A Method for Fault Detection and Isolation Using Neural Networks - Neural Networks, 1996., IEEE International Conference on

These two approaches both require the availability of large amounts of observed data of the monitored system in the healthy mode and in al l the possible faulty modes, which is rarely feasible in practice. In this paper, we propose a different approach. Informally speaking, it considers neural networks as parametric models and apply to them the asymptotic local approach to change detection [3, 111. For the purpose of fault detection, it only requires a nominal model of the monitored system, which can be trained with observed data of the monitored system in healthy mode. For the purpose of fault isolation, some partial physical knowledge of the monitored system is assumed to be available and expressed in the form of a small physical model, then the relationship between the nominal model and the small physical model is estimated. mode and faulty modes.