Model-based fault detection in a high-pressure heater line

Abstract

In this chapter, we have presented a complete overview of various methodologies, starting with the construction of models of a power plant in normal and faulty conditions and continuing with its tuning by means of on-line grey-box identification. Modelling provides the basis for designing estimators that give estimates of the state variables of the system for the purpose of fault diagnostics. Both grey-box identification and estimation were based on the combined use of neural network and stochastic approximation. The solutions of the problems have been obtained by reducing the original functional optimisation problem to a non-linear programming one. This reduction was made possible by the use of feed forward neural networks, which, as is well known, exhibit excellent approximation properties. The low computational demand and good performance obtained with respect to the EKF suggest that the proposed estimator is well suited to being used in real power plant applications concerning process monitoring and fault detection over the long term. It is worth recalling that a lot of work has to be devoted to the construction of a model, its identification, and to the learning required for the design of the neural estimator. This effort provides a flexible, powerful, and accurate tool to supervise and predict how the plant performs on-line and to prevent dangerous and undesirable situations.