Non-linear multiscale principal component analysis for fault detection: application to pollution parameters

In the general frame of process surveillance, principal component analysis (PCA) has been often selected due to its simplicity and ability to capture the linear relations between the stationary process variables. However, the method showed limitations when dealing with industrial data that generally presents non-linear and multiscale features. The approach proposed in this study rests on the modelling using non-linear PCA coupled with artificial neural networks (ANNs) to extract the non-linear inter-correlation between variables and on the wavelet analysis to decompose each sensor signal into a set of coefficients at different scales. The contribution of each variable for each scale is then collected in separated matrices and a non-linear PCA model is constructed for each matrix. The proposed approach is applied to fault detection of pollution parameters affecting the region of Annaba in Algeria. The performance of the approach is then illustrated and compared with those of classic PCA and multiscale PCA (MSPCA).