The author describes a method in which a neural network is used to model the relationship between two or more sensor outputs at a time when the component or system is known to be performing satisfactorily. The neural network is then used to predict one or more of the sensor signals using the other sensor signals as inputs. The predicted signal is then compared with the corresponding actual signal. If there is a significant difference (beyond normal statistical variations), then the relationship between the sensor signals has changed, indicating that something in the component has changed since the neural network was trained (i.e. since the component or system was working satisfactorily). Several industrial applications of this technique (especially in nuclear power plants) are discussed.
展开▼
