Fault Diagnosis using Correspondence Analysis: Implementation issues and analysis

This paper presents an approach based on the use of the correspondence analysis (CA) algorithm for the task of fault detection and diagnosis. The CA algorithm analyzes the joint row-column association to represent the information content in the data matrix X. Decomposition of the information represented by this metric is shown to capture dynamic information more efficiently [1] and therefore yield superior performance from the viewpoints of data compression, discrimination and classification as well as early detection of faults. In this paper, we are discussing certain implementation issues, such as dimensional homogeneity, before correspondence analysis can be applied to any data set. We also demonstrate how these conditions can be met for the data sets obtained from an online plant. We demonstrate performance improvements over PCA and DPCA on the Tennessee Eastman problem, which is a representative benchmark problem used in the literature. CA is shown to yield vastly superior performance for the monitoring of the TE problem, when compared with PCA and DPCA.