Data reduction and fault diagnosis using principle of distributional equivalence

Historical data based fault diagnosis methods exploit two key strengths of the multivariate statistical tool being used: i) data compression ability, and ii) discriminatory ability. It has been shown that correspondence analysis (CA) is superior to principal components analysis (PCA) on both these counts[1], and hence is more suited for the task of fault detection and isolation(FDI). In this paper, we propose a methodology for fault diagnosis that can facilitate significant data reduction as well as better discrimination. The proposed methodology is based on the principle of distributional equivalence (PDE). The PDE is a property unique to CA and can be very useful in analyzing large datasets. The principle, when applied to historical data sets for FDI, can significantly reduce the data matrix size without significantly affecting the discriminatory ability of the CA algorithm. The data reduction ability of the proposed methodology is demonstrated using a simulation case study involving benchmark quadruple tank laboratory process. The above aspect is also validated for large scale system using benchmark Tennessee Eastman process simulation case study.