Neuronal principal component analysis for the diagnosis of a non linear system

This paper present a detection and diagnosis sensor faults based on a Neuronal Non Linear Principal Component Analysis (NNLPCA) and on a Fisher Discriminant Analysis (FDA). This method is validated in simulation on a non linear system: an experimental greenhouse. Several results are presented. The neuronal approach of the NLPCA is used to underline the correlations between the variables and to estimate the non linear principal components. This NLPCA model allows to estimate the prediction error (SPE: Squared Prediction Error) and to define data classes with and without fault. The classes associated to data with fault are isolated by applying a FDA.