Abstract:

For the nature of multimode and non-Gaussian distribution in industrial process data, a fault detection method was proposed for multimode processes based on independent component analysis mixture model (ICAMM). In this method, Bayesian inference and independent component analysis (ICA) were combined to create a probability mixture model; the mode classification of each observation by Bayesian inference and ICA model parameters' setting were accomplished simultaneously; and the global monitoring statistics were established within the Bayesian framework to monitor real-time process changes. In order to solve the problem that traditional variable contribution plot could not indicate the relationships of information transmission among fault variables after fault detection, a fault recognition method for multimode processes was further proposed on the basis of information transfer contribution plot. Three steps were developed in the fault recognition method, including the calculation of variable contributions to the independent component analysis mixture model, the determination of cause-and-effect relationships of fault variables through variable transfer capability and the nearest neighbor transfer entropy, and the finding of fault source variables and fault propagation process. Simulation study on a numerical example and continuous stirring tank reactor (CSTR) system showed effectiveness of the proposed approach.

Key words: ICA mixture model, multimode process, posterior probability, transfer entropy, contribution plot

摘要：

针对工业过程数据的多模态和非高斯特性，提出一种基于独立元混合模型（independent component analysis mixture model,ICAMM）的多工况过程故障诊断方法。该方法将独立元分析与贝叶斯估计结合，同时完成各个工况的数据聚类和模型参数求取，并建立基于贝叶斯框架下的集成监控统计量实时监控过程变化。在检测到故障后，针对传统的变量贡献图方法无法表征变量之间信息传递关系的缺点，提出基于信息传递贡献图的故障识别方法。该方法首先计算各变量对独立元混合模型统计量的贡献度，进一步通过最近邻传递熵描述故障变量之间的传递性，挖掘故障变量之间的因果关系，从而确定故障源变量和故障传播过程。最后对一个数值系统和连续搅拌反应釜（CSTR）过程进行仿真研究，结果验证了本文所提出方法的有效性。

关键词: ICA混合模型, 多工况过程, 后验概率, 传递熵, 贡献图