Advanced Data-driven Fault Detection in Chemical Processes

—Accurate first-principle models are sometimes difficult to obtain for complex high-dimensional processes, which has driven interest in the development of robust data-driven frameworks capable of understanding process dynamics using only data samples gathered from its variables. Data-driven process monitoring in industrial processes is a multi-stage framework that involves modeling, fault detection and fault classification. Fault detection is a critical part of process monitoring, where the objective is to flag unexpected operating behavior quickly and accurately. In this paper, a novel extension of the Generalized Likelihood Ratio (GLR) chart is proposed, denoted as the Maximum Multivariate GLR (MMGLR) chart. Linear and nonlinear data-driven models, namely principal component analysis (PCA) and Bayesian-optimized neural networks (BONN), are combined with different statistical charts towards the detection of multiple fault types in chemical processes. The results show that the MMGLR charts have a better detection accuracy than conventional charts, and that neural networks are more robust models and fault detectors than PCA.