Polar classification with correspondence analysis for fault isolation

Data collected from operating plants can be mined to extract information related to both normal and fault modes of operation. Correspondence analysis (CA), that decomposes a measure of row–column association, to generate the lower dimensional space has been recently proposed [1] for this task. CA represents the association between samples and variables in terms of angle based measures on a biplot. Thus, toward clearer resolution of the faults, polar clustering and classification procedures are necessary. In this paper, we develop a methodology to mine the operating data and build such clusters. We demonstrate the application of this methodology on data generated from simulations and experiments involving representative systems,for detecting parametric changes and resolving sensor and actuator biases.