基于微粒群优化的投影寻踪理论，解决了过程数据不满足正态分布时的统计性能监控问题。该方法将数据分解成一系列相互独立、非高斯递减的变量(独立成分)后进行浙江大学博士学位论文监控，其独立成分的数目及相应的统计控制限可以很方便地确定。通过在工业青霉素仿真平台上的应用，证实了该算法的可行性和优越性。 (5)采用蒙特卡罗实验方法研究了数据动态性对监控系统性能的影响。证明了动态PCA算法不能保证数据间的不相关性。针对数值例子的研究结果表明，PCA算法、动态PCA算法及ARMA一PCA算法在数据出现相关性的情况下，其误警率都会不同程度地偏离理论值。为此，提出了一种基于状态子空间建模的多维时间序列分析算法(SI一PC人)，给出了一种有效判断模型阶次的准则。该方法能够保证模型余差间的不相关性，从而使得实际监控的误警率与理论值更加吻合。通过对TE过程的监控分析表明，该方法能够有效地去除数据的动态性，从而获得更可靠的性能监控结果。 (6)对基于信号重构的传感器故障诊断算法进行了研究，给出了信号重构算法的一般形式和重构信号的统计分布情况，并推导了基于信号重构算法进行故障诊断的可检测与可分离性条件，定义了模型空间和余差空间的故障识别指标。通过CSTR仿真对象的应用比较了不同信号重构算法的差异，验证了基于信号重构的传感器故障诊断算法的有效性。 最后，在总结全文的基础上，对未来的研究课题和工作进行了展望。关键词:间歇过程统计性能监控统计质量控制故障诊断主元分析更多还原

【Abstract】 Batch processes are widely used in process industry. Due to the flexibility, batch processes play important role in the production of low-volume, high-value products such as pharmaceuticals, colorants, flavors and biochemical products. Based on the historical operation data, batch SPMC (statistical performance monitoring & control) establishes monitoring and fault diagnosis models with statistical approach and applies them in the on-line monitoring of batch process operation. Batch SMPC, enabling the detection and correction of abnormal process behaviors, guarantees the efficient, safe and stable operations of batch processes and thus improves the quality consistency and the profits of enterprise. Because batch SPMC only relies on process data and thus general-purposed, it has become one of the most active research areas in process control.In this thesis, some important problems of batch SMPC are widely studied and some new SMPC algorithms are proposed with respect to the characteristic of batch process data.The main contributions are as follows:(1) Mote-Carlo experiments are involved to analyze the properties of the statistics of PCA. It is shown that the actual false alarms rate will be different with the theoretical value if the number of modeling data is too few or the number of principal components is incorrect. After analyzing the cause of the difference, a new two-stage PCA algorithm is proposed and demonstration of the distributions of relative statistics is given. Further experiments demonstrate that two-stage PCA overcomes the problem of tradition PCA.(2) A robust Multiway Principal Component Analysis (MPCA) is developed to construct monitoring models when outliers are present in historical data. By involving the robust estimation theory and projection pursuit algorithm, robust MPCA eliminates the effect of outliers efficiently while tradition MPCA is strongly affected by them. Applications on DuPont batch polymerization process reveal the efficiency of robust MPCA.(3) A new step-by-step batch SPMC approach is proposed for the monitoring of multi-stage batch processes. It overcomes the need to estimating or filling in the unknown part of the process variable trajectory deviations from the current time until the end. The adaptive rate is easily controlled through a parameter that controls the weight of past data in a summation manner. This algorithm is evaluated on industrial fermentation process data and compared with the traditional MPCA. The method describes the behavior of batch process more precisely and has significant benefits especially when monitoring multi-stage batch process where the latent vector structure may change at several points.(4) A PSO-NGPP batch monitoring approach is proposed without assuming that thelatent variables subject to normal distribution. The approach is based on non-Gauss evaluation criterion, projection pursuit and partial swarm optimization. PSO-NGPP decomposes the original signals into a series of independent components whose distributions can be easily obtained by kernel density estimation. Application results on the penicillin fermentation benchmark process demonstrate the power and advantages of presented method.(5) The influence of auto & cross correlations on statistical process control (SPC) is investigated in detail via Monte Carlo experiments. It is revealed that non independent identically distribution will cause the actual FAR to deviate from its theoretic value apparently and result in unexpected consecutive false alarms for normal operating process. Dynamic PCA and ARMA-PCA are demonstrated to be inefficient to remove the auto & cross correlations which still influence the FAR. Subspace identification based PCA (SI-PCA) is proposed to improve the monitoring of dynamic processes which can remove the auto & cross correlations efficiently and avoid consecutive false alarms. A criterion to determine the order of SI model is also introduced. The applications in Tennessee Eastman challenge process illustrate the advantages of the proposed approach.(6) Senso更多还原