基于多变量统计的乘性故障诊断方法研究

Driven by the demands for the high-efficient operation and stable product quality of industrial processes, fault diagnosis has become a hot topic of research both in industry and academia. Compared with the commonly encountered additive fault, the multiplicative fault that reflects changes in the parameter of system models is more complicated and hard to model, which has been rarely studied in fault diagnosis field. Moreover, in the real industry, multiplicative faults are commonly existing and also threatening the safety and product quality of industrial processes. Based on these observations, this project seeks to develop multivariate statistics-based methods for the multiplicative fault diagnosis problem, and accomplish in detail the following tasks: (1) Using the hybrid of analytical models and statistical properties of process data to reveal the mechanism of how multiplicative faults affect systems; (2) Develop high-efficient multiplicative fault detection methods based on detection statistics, aided by multivariate data analytical methods, and evaluated by the fault detection rate index; (3) Develop contribution plot-based multiplicative fault identification methods based on the multiplicative fault model using fault reconstruction and estimation techniques. The project seeks to construct a fault diagnosis framework consisting from knowing multiplicative faults to detecting multiplicative faults until identifying multiplicative faults, which builds the theoretical foundation for effectively diagnosing multiplicative faults in real industrial processes.

在工业过程安全高效运行与产品质量合格稳定的实际需求的驱动之下，故障诊断技术已成为工业界与学术界的研究热点。与通常考虑的加性故障相比，乘性故障主要体现在系统模型参数的变化上，对整个过程的影响更为复杂也更加难以建模，学术界对其研究较少。但实际过程中乘性故障也较为普遍，并同样威胁着系统的安全运行与产品质量。针对这一现状，本课题拟采用多变量统计的方法对乘性故障诊断问题展开研究，拟解决以下三个方面的关键问题：（1）通过机理模型与数据统计特性相结合的方法揭示乘性故障对系统影响的机理；（2）建立以检测统计量为根本，多变量数据建模方法为辅助，故障检测率为评价指标的高性能乘性故障检测方法；（3）研究以乘性故障模型为基础，乘性故障重构与估计为关键技术，基于贡献图的乘性故障辨识方法。本课题旨在建立一套从故障作用机理分析到故障检测再到故障辨识的故障诊断框架，为实际工业过程中的乘性故障诊断问题提供理论依据。

工业过程的安全高效运行通过构建过程监控和故障诊断机制来保障。本项目以对工业过程影响大、常发生，但又经常被忽视的乘性故障为对象，重点研究该类故障的检测与诊断问题。项目以数据驱动技术为核心，以多变量统计方法为抓手，对乘性故障建模、检测和诊断等问题展开详细研究，主要解决了以下三个方面的关键问题：（1）通过机理模型与数据统计特性相结合的方法揭示了乘性故障系统影响的机理，给出了乘性故障与加性故障对系统的不同影响，并从数据特征的角度证明了乘性故障对被控变量高阶统计量的影响；（2）通过对多种故障检测统计量检测性能评估的研究解揭示了乘性故障检测的难点，并给出了如何提高乘性故障检测率的有效途径，证明了在有足够数据的条件下，基于Wishart分布、信息熵、信息判据等方法可以提高乘性故障检测性能；（3）通过对多变量数据建模方法的研究揭示了PCA、PLS、CCA及其动态和非线性扩展方法在挖掘数据内部特征的异同，给出了三类方法的内在等价关系，提出了针对多模态数据的共性-个性PCA、PLS和基于深度置信网络的特征提取方法，实现了过程信息的深度提取，提出了针对厂级互联系统的分布式CCA模型，有效地扩展了多变量建模方法的适用范围。本课题建立了一套从乘性故障作用机理分析到乘性故障检测再到故障辨识的故障诊断框架，本项目理论研究结果利用热连轧过程真实数据进行验证。同时基于以上理论成果，项目组依托热连轧虚拟仿真平台和热连轧过程数据库开发了故障诊断与过程监控性能评估平台和热连轧过程监控与故障诊断平台，搭建了从理论研究到实际应用的桥梁。