A Kernel Direct Decomposition-Based Monitoring Approach for Nonlinear Quality-Related Fault Detection

This article considers the issue of quality-related process monitoring. A novel kernel direct decomposition (KDD) algorithm is proposed and a KDD-based nonlinear quality-related fault detection approach is designed. The proposed KDD algorithm first maps original process variables into feature space to deal with the nonlinearities among these variables. Feature matrix is then directly decomposed into two orthogonal parts according to its full correlation with output matrix without building any regression model. Compared with conventional nonlinear methods, the KDD-based approach has the following advantages: 1) it is simpler in design as it omits the steps of constructing a regression model like kernel partial least squares (KPLS); 2) its performance is more stable because it extracts the full correlation information of feature matrix unlike KPLS-based methods which only use the partial correlation information of several selected latent variables; and 3) it has a simpler diagnosis logic since it only uses two statistics to determine the type of fault while most existing methods need four. Simulations on a literature example and a simulated industrial process are used to demonstrate the advantages of the new method.