A Novel Non-Gaussian Dynamic Process Monitoring Method for Key Performance Indicators

Abstract

Key performance indicators (KPIs) play a crucial role in ensuring enterprise economic benefits and product quality in modern industrial production. However, the existence of dynamic and non-Gaussian disturbances between process and KPI variables presents significant challenges for conventional static fault detection techniques based on Gaussian assumptions. In this article, a novel KPI-related process monitoring framework for non-Gaussian dynamic processes is developed, referred to as a class of Gaussianized slow feature analysis monitoring (GSFAM) method. First, a dual-space information entropy Gaussianization (DEG) method is constructed, which realizes simultaneous marginal Gaussianization transform and simultaneous rotational independence of the dual data space by maximizing mutual information (MI) and minimizing residual entropy. Then, based on the data Gaussianization, a novel KPI-related enhanced probabilistic slow feature analysis (EPSFA) model is constructed, whose extracted slow features account for both predictive performance and smoothness, in view of the inherent dynamic characteristics of chemical processes. Furthermore, a class of non-Gaussian dynamic monitoring strategies with high transferability is formed by integrating the data Gaussianization strategy and the KPI-related EPSFA model. Experiments on the Tennessee Eastman process, the fluidized catalytic cracking unit, and a real closed-loop water circulation system verify the effectiveness of GSFAM method.