Fault Detection Using Projection Pursuit Regression (ppr): a Classification versus an Estimation Based Approach

Two fault detection approaches are compared using a Projection Pursuit Regression (PPR) algorithm: ia classification approach where the fault detection PPR model is trained based on the class numbers and iian estimation approach where the PPR model is trained to predict the value of the process variable that define the class boundaries and then the corresponding class is identified by comparing the estimated value versus the limits of the fault classes. The comparison is carried on for simple illustration examples, to elucidate the main issues, and for a copolymerization process. The classification approach is found superior provided that the training data closest to the boundaries are located at equidistant locations from these boundaries.