Search

The problem of compensation of modelling errors for the purpose of robust fault detection based on parity relations is addressed. The idea is to approximate unmodelled non-linear dynamics by a neural network model and then to remove the effects of unmodelled dynamics from the primary residuals. The design of such a compensator takes two steps. In the first, a subset of the most informative regressors is selected. The second step entails structure determination and parameter estimation by means of numerical optimization of a criterion function. The criterion reflects a compromise between the quality of approximation and the complexity of the model structure. The results from the study on a three-tank test rig are presented and a comparison between compensated and uncompensated residuals made. It is shown that the compensated residuals represent a good basis for reliable, yet sensitive enough, fault detection and isolation. [ABSTRACT FROM AUTHOR]