基于邻域保持嵌入算法的间歇过程监控与故障诊断研究

Production safety and reliability of widely used batch process are more significance. So research on monitoring and fault diagnosis of batch process has vitally important value. Because its characteristics of strong nonlinearity, dynamics multi-distribution, multi-mode and three-dimensional data, batch production process is more complex than continuous production process. It is very difficult to monitor and diagnose faults. Aiming at the characteristics of three-dimensional data, strong nonlinearity, dynamics, multi-distribution and multi-mode for this project, the existing fault diagnosis technologies are analyzed from the various aspects. A series of monitoring and fault diagnosis algorithms of batch process based on NPE are proposed. These improved algorithms will be suited to features of batch process to reduce fault alarm rate and fault missing rate to improve accuracy of fault diagnosis. These improved algorithms will be used to verify their feasibility and superiority in the penicillin fermentation batch process and DuPont polymerization reactor. Then they will be applied to practical batch processes to test the practicability.

广泛应用的间歇生产过程，人们越来越重视其安全性和可靠性，因此研究间歇生产过程监控和故障诊断具有极其重要的价值。但由于间歇过程与连续过程相比更加复杂，具有强非线性、动态性、多分布和多模态等特点，且数据呈现三维特性，对它进行监控和故障诊断也更加困难。本项目针对间歇过程中出现的数据三维性、非线性、多模态、动态性和多分布特性，从各个方面对已有的故障诊断技术进行分析，提出基于改进NPE的一系列间歇过程监控与故障诊断算法，使这些算法适应更多的间歇过程特点，降低故障的误报率和漏报率，从而提高对间歇过程故障诊断的准确性，并在青霉素发酵间歇过程中和DuPont工业间歇聚合反应器生产中验证这些算法的可行性和优越性，再进一步应用到实际的间歇生产过程中，检验这些算法的实用性。

由于间歇过程作为一种常见的工业生产方式，具有强非线性、动态性、多分布和多模态等特点，对它进行监控和故障诊断更具挑战性。而间歇过程的监控与故障诊断研究存在对特征信息的提取会不充分，必将丢失大量的过程信息，其它特性必然会造成故障检测的误报率和漏报率升高，使得故障诊断准确率下降，导致不能有效地对过程进行监控。邻域保持嵌入（Neighborhood Preserving Embedding, NPE）算法是一种典型的流形学习算法，可以提取原始变量空间中隐藏的低维流形信息，保持数据集的局部邻域结构不变，发现隐藏在高维数据集中的更多信息，得到精确的线性投影矩阵，因此本项目同时针对间歇过程的多个特性的不利影响，利用邻域保持嵌入算法的优势，克服其存在的缺点，提出了一系列改进NPE算法提高间歇过程监控与故障诊断的准确性，在青霉素发酵间歇过程等中验证这些算法的可行性和优越性，能较好地解决了间歇过程监控与故障诊断问题。