Component and instrument failures detection using continuous mapping neural network

Abstract

Two schemes for distinction between component failure detection (CFD) and instrument failure detection (IFD) are suggested. These schemes are based on the use of a dynamic neural network capable of performing continuous mapping between two sets in the real number domain, R/sup n/ to R/sup m/. The first scheme requires only one dynamic neural net which takes both computed input signals and measurable signals provided by sensors and predicts a combination of the plant parameters and inaccessible signals. In the second scheme two dynamic neural nets are connected in a reverse direction. The first neural net is used to detect possible failures in plant components or sensors, while the second net is used for the distinction between the two different kinds of failures. Both nets use only measured signals. One net literally simulates the inverse dynamics of the plant and possesses a physical meaning; therefore, it can be trained online. The other requires a dynamic model of the plant. Simulations based on a robot arm are presented. >