Neural state estimators for direct model-based fault diagnosis

A model-based fault-detection scheme for nonlinear systems is proposed, which is based on finite-memory state estimation. The faults are diagnosed by means of the estimates of the fault vector (used to describe plant, actuator, and sensor faults). The fault finite-memory estimator is stated in a general nonlinear setup and the optimal functions solving the estimation problem are approximated by means of feedforward neural nets. The optimization of the neural weights consists of two phases. In the first phase, any possible " a priori" knowledge on the statistics of the random variables is used to "initialize" (off-line) the neural estimation functions. In the second one, the optimization (or training) continues online. Both off-line and online phases rely upon stochastic approximation algorithms.