Abstract

This paper focuses on the development of neural-based condition-monitoring and fault-diagnosis (CMFD) systems. Specifically, we consider the impact of the limited availability of `faulty' training data in real CMFD applications. Where limited data are available we demonstrate two ways in which performance may, in some circumstances, be improved: (1) by using fewer training data made up of roughly equal numbers of,normal' and `fault' samples; or (2) by using a `duplicate-data' training algorithm.