Abstract:

When sensor parameters cannot be directly measured, estimation of their performance can be carried out using other plant variables that can be measured, so long as they are correlated with the sensor parameters. The correlations can cause computational difficulties. Well-established techniques exist for dealing with these difficulties in the case where the relationship between the predictor variables and the response variable is approximately linear, but for cases where the relationship is nonlinear, the situation is less-well understood. This paper demonstrates that estimation of sensor performance in the nonlinear case can be reliably achieved using elementary neural networks that are subjected to the method of boosting. We first apply this method to a set of data from a nuclear power plant (NPP) in Florida that has been widely studied elsewhere. We show that for this data, which is close to linear, our boosting method provides estimates that are competitive with those obtained using regularized linear models. We then consider a data set from a NPP in the Netherlands where the relationship between predictor and response variables is considerably nonlinear and show that the boosting method gives excellent estimates.
Received on January 14, 2005
References: 15