Fault localization using principal component analysis based on a new contribution to the squared prediction error

In Multivariate Statistic Process Control (MSPC), the projection methods, in particular, the Principal Component Analysis (PCA) proved their efficiencies to the problems of fault diagnosis. In this paper, we explained the use of this tool and its major interest. A projected observation in the PCA space has a Score Distance (SD) in the principal component subspace and an Orthogonal Distance (OD) or SPE (Squared Prediction Error) in the residual subspace. It is known that each of these two subspaces has a control limit. Hence, fault detection is well accomplished by studying simultaneously the two distances. The main approach for fault localization based on PCA is the contribution plots to these distances. We note that each variable has two contributions. We showed that the contribution to the SPE, proposed in the literature presents some limitations. That’s why we proposed a new form for this contribution. With this study, we contributed to the realization of a more efficient diagnosis.