Model-Based Fault Detection in a High-Pressure Heater Line

The section of a real power plant is considered as benchmark to sliding-window neural state estimator for the purpose of fault diagnosis. First, a complete description of the model is given with a number of details. Second, the design procedure of the neural estimator is presented that relies on a receding-horizon strategy and employ estimation functions with fixed structures in which a certain number of parameter shave to be optimized. Third, multilayer feedforward neural networks are chosen as approximating structures, whose parameters (i.e., the network weights) are chosen by using stochastic approximation. Fourth, simulation results are shown to compare the behavior of the proposed estimator with the EKF in different scenarios