Empirical process modeling in fast breeder reactors

Abstract A non-linear multi-input/single output (MISO) empirical model is introduced for monitoring vital system parameters in a nuclear reactor environment. The proposed methodology employs a scheme of non-parametric smoothing that models the local dynamics of each fitting point individually, as opposed to global modeling techniques-such as multi-layer perceptrons (MLPs)-that attempt to capture the dynamics of the entire design space. The stimulation for employing local models in monitoring rises from one's desire to capture localized idiosyncrasies of the dynamic system utilizing independent estimators. This approach alleviates the effect of negative interference between old and new observations enhancing the model prediction capabilities. Modeling the behavior of any given system comes down to a trade off between variance and bias. The building blocks of the proposed approach are tailored to each data set through two separate, adaptive procedures in order to optimize the bias-variance reconciliation. Hetero-associative schemes of the technique presented exhibit insensitivity to sensor noise and provide the operator with accurate predictions of the actual process signals. A comparison between the local model and MLP prediction capabilities is performed and the results appear in favor of the first method. The data used to demonstrate the potential of local regression have been obtained during two startup periods of the Monju fast breeder reactor (FBR).