Virtual industrial sensors trough neural networks. Demonstration examples in nuclear power plants

Variables measured in complex industrial plants (like nuclear power plants) are related in a complex and non-explicit manner. Nevertheless, this relation can be exploited to aid the plant instrumentation system trough the use of suitable tools. Neural network technology provides such kind of tools, useful in many applications. In this work we try demonstrate the suitability of this approach by showing two examples of virtual sensors (i.e. neural networks for variable estimation) developed for pressurized water reactor nuclear power plants. Questions addressed have been input variable selection, data normalization, network architecture optimization, training set sizing, etc