Abstract

This paper develops a technique for determining a reliable threshold for RBF classifiers. A two-phase approach is proposed to RBF classifier use in situations where unknown faults may occur: the first phase deals with the possibility of unknown faults; in the second phase, the classifier threshold is modified through retraining using all available data, including newly collected data about unknown faults. The approach is easy to use and is demonstrated to be particularly effective in classification problems where novelty detection capability is required.