Model-Based Fault-Detection in a Real Power Plant Using Neural Networks and Stochastic Approximation

Abstract

This paper describes a model-based method to solving some very difficult fault-diagnosis problems in a real system, i.e., the four heaters of a feedwater high-pressure line of a 320 MW power plant. The proposed approach exhibits very general features so that it can be applied to different types of plants. In the paper, it is shown how the combined use of feedforward neural networks and stochastic approximation allows to tune on-line a model of the plant on the basis of the real behaviour of the process. Then, after the tuning phase, the mismatch between the actual process and the mathematical model is reduced, which is crucial for fault diagnosis because these discrepancies are one of the main causes of false alarms. This, in turn, allows the connection of the simulator of the model in parallel with the real plant, thus providing crucial information on the process behaviour that allow to solve efficiently some difficult fault-detection problems.