Dynamic system fault diagnosis based on neural network modelling

A fault diagnosis scheme for nonlinear unknown systems comprising two stages: residual generation and residual evaluation is presented. The essential step to this goal is to build up a bank of recurrent neural network models that perform long-term predictions. This scheme has advantages over conventional FDD schemes in that it can deal not only with the nonlinearity presented in the system, but it also provides useful information about the deterioration level of system. An application of the scheme for detection and identification of faults occurring in a gas turbine engine is presented. Simulation studies show the effectiveness of proposed scheme. Copyright (C) 1998 IFAC.