Sensor fault diagnosis for systems with unknown nonlinearity using neural network based nonlinear observers

A nonlinear observer for fault detection and isolation (FDI) of systems with unknown nonlinearity is presented. The nonlinear compensation term in the observer design is obtained by a `deconvolution' method and a B-spline neural network. The problem with the use of one-step ahead prediction error of the observer in FDI is discussed, and an alternative approach based on multi-step ahead prediction is proposed. A nonlinear `dedicated observer' scheme for the FDI using multiple measurements is also discussed.