Fault Detection in Multimode Processes Through Improved Nonlinear External Analysis Regression

Abstract

External analysis serve as a cogent approach for multimode process detection in recent years. However, external analysis approach may not detect faults well because of the imprecise extraction of relations between variables. This paper proposes a fault detection approach for multimode processes, called improved nonlinear external analysis regression. External regression models between external variables and main/quality variables are established to remove mode-change-related information in the main/quality variables, ensuring that the following work is performed under a single mode. The remaining information in the main and quality variables is employed to develop an internal regression model for fault detection. Compared with existing approaches, the proposed approach has the following advantages: (1) In external regression models, applying kernel orthogonal projections to latent structures resulted in a relatively smaller number of loadings, reduced model complexity and, not least, efficient extraction of mode-change-related information in main and quality variables. (2) Internal regression model has the capacity to improve separation performance of output-related information and output-unrelated information. (3) Two comprehensive and perspicuous detection statistics are designed to accurately detect process faults. To experimentally verify the stability and superiority of the method, it is applied to a penicillin fermentation process for fault detection.