Distributed fault detection method based on locality preserving CCA and its application to industry process

Abstract

The complex interrelationships between variables and units in a plant lead to challenging monitoring, and it is necessary to design distributed monitoring schemes. Among the steps of distributed data-driven fault detection methods, process decomposition is important to ensure monitoring performance. In this paper, a data-driven mutual information approach is used for process decomposition. Traditional canonical correlation analysis (CCA) focuses on global structural information but ignores local information which is also important for process monitoring. In the locality preserving canonical correlation analysis approach, local structure information is integrated into CCA to produce a new optimization objective involving both global and local structure information for better extraction of data features. In this paper, we propose a datadriven Distributed Locality Preserving Canonical Correlation Analysis (D-LPCCA) fault detection method for addressing plant-wide process fault detection that contains nonlinear correlations. In order to set a detection limit that can better distinguish between normal and abnormal values, a kernel density estimation method is used in this paper. Finally, the effectiveness of the method proposed in this paper is verified by the Tennessee Eastman simulation platform.