Optimal experimental design for training of a fault diagnosis algorithm

Abstract

This paper focuses on optimal experimental design (OED) to train a projection pursuit regression (PPR) model, used for fault diagnosis. Two new experimental design methods, referred to as the Gaussian probability design and the fuzzy boundary design, are proposed and compared with a conventional factorial design. The Gaussian probability design is based on an assessment of the probability that the experimental data near a class boundary belongs to a specific class. The fuzzy boundary design is based on the identification of fuzzy class boundaries by a group of PPR models, which are developed from subsets of an initial training data set using a bootstrapping approach. Both the Gaussian probability design and the fuzzy boundary design methods automatically search for regions, where training data are sparse, and add pairs of training data on both sides of a class boundary where the data density is the lowest. The proposed design methods outperform the factorial design by reducing the fault misclassifications with the same amount of training data.