Adaptive error compensation for robust fault detection

The estimation of system states or outputs through a linear observer provides redundant relations used for fault detection. However, such estimates are often inaccurate due to the model-mismatch or uncertainties. In this paper an approach is presented for robust residual generation in the presence of unmodelled nonlinearity. This is achieved by eliminating the effect of modelling error from the residual by using a neural network-based compensation scheme. The details of design procedures are described, and the learning algorithm is derived based on the gradient calculation. The detection and isolation of the small faults occurring in the gas turbine engine system is used as an illustrative example. Simulation studies show the effectiveness of the proposed scheme.