Application of randomized algorithms to the analysis and synthesis of model-based and data-driven fault diagnosis systems

本研究项目致力于分析和综合基于观测器和数据驱动的故障检测系统(FD系统)。就我们的目的而言，所谓的随机算法(RA)是研究和调查的数学和系统基础。该项目的第一个目标是为基于观察者的FD系统的分析和综合勾画一个基于RA的框架。在这个框架的基础上，将开发计算虚警率(FAR)和漏检率(MDR)的方法和算法。此外，还将讨论在FAR和MDR的背景下为优化基于观察者的FD-系统而开发的统计学习方法。考虑到可用过程数据可能不能覆盖所有故障操作场景的实际情况，数据驱动的故障诊断方法在这些情况下可能会失败，无法成功地进行故障隔离和识别。本项目的第二个目标是开发基于RA的方法和算法，用于生成故障数据并进一步识别适用的故障模型。这些方法将被应用于主成分分析和偏最小二乘等多元分析方法，并进一步应用于数据驱动的基于观测器的FD系统。该项目的最后一个目标是在RA方法的基础上开发一个用于数据驱动和基于模型的FD-系统的评估平台，用于对FD-性能的可比性评估，如FAR、MDR和检测时间。所获得的理论结果将在田纳西·伊斯曼基准工艺和供热交换控制实验室系统上进行验证和演示。最后，在MatLab编程环境下开发了一个软件工具，用于基于RA的基于观测器的FD系统的分析和综合，并建立了带Web访问的FD评估平台。

This research project deals with the analysis and synthesis of observer-based and data-driven fault detection systems (FD-systems). For our purpose, the so-called randomized Algorithms (RA) serves as the mathematical and systematic fundamentals of the study and investigation. The first objective of the project is to sketch a RA-based framework for the analysis and synthesis of observer-based FD-systems. On the basis of this framework, methods and algorithms will be developed for the computation of false alarm rate (FAR) and missed detection rate (MDR). Moreover, development of statistical learning methods for the optimization of observer-based FD-Systems in the context of the FAR and MDR will be addressed. Considering practical situations that the available process data may not cover all faulty operation scenarios, data-driven FD-methods may fail in these situations for a successful fault isolation and identification. It is the second objective of this project to develop methods and algorithms, which are based on the RA and used for generating fault data and further for identifying applicable fault models. These methods will be applied to the multivariate analysis (MVA) methods like PCA (Principle Component Analysis) and PLS (Partial Least Squares) and further to the data-driven observer-based FD-systems. The last objective of the project consists in developing, on the basis of the RA methods, an evaluation platform for the data-driven and model-based FD-systems, which is used for a comparable evaluation of FD-performance like FAR, MDR and detection time. The achieved theoretical results will be tested and demonstrated on the benchmark process Tennessee Eastman and on the laboratory system of heating exchange control. Finally, a software tool in the MATLAB-programming environment will be developed for the RA-based analysis and synthesis of observer-based FD-systems, and the FD-evaluation platform will be established with a web-access.